maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/clang/test/OpenMP/task_codegen.cpp

4695 lines
380 KiB
C++
Raw Normal View History

[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --function-signature --include-generated-funcs --replace-value-regex "__omp_offloading_[0-9a-z]+_[0-9a-z]+" "reduction_size[.].+[.]" "pl_cond[.].+[.|,]" --prefix-filecheck-ir-name _
// RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp -x c++ -emit-llvm %s -o - -DUNTIEDRT | FileCheck %s --check-prefix=CHECK1
[OPENMP]Add support for allocate vars in untied tasks. Local vars, marked with pragma allocate, mustbe allocate by the call of the runtime function and cannot be allocated as other local variables. Instead, we allocate a space for the pointer in private record and store the address, returned by kmpc_alloc call in this pointer. So, for untied tasks ``` #pragma omp task untied { S s; #pragma omp allocate(s) allocator(allocator) s = x; } ``` compiler generates something like this: ``` struct task_with_privates { S *ptr; }; void entry(task_with_privates *p) { S *s = p->s; switch(partid) { case 1: p->s = (S*)kmpc_alloc(); kmpc_omp_task(); br exit; case 2: *s = x; kmpc_omp_task(); br exit; case 2: ~S(s); kmpc_free((void*)s); br exit; } exit: } ``` Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D86558
2020-09-15 23:21:47 +08:00
// RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-apple-darwin10 -emit-pch -o %t %s -DUNTIEDRT
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK2
// RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp -fopenmp-enable-irbuilder -x c++ -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK3
[OpenMP] Lower taskwait using OpenMP IR Builder The code generation is exactly the same as it was. But not that the special handling of untied tasks is still handled by emitUntiedSwitch in clang. Differential Revision: https://reviews.llvm.org/D69828
2019-12-12 17:19:10 +08:00
// RUN: %clang_cc1 -fopenmp -fopenmp-enable-irbuilder -x c++ -triple x86_64-apple-darwin10 -emit-pch -o %t %s
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// RUN: %clang_cc1 -fopenmp -fopenmp-enable-irbuilder -x c++ -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK4
[OPENMP] Support for -fopenmp-simd option with compilation of simd loops only. Added support for -fopenmp-simd option that allows compilation of simd-based constructs without emission of OpenMP runtime calls. llvm-svn: 321560
2017-12-30 02:07:07 +08:00
[OpenMP][NFC] Remove SIMD check lines for non-simd tests If a test does not contain an " simd" but -fopenmp-simd RUN lines we can just check that we do not create __kmpc|__tgt calls. Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101973
2021-05-19 10:52:53 +08:00
// RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp-simd -x c++ -emit-llvm %s -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
[OPENMP] Support for -fopenmp-simd option with compilation of simd loops only. Added support for -fopenmp-simd option that allows compilation of simd-based constructs without emission of OpenMP runtime calls. llvm-svn: 321560
2017-12-30 02:07:07 +08:00
// RUN: %clang_cc1 -fopenmp-simd -x c++ -triple x86_64-apple-darwin10 -emit-pch -o %t %s
[OpenMP][NFC] Remove SIMD check lines for non-simd tests If a test does not contain an " simd" but -fopenmp-simd RUN lines we can just check that we do not create __kmpc|__tgt calls. Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101973
2021-05-19 10:52:53 +08:00
// RUN: %clang_cc1 -fopenmp-simd -x c++ -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
// expected-no-diagnostics
#ifndef HEADER
#define HEADER
[OPENMP]Add support for allocate vars in untied tasks. Local vars, marked with pragma allocate, mustbe allocate by the call of the runtime function and cannot be allocated as other local variables. Instead, we allocate a space for the pointer in private record and store the address, returned by kmpc_alloc call in this pointer. So, for untied tasks ``` #pragma omp task untied { S s; #pragma omp allocate(s) allocator(allocator) s = x; } ``` compiler generates something like this: ``` struct task_with_privates { S *ptr; }; void entry(task_with_privates *p) { S *s = p->s; switch(partid) { case 1: p->s = (S*)kmpc_alloc(); kmpc_omp_task(); br exit; case 2: *s = x; kmpc_omp_task(); br exit; case 2: ~S(s); kmpc_free((void*)s); br exit; } exit: } ``` Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D86558
2020-09-15 23:21:47 +08:00
enum omp_allocator_handle_t {
omp_null_allocator = 0,
omp_default_mem_alloc = 1,
omp_large_cap_mem_alloc = 2,
omp_const_mem_alloc = 3,
omp_high_bw_mem_alloc = 4,
omp_low_lat_mem_alloc = 5,
omp_cgroup_mem_alloc = 6,
omp_pteam_mem_alloc = 7,
omp_thread_mem_alloc = 8,
KMP_ALLOCATOR_MAX_HANDLE = __UINTPTR_MAX__
};
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
struct S {
int a;
S() : a(0) {}
S(const S &s) : a(s.a) {}
~S() {}
};
int a;
int main() {
char b;
[OPENMP] Codegen for 'depend' clause (OpenMP 4.0). If task directive has associated 'depend' clause then function kmp_int32 __kmpc_omp_task_with_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_task_t * new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list,kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called instead of __kmpc_omp_task(). If this directive has associated 'if' clause then also before a call of kmpc_omp_task_begin_if0() a function void __kmpc_omp_wait_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called. Array sections are not supported yet. llvm-svn: 240532
2015-06-24 19:01:36 +08:00
S s[2];
[OPENMP 4.0] Codegen for array sections. Added codegen for array section in 'depend' clause of 'task' directive. It emits to pointers, one for the begin of array section and another for the end of array section. Size of the section is calculated as (end + 1 - start) * sizeof(basic_element_type). llvm-svn: 246422
2015-08-31 15:32:19 +08:00
int arr[10][a];
[OPENMP 4.0] Fixed codegen for destructors in task-based directives. If private variables require destructors call at the deletion of the task, additional flag in task flags must be set. Patch fixes this problem. llvm-svn: 269039
2016-05-10 18:36:51 +08:00
#pragma omp task shared(a, b, s) priority(b)
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
{
a = 15;
b = a;
[OPENMP] Codegen for 'depend' clause (OpenMP 4.0). If task directive has associated 'depend' clause then function kmp_int32 __kmpc_omp_task_with_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_task_t * new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list,kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called instead of __kmpc_omp_task(). If this directive has associated 'if' clause then also before a call of kmpc_omp_task_begin_if0() a function void __kmpc_omp_wait_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called. Array sections are not supported yet. llvm-svn: 240532
2015-06-24 19:01:36 +08:00
s[0].a = 10;
}
[OPENMP 4.0] Codegen for array sections. Added codegen for array section in 'depend' clause of 'task' directive. It emits to pointers, one for the begin of array section and another for the end of array section. Size of the section is calculated as (end + 1 - start) * sizeof(basic_element_type). llvm-svn: 246422
2015-08-31 15:32:19 +08:00
#pragma omp task shared(a, s) depend(in : a, b, s, arr[:])
[OPENMP] Codegen for 'depend' clause (OpenMP 4.0). If task directive has associated 'depend' clause then function kmp_int32 __kmpc_omp_task_with_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_task_t * new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list,kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called instead of __kmpc_omp_task(). If this directive has associated 'if' clause then also before a call of kmpc_omp_task_begin_if0() a function void __kmpc_omp_wait_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called. Array sections are not supported yet. llvm-svn: 240532
2015-06-24 19:01:36 +08:00
{
a = 15;
s[1].a = 10;
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
}
#pragma omp task untied
{
[OPENMP]Fix PR39422: variables are not firstprivatized in task context. According to the OpenMP standard, In a task generating construct, if no default clause is present, a variable for which the data-sharing attribute is not determined by the rules above is firstprivatized. Compiler tries to implement this, but if the variable is not directly used in the task context, this variable may not be firstprivatized. Patch fixes this problem. llvm-svn: 345277
2018-10-25 23:35:27 +08:00
#pragma omp critical
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
a = 1;
}
[OPENMP 4.0] Codegen for array sections. Added codegen for array section in 'depend' clause of 'task' directive. It emits to pointers, one for the begin of array section and another for the end of array section. Size of the section is calculated as (end + 1 - start) * sizeof(basic_element_type). llvm-svn: 246422
2015-08-31 15:32:19 +08:00
#pragma omp task untied depend(out : s[0], arr[4:][b])
[OPENMP] Codegen for 'depend' clause (OpenMP 4.0). If task directive has associated 'depend' clause then function kmp_int32 __kmpc_omp_task_with_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_task_t * new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list,kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called instead of __kmpc_omp_task(). If this directive has associated 'if' clause then also before a call of kmpc_omp_task_begin_if0() a function void __kmpc_omp_wait_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called. Array sections are not supported yet. llvm-svn: 240532
2015-06-24 19:01:36 +08:00
{
a = 1;
}
[OpenMP] Adding support to the mutexinoutset dep-type Summary: this commit adds support to a new dependence type introduced in OpenMP 5.0. The LLVM OpenMP RTL already supports this feature, so we only need to modify CLANG to take advantage of them. Differential Revision: https://reviews.llvm.org/D57576 llvm-svn: 353018
2019-02-04 15:33:19 +08:00
#pragma omp task untied depend(mutexinoutset: s[0], arr[4:][b])
{
a = 1;
}
[OPENMP 4.0] Codegen for array sections. Added codegen for array section in 'depend' clause of 'task' directive. It emits to pointers, one for the begin of array section and another for the end of array section. Size of the section is calculated as (end + 1 - start) * sizeof(basic_element_type). llvm-svn: 246422
2015-08-31 15:32:19 +08:00
#pragma omp task final(true) depend(inout: a, s[1], arr[:a][3:])
[OPENMP] Codegen for 'depend' clause (OpenMP 4.0). If task directive has associated 'depend' clause then function kmp_int32 __kmpc_omp_task_with_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_task_t * new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list,kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called instead of __kmpc_omp_task(). If this directive has associated 'if' clause then also before a call of kmpc_omp_task_begin_if0() a function void __kmpc_omp_wait_deps ( ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) must be called. Array sections are not supported yet. llvm-svn: 240532
2015-06-24 19:01:36 +08:00
{
a = 2;
}
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
#pragma omp task final(true)
{
a = 2;
}
const bool flag = false;
#pragma omp task final(flag)
{
a = 3;
}
[OPENMP] Preserve alignment of the original variables for the captured references. Patch makes codegen to preserve alignment of the shared variables captured and used in OpenMP regions. llvm-svn: 247401
2015-09-11 18:29:41 +08:00
int c __attribute__((aligned(128)));
#pragma omp task final(b) shared(c)
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
{
a = 4;
[OPENMP] Preserve alignment of the original variables for the captured references. Patch makes codegen to preserve alignment of the shared variables captured and used in OpenMP regions. llvm-svn: 247401
2015-09-11 18:29:41 +08:00
c = 5;
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
}
[OPENMP]Add support for allocate vars in untied tasks. Local vars, marked with pragma allocate, mustbe allocate by the call of the runtime function and cannot be allocated as other local variables. Instead, we allocate a space for the pointer in private record and store the address, returned by kmpc_alloc call in this pointer. So, for untied tasks ``` #pragma omp task untied { S s; #pragma omp allocate(s) allocator(allocator) s = x; } ``` compiler generates something like this: ``` struct task_with_privates { S *ptr; }; void entry(task_with_privates *p) { S *s = p->s; switch(partid) { case 1: p->s = (S*)kmpc_alloc(); kmpc_omp_task(); br exit; case 2: *s = x; kmpc_omp_task(); br exit; case 2: ~S(s); kmpc_free((void*)s); br exit; } exit: } ``` Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D86558
2020-09-15 23:21:47 +08:00
#pragma omp task untied firstprivate(c) allocate(omp_pteam_mem_alloc:c)
[OPENMP] Codegen for untied tasks. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. Patch adds proper codegen for untied tasks. llvm-svn: 266853
2016-04-20 12:01:36 +08:00
{
[OPENMP]Add support for allocate vars in untied tasks. Local vars, marked with pragma allocate, mustbe allocate by the call of the runtime function and cannot be allocated as other local variables. Instead, we allocate a space for the pointer in private record and store the address, returned by kmpc_alloc call in this pointer. So, for untied tasks ``` #pragma omp task untied { S s; #pragma omp allocate(s) allocator(allocator) s = x; } ``` compiler generates something like this: ``` struct task_with_privates { S *ptr; }; void entry(task_with_privates *p) { S *s = p->s; switch(partid) { case 1: p->s = (S*)kmpc_alloc(); kmpc_omp_task(); br exit; case 2: *s = x; kmpc_omp_task(); br exit; case 2: ~S(s); kmpc_free((void*)s); br exit; } exit: } ``` Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D86558
2020-09-15 23:21:47 +08:00
S s1, s2;
#ifdef UNTIEDRT
#pragma omp allocate(s2) allocator(omp_pteam_mem_alloc)
#endif
s2.a = 0;
[OPENMP] Codegen for untied tasks. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. Patch adds proper codegen for untied tasks. llvm-svn: 266853
2016-04-20 12:01:36 +08:00
#pragma omp task
[OPENMP]Add support for allocate vars in untied tasks. Local vars, marked with pragma allocate, mustbe allocate by the call of the runtime function and cannot be allocated as other local variables. Instead, we allocate a space for the pointer in private record and store the address, returned by kmpc_alloc call in this pointer. So, for untied tasks ``` #pragma omp task untied { S s; #pragma omp allocate(s) allocator(allocator) s = x; } ``` compiler generates something like this: ``` struct task_with_privates { S *ptr; }; void entry(task_with_privates *p) { S *s = p->s; switch(partid) { case 1: p->s = (S*)kmpc_alloc(); kmpc_omp_task(); br exit; case 2: *s = x; kmpc_omp_task(); br exit; case 2: ~S(s); kmpc_free((void*)s); br exit; } exit: } ``` Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D86558
2020-09-15 23:21:47 +08:00
a = c = 4;
[OPENMP] Codegen for untied tasks. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. Patch adds proper codegen for untied tasks. llvm-svn: 266853
2016-04-20 12:01:36 +08:00
#pragma omp taskyield
s1 = S();
[OPENMP]Add support for allocate vars in untied tasks. Local vars, marked with pragma allocate, mustbe allocate by the call of the runtime function and cannot be allocated as other local variables. Instead, we allocate a space for the pointer in private record and store the address, returned by kmpc_alloc call in this pointer. So, for untied tasks ``` #pragma omp task untied { S s; #pragma omp allocate(s) allocator(allocator) s = x; } ``` compiler generates something like this: ``` struct task_with_privates { S *ptr; }; void entry(task_with_privates *p) { S *s = p->s; switch(partid) { case 1: p->s = (S*)kmpc_alloc(); kmpc_omp_task(); br exit; case 2: *s = x; kmpc_omp_task(); br exit; case 2: ~S(s); kmpc_free((void*)s); br exit; } exit: } ``` Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D86558
2020-09-15 23:21:47 +08:00
s2.a = 10;
[OPENMP] Codegen for untied tasks. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. Patch adds proper codegen for untied tasks. llvm-svn: 266853
2016-04-20 12:01:36 +08:00
#pragma omp taskwait
}
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
return a;
}
[OPENMP] Codegen for untied tasks. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. Patch adds proper codegen for untied tasks. llvm-svn: 266853
2016-04-20 12:01:36 +08:00
[OPENMP]Fix PR37671: Privatize local(private) variables in untied tasks. In untied tasks, need to allocate the space for local variales, declared in task region, when the memory for task data is allocated. THe function can be interrupted and we can exit from the function in untied task switch. Need to keep the state of the local variables in this case. Also, the compiler should not call cleanup when exiting in untied task switch until the real exit out of the declaration scope is met during execution. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D84457
2020-08-12 22:54:36 +08:00
[OPENMP] Codegen for untied tasks. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. Patch adds proper codegen for untied tasks. llvm-svn: 266853
2016-04-20 12:01:36 +08:00
[OPENMP]Add support for allocate vars in untied tasks. Local vars, marked with pragma allocate, mustbe allocate by the call of the runtime function and cannot be allocated as other local variables. Instead, we allocate a space for the pointer in private record and store the address, returned by kmpc_alloc call in this pointer. So, for untied tasks ``` #pragma omp task untied { S s; #pragma omp allocate(s) allocator(allocator) s = x; } ``` compiler generates something like this: ``` struct task_with_privates { S *ptr; }; void entry(task_with_privates *p) { S *s = p->s; switch(partid) { case 1: p->s = (S*)kmpc_alloc(); kmpc_omp_task(); br exit; case 2: *s = x; kmpc_omp_task(); br exit; case 2: ~S(s); kmpc_free((void*)s); br exit; } exit: } ``` Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D86558
2020-09-15 23:21:47 +08:00
[OPENMP] Codegen for untied tasks. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. Patch adds proper codegen for untied tasks. llvm-svn: 266853
2016-04-20 12:01:36 +08:00
[OPENMP]Fix PR37671: Privatize local(private) variables in untied tasks. In untied tasks, need to allocate the space for local variales, declared in task region, when the memory for task data is allocated. THe function can be interrupted and we can exit from the function in untied task switch. Need to keep the state of the local variables in this case. Also, the compiler should not call cleanup when exiting in untied task switch until the real exit out of the declaration scope is met during execution. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D84457
2020-08-12 22:54:36 +08:00
// s1 = S();
[OPENMP] Codegen for untied tasks. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. Patch adds proper codegen for untied tasks. llvm-svn: 266853
2016-04-20 12:01:36 +08:00
[OPENMP]Fix PR37671: Privatize local(private) variables in untied tasks. In untied tasks, need to allocate the space for local variales, declared in task region, when the memory for task data is allocated. THe function can be interrupted and we can exit from the function in untied task switch. Need to keep the state of the local variables in this case. Also, the compiler should not call cleanup when exiting in untied task switch until the real exit out of the declaration scope is met during execution. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D84457
2020-08-12 22:54:36 +08:00
[OPENMP]Fix PR39366: do not try to private field if it is not captured. The compiler is crashing if we trying to post-capture the fields implicitly captured inside of the task constructs. Seems, this kind of processing is not supported and such fields should not be firstprivatized. llvm-svn: 345177
2018-10-25 02:53:12 +08:00
struct S1 {
int a;
S1() { taskinit(); }
void taskinit() {
#pragma omp task
a = 0;
}
} s1;
[OPENMP]Fix PR47621: Variable used by task inside a template function is not made firstprivate by default Need to fix a check for the variable if it is declared in the inner OpenMP region to be able to firstprivatize it. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D88240
2020-09-25 02:46:06 +08:00
#ifdef UNTIEDRT
// FIXME: There is a buffer overflow in IrBuilder mode.
template <typename T = void>
void foobar() {
float a;
#pragma omp parallel
#pragma omp single
{
double b;
#pragma omp task
a += b;
}
}
void xxxx() {
foobar();
}
// Copy `a` to the list of shared variables
// Allocate task.
// Copy shared vars.
// Copy firstprivate value of `b`.
#endif // UNTIEDRT
[OPENMP] Initial codegen for 'omp task' directive. The task region is emmitted in several steps: Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the function: kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { TaskFunction(gtid, tt->part_id, tt->shareds); return 0; } Copy a list of shared variables to field shareds of the resulting structure kmp_task_t returned by the previous call (if any). Copy a pointer to destructions function to field destructions of the resulting structure kmp_task_t. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t *new_task), where new_task is a resulting structure from previous items. Differential Revision: http://reviews.llvm.org/D7560 llvm-svn: 231762
2015-03-10 15:28:44 +08:00
#endif
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-LABEL: define {{[^@]+}}@main
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-SAME: () #[[ATTR0:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[RETVAL:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[B:%.*]] = alloca i8, align 1
// CHECK1-NEXT: [[S:%.*]] = alloca [2 x %struct.S], align 4
// CHECK1-NEXT: [[SAVED_STACK:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__VLA_EXPR0:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 8
// CHECK1-NEXT: [[AGG_CAPTURED1:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 8
// CHECK1-NEXT: [[DOTDEP_ARR_ADDR:%.*]] = alloca [4 x %struct.kmp_depend_info], align 8
// CHECK1-NEXT: [[DEP_COUNTER_ADDR:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[AGG_CAPTURED3:%.*]] = alloca [[STRUCT_ANON_2:%.*]], align 1
// CHECK1-NEXT: [[AGG_CAPTURED4:%.*]] = alloca [[STRUCT_ANON_4:%.*]], align 1
// CHECK1-NEXT: [[DOTDEP_ARR_ADDR5:%.*]] = alloca [2 x %struct.kmp_depend_info], align 8
// CHECK1-NEXT: [[DEP_COUNTER_ADDR11:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[AGG_CAPTURED12:%.*]] = alloca [[STRUCT_ANON_6:%.*]], align 1
// CHECK1-NEXT: [[DOTDEP_ARR_ADDR13:%.*]] = alloca [2 x %struct.kmp_depend_info], align 8
// CHECK1-NEXT: [[DEP_COUNTER_ADDR19:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[AGG_CAPTURED20:%.*]] = alloca [[STRUCT_ANON_8:%.*]], align 1
// CHECK1-NEXT: [[DOTDEP_ARR_ADDR21:%.*]] = alloca [3 x %struct.kmp_depend_info], align 8
// CHECK1-NEXT: [[DEP_COUNTER_ADDR27:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[AGG_CAPTURED28:%.*]] = alloca [[STRUCT_ANON_10:%.*]], align 1
// CHECK1-NEXT: [[FLAG:%.*]] = alloca i8, align 1
// CHECK1-NEXT: [[AGG_CAPTURED29:%.*]] = alloca [[STRUCT_ANON_12:%.*]], align 1
// CHECK1-NEXT: [[C:%.*]] = alloca i32, align 128
// CHECK1-NEXT: [[AGG_CAPTURED30:%.*]] = alloca [[STRUCT_ANON_14:%.*]], align 8
// CHECK1-NEXT: [[AGG_CAPTURED31:%.*]] = alloca [[STRUCT_ANON_16:%.*]], align 1
// CHECK1-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
// CHECK1-NEXT: store i32 0, i32* [[RETVAL]], align 4
// CHECK1-NEXT: [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i32 0, i32 0
// CHECK1-NEXT: [[ARRAYCTOR_END:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAY_BEGIN]], i64 2
// CHECK1-NEXT: br label [[ARRAYCTOR_LOOP:%.*]]
// CHECK1: arrayctor.loop:
// CHECK1-NEXT: [[ARRAYCTOR_CUR:%.*]] = phi %struct.S* [ [[ARRAY_BEGIN]], [[ENTRY:%.*]] ], [ [[ARRAYCTOR_NEXT:%.*]], [[ARRAYCTOR_LOOP]] ]
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[ARRAYCTOR_CUR]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[ARRAYCTOR_NEXT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYCTOR_CUR]], i64 1
// CHECK1-NEXT: [[ARRAYCTOR_DONE:%.*]] = icmp eq %struct.S* [[ARRAYCTOR_NEXT]], [[ARRAYCTOR_END]]
// CHECK1-NEXT: br i1 [[ARRAYCTOR_DONE]], label [[ARRAYCTOR_CONT:%.*]], label [[ARRAYCTOR_LOOP]]
// CHECK1: arrayctor.cont:
// CHECK1-NEXT: [[TMP1:%.*]] = load i32, i32* @a, align 4
// CHECK1-NEXT: [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
// CHECK1-NEXT: [[TMP3:%.*]] = call i8* @llvm.stacksave()
// CHECK1-NEXT: store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = mul nuw i64 10, [[TMP2]]
// CHECK1-NEXT: [[VLA:%.*]] = alloca i32, i64 [[TMP4]], align 16
// CHECK1-NEXT: store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK1-NEXT: store i8* [[B]], i8** [[TMP5]], align 8
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
// CHECK1-NEXT: store [2 x %struct.S]* [[S]], [2 x %struct.S]** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP7:%.*]] = load i8, i8* [[B]], align 1
// CHECK1-NEXT: [[CONV:%.*]] = sext i8 [[TMP7]] to i32
// CHECK1-NEXT: [[TMP8:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 33, i64 40, i64 16, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast i8* [[TMP8]] to %struct.kmp_task_t_with_privates*
// CHECK1-NEXT: [[TMP10:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP9]], i32 0, i32 0
// CHECK1-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP10]], i32 0, i32 0
// CHECK1-NEXT: [[TMP12:%.*]] = load i8*, i8** [[TMP11]], align 8
// CHECK1-NEXT: [[TMP13:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
// CHECK1-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP12]], i8* align 8 [[TMP13]], i64 16, i1 false)
// CHECK1-NEXT: [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP10]], i32 0, i32 4
// CHECK1-NEXT: [[TMP15:%.*]] = bitcast %union.kmp_cmplrdata_t* [[TMP14]] to i32*
// CHECK1-NEXT: store i32 [[CONV]], i32* [[TMP15]], align 8
// CHECK1-NEXT: [[TMP16:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP8]])
// CHECK1-NEXT: [[TMP17:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[AGG_CAPTURED1]], i32 0, i32 0
// CHECK1-NEXT: store [2 x %struct.S]* [[S]], [2 x %struct.S]** [[TMP17]], align 8
// CHECK1-NEXT: [[TMP18:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.1*)* @.omp_task_entry..2 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP19:%.*]] = bitcast i8* [[TMP18]] to %struct.kmp_task_t_with_privates.1*
// CHECK1-NEXT: [[TMP20:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_1:%.*]], %struct.kmp_task_t_with_privates.1* [[TMP19]], i32 0, i32 0
// CHECK1-NEXT: [[TMP21:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP20]], i32 0, i32 0
// CHECK1-NEXT: [[TMP22:%.*]] = load i8*, i8** [[TMP21]], align 8
// CHECK1-NEXT: [[TMP23:%.*]] = bitcast %struct.anon.0* [[AGG_CAPTURED1]] to i8*
// CHECK1-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP22]], i8* align 8 [[TMP23]], i64 8, i1 false)
// CHECK1-NEXT: [[TMP24:%.*]] = getelementptr inbounds [4 x %struct.kmp_depend_info], [4 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR]], i64 0, i64 0
// CHECK1-NEXT: [[TMP25:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO:%.*]], %struct.kmp_depend_info* [[TMP24]], i64 0
// CHECK1-NEXT: [[TMP26:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP25]], i32 0, i32 0
// CHECK1-NEXT: store i64 ptrtoint (i32* @a to i64), i64* [[TMP26]], align 8
// CHECK1-NEXT: [[TMP27:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP25]], i32 0, i32 1
// CHECK1-NEXT: store i64 4, i64* [[TMP27]], align 8
// CHECK1-NEXT: [[TMP28:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP25]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 1, i8* [[TMP28]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP29:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i64 1
// CHECK1-NEXT: [[TMP30:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP29]], i32 0, i32 0
// CHECK1-NEXT: [[TMP31:%.*]] = ptrtoint i8* [[B]] to i64
// CHECK1-NEXT: store i64 [[TMP31]], i64* [[TMP30]], align 8
// CHECK1-NEXT: [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP29]], i32 0, i32 1
// CHECK1-NEXT: store i64 1, i64* [[TMP32]], align 8
// CHECK1-NEXT: [[TMP33:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP29]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 1, i8* [[TMP33]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP34:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i64 2
// CHECK1-NEXT: [[TMP35:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP34]], i32 0, i32 0
// CHECK1-NEXT: [[TMP36:%.*]] = ptrtoint [2 x %struct.S]* [[S]] to i64
// CHECK1-NEXT: store i64 [[TMP36]], i64* [[TMP35]], align 8
// CHECK1-NEXT: [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP34]], i32 0, i32 1
// CHECK1-NEXT: store i64 8, i64* [[TMP37]], align 8
// CHECK1-NEXT: [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP34]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 1, i8* [[TMP38]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP39:%.*]] = mul nsw i64 0, [[TMP2]]
// CHECK1-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP39]]
// CHECK1-NEXT: [[TMP40:%.*]] = mul nsw i64 9, [[TMP2]]
// CHECK1-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP40]]
// CHECK1-NEXT: [[TMP41:%.*]] = getelementptr i32, i32* [[ARRAYIDX2]], i32 1
// CHECK1-NEXT: [[TMP42:%.*]] = ptrtoint i32* [[ARRAYIDX]] to i64
// CHECK1-NEXT: [[TMP43:%.*]] = ptrtoint i32* [[TMP41]] to i64
// CHECK1-NEXT: [[TMP44:%.*]] = sub nuw i64 [[TMP43]], [[TMP42]]
// CHECK1-NEXT: [[TMP45:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i64 3
// CHECK1-NEXT: [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP45]], i32 0, i32 0
// CHECK1-NEXT: [[TMP47:%.*]] = ptrtoint i32* [[ARRAYIDX]] to i64
// CHECK1-NEXT: store i64 [[TMP47]], i64* [[TMP46]], align 8
// CHECK1-NEXT: [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP45]], i32 0, i32 1
// CHECK1-NEXT: store i64 [[TMP44]], i64* [[TMP48]], align 8
// CHECK1-NEXT: [[TMP49:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP45]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 1, i8* [[TMP49]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i64 4, i64* [[DEP_COUNTER_ADDR]], align 8
// CHECK1-NEXT: [[TMP50:%.*]] = bitcast %struct.kmp_depend_info* [[TMP24]] to i8*
// CHECK1-NEXT: [[TMP51:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP18]], i32 4, i8* [[TMP50]], i32 0, i8* null)
// CHECK1-NEXT: [[TMP52:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.3*)* @.omp_task_entry..4 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP53:%.*]] = bitcast i8* [[TMP52]] to %struct.kmp_task_t_with_privates.3*
// CHECK1-NEXT: [[TMP54:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_3:%.*]], %struct.kmp_task_t_with_privates.3* [[TMP53]], i32 0, i32 0
// CHECK1-NEXT: [[TMP55:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP54]], i32 0, i32 2
// CHECK1-NEXT: store i32 0, i32* [[TMP55]], align 8
// CHECK1-NEXT: [[TMP56:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP52]])
// CHECK1-NEXT: [[TMP57:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.5*)* @.omp_task_entry..6 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP58:%.*]] = bitcast i8* [[TMP57]] to %struct.kmp_task_t_with_privates.5*
// CHECK1-NEXT: [[TMP59:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_5:%.*]], %struct.kmp_task_t_with_privates.5* [[TMP58]], i32 0, i32 0
// CHECK1-NEXT: [[TMP60:%.*]] = getelementptr inbounds [2 x %struct.kmp_depend_info], [2 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR5]], i64 0, i64 0
// CHECK1-NEXT: [[ARRAYIDX6:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 0
// CHECK1-NEXT: [[TMP61:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i64 0
// CHECK1-NEXT: [[TMP62:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP61]], i32 0, i32 0
// CHECK1-NEXT: [[TMP63:%.*]] = ptrtoint %struct.S* [[ARRAYIDX6]] to i64
// CHECK1-NEXT: store i64 [[TMP63]], i64* [[TMP62]], align 8
// CHECK1-NEXT: [[TMP64:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP61]], i32 0, i32 1
// CHECK1-NEXT: store i64 4, i64* [[TMP64]], align 8
// CHECK1-NEXT: [[TMP65:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP61]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 3, i8* [[TMP65]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP66:%.*]] = load i8, i8* [[B]], align 1
// CHECK1-NEXT: [[TMP67:%.*]] = sext i8 [[TMP66]] to i64
// CHECK1-NEXT: [[TMP68:%.*]] = mul nsw i64 4, [[TMP2]]
// CHECK1-NEXT: [[ARRAYIDX7:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP68]]
// CHECK1-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX7]], i64 [[TMP67]]
// CHECK1-NEXT: [[TMP69:%.*]] = load i8, i8* [[B]], align 1
// CHECK1-NEXT: [[TMP70:%.*]] = sext i8 [[TMP69]] to i64
// CHECK1-NEXT: [[TMP71:%.*]] = mul nsw i64 9, [[TMP2]]
// CHECK1-NEXT: [[ARRAYIDX9:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP71]]
// CHECK1-NEXT: [[ARRAYIDX10:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX9]], i64 [[TMP70]]
// CHECK1-NEXT: [[TMP72:%.*]] = getelementptr i32, i32* [[ARRAYIDX10]], i32 1
// CHECK1-NEXT: [[TMP73:%.*]] = ptrtoint i32* [[ARRAYIDX8]] to i64
// CHECK1-NEXT: [[TMP74:%.*]] = ptrtoint i32* [[TMP72]] to i64
// CHECK1-NEXT: [[TMP75:%.*]] = sub nuw i64 [[TMP74]], [[TMP73]]
// CHECK1-NEXT: [[TMP76:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i64 1
// CHECK1-NEXT: [[TMP77:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP76]], i32 0, i32 0
// CHECK1-NEXT: [[TMP78:%.*]] = ptrtoint i32* [[ARRAYIDX8]] to i64
// CHECK1-NEXT: store i64 [[TMP78]], i64* [[TMP77]], align 8
// CHECK1-NEXT: [[TMP79:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP76]], i32 0, i32 1
// CHECK1-NEXT: store i64 [[TMP75]], i64* [[TMP79]], align 8
// CHECK1-NEXT: [[TMP80:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP76]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 3, i8* [[TMP80]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i64 2, i64* [[DEP_COUNTER_ADDR11]], align 8
// CHECK1-NEXT: [[TMP81:%.*]] = bitcast %struct.kmp_depend_info* [[TMP60]] to i8*
// CHECK1-NEXT: [[TMP82:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP59]], i32 0, i32 2
// CHECK1-NEXT: store i32 0, i32* [[TMP82]], align 8
// CHECK1-NEXT: [[TMP83:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP57]], i32 2, i8* [[TMP81]], i32 0, i8* null)
// CHECK1-NEXT: [[TMP84:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.7*)* @.omp_task_entry..8 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP85:%.*]] = bitcast i8* [[TMP84]] to %struct.kmp_task_t_with_privates.7*
// CHECK1-NEXT: [[TMP86:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_7:%.*]], %struct.kmp_task_t_with_privates.7* [[TMP85]], i32 0, i32 0
// CHECK1-NEXT: [[TMP87:%.*]] = getelementptr inbounds [2 x %struct.kmp_depend_info], [2 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR13]], i64 0, i64 0
// CHECK1-NEXT: [[ARRAYIDX14:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 0
// CHECK1-NEXT: [[TMP88:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i64 0
// CHECK1-NEXT: [[TMP89:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP88]], i32 0, i32 0
// CHECK1-NEXT: [[TMP90:%.*]] = ptrtoint %struct.S* [[ARRAYIDX14]] to i64
// CHECK1-NEXT: store i64 [[TMP90]], i64* [[TMP89]], align 8
// CHECK1-NEXT: [[TMP91:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP88]], i32 0, i32 1
// CHECK1-NEXT: store i64 4, i64* [[TMP91]], align 8
// CHECK1-NEXT: [[TMP92:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP88]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 4, i8* [[TMP92]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP93:%.*]] = load i8, i8* [[B]], align 1
// CHECK1-NEXT: [[TMP94:%.*]] = sext i8 [[TMP93]] to i64
// CHECK1-NEXT: [[TMP95:%.*]] = mul nsw i64 4, [[TMP2]]
// CHECK1-NEXT: [[ARRAYIDX15:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP95]]
// CHECK1-NEXT: [[ARRAYIDX16:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX15]], i64 [[TMP94]]
// CHECK1-NEXT: [[TMP96:%.*]] = load i8, i8* [[B]], align 1
// CHECK1-NEXT: [[TMP97:%.*]] = sext i8 [[TMP96]] to i64
// CHECK1-NEXT: [[TMP98:%.*]] = mul nsw i64 9, [[TMP2]]
// CHECK1-NEXT: [[ARRAYIDX17:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP98]]
// CHECK1-NEXT: [[ARRAYIDX18:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX17]], i64 [[TMP97]]
// CHECK1-NEXT: [[TMP99:%.*]] = getelementptr i32, i32* [[ARRAYIDX18]], i32 1
// CHECK1-NEXT: [[TMP100:%.*]] = ptrtoint i32* [[ARRAYIDX16]] to i64
// CHECK1-NEXT: [[TMP101:%.*]] = ptrtoint i32* [[TMP99]] to i64
// CHECK1-NEXT: [[TMP102:%.*]] = sub nuw i64 [[TMP101]], [[TMP100]]
// CHECK1-NEXT: [[TMP103:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i64 1
// CHECK1-NEXT: [[TMP104:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP103]], i32 0, i32 0
// CHECK1-NEXT: [[TMP105:%.*]] = ptrtoint i32* [[ARRAYIDX16]] to i64
// CHECK1-NEXT: store i64 [[TMP105]], i64* [[TMP104]], align 8
// CHECK1-NEXT: [[TMP106:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP103]], i32 0, i32 1
// CHECK1-NEXT: store i64 [[TMP102]], i64* [[TMP106]], align 8
// CHECK1-NEXT: [[TMP107:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP103]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 4, i8* [[TMP107]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i64 2, i64* [[DEP_COUNTER_ADDR19]], align 8
// CHECK1-NEXT: [[TMP108:%.*]] = bitcast %struct.kmp_depend_info* [[TMP87]] to i8*
// CHECK1-NEXT: [[TMP109:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP86]], i32 0, i32 2
// CHECK1-NEXT: store i32 0, i32* [[TMP109]], align 8
// CHECK1-NEXT: [[TMP110:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP84]], i32 2, i8* [[TMP108]], i32 0, i8* null)
// CHECK1-NEXT: [[TMP111:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 3, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.9*)* @.omp_task_entry..10 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP112:%.*]] = bitcast i8* [[TMP111]] to %struct.kmp_task_t_with_privates.9*
// CHECK1-NEXT: [[TMP113:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_9:%.*]], %struct.kmp_task_t_with_privates.9* [[TMP112]], i32 0, i32 0
// CHECK1-NEXT: [[TMP114:%.*]] = getelementptr inbounds [3 x %struct.kmp_depend_info], [3 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR21]], i64 0, i64 0
// CHECK1-NEXT: [[TMP115:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i64 0
// CHECK1-NEXT: [[TMP116:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP115]], i32 0, i32 0
// CHECK1-NEXT: store i64 ptrtoint (i32* @a to i64), i64* [[TMP116]], align 8
// CHECK1-NEXT: [[TMP117:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP115]], i32 0, i32 1
// CHECK1-NEXT: store i64 4, i64* [[TMP117]], align 8
// CHECK1-NEXT: [[TMP118:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP115]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 3, i8* [[TMP118]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[ARRAYIDX22:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 1
// CHECK1-NEXT: [[TMP119:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i64 1
// CHECK1-NEXT: [[TMP120:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP119]], i32 0, i32 0
// CHECK1-NEXT: [[TMP121:%.*]] = ptrtoint %struct.S* [[ARRAYIDX22]] to i64
// CHECK1-NEXT: store i64 [[TMP121]], i64* [[TMP120]], align 8
// CHECK1-NEXT: [[TMP122:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP119]], i32 0, i32 1
// CHECK1-NEXT: store i64 4, i64* [[TMP122]], align 8
// CHECK1-NEXT: [[TMP123:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP119]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 3, i8* [[TMP123]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP124:%.*]] = mul nsw i64 0, [[TMP2]]
// CHECK1-NEXT: [[ARRAYIDX23:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP124]]
// CHECK1-NEXT: [[ARRAYIDX24:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX23]], i64 3
// CHECK1-NEXT: [[TMP125:%.*]] = load i32, i32* @a, align 4
// CHECK1-NEXT: [[TMP126:%.*]] = sext i32 [[TMP125]] to i64
// CHECK1-NEXT: [[LEN_SUB_1:%.*]] = sub nsw i64 [[TMP126]], 1
// CHECK1-NEXT: [[TMP127:%.*]] = load i32, i32* @a, align 4
// CHECK1-NEXT: [[TMP128:%.*]] = sext i32 [[TMP127]] to i64
// CHECK1-NEXT: [[LB_ADD_LEN:%.*]] = add nsw i64 -1, [[TMP128]]
// CHECK1-NEXT: [[TMP129:%.*]] = mul nsw i64 [[LB_ADD_LEN]], [[TMP2]]
// CHECK1-NEXT: [[ARRAYIDX25:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP129]]
// CHECK1-NEXT: [[ARRAYIDX26:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX25]], i64 [[LEN_SUB_1]]
// CHECK1-NEXT: [[TMP130:%.*]] = getelementptr i32, i32* [[ARRAYIDX26]], i32 1
// CHECK1-NEXT: [[TMP131:%.*]] = ptrtoint i32* [[ARRAYIDX24]] to i64
// CHECK1-NEXT: [[TMP132:%.*]] = ptrtoint i32* [[TMP130]] to i64
// CHECK1-NEXT: [[TMP133:%.*]] = sub nuw i64 [[TMP132]], [[TMP131]]
// CHECK1-NEXT: [[TMP134:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i64 2
// CHECK1-NEXT: [[TMP135:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP134]], i32 0, i32 0
// CHECK1-NEXT: [[TMP136:%.*]] = ptrtoint i32* [[ARRAYIDX24]] to i64
// CHECK1-NEXT: store i64 [[TMP136]], i64* [[TMP135]], align 8
// CHECK1-NEXT: [[TMP137:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP134]], i32 0, i32 1
// CHECK1-NEXT: store i64 [[TMP133]], i64* [[TMP137]], align 8
// CHECK1-NEXT: [[TMP138:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP134]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK1-NEXT: store i8 3, i8* [[TMP138]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i64 3, i64* [[DEP_COUNTER_ADDR27]], align 8
// CHECK1-NEXT: [[TMP139:%.*]] = bitcast %struct.kmp_depend_info* [[TMP114]] to i8*
// CHECK1-NEXT: [[TMP140:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP111]], i32 3, i8* [[TMP139]], i32 0, i8* null)
// CHECK1-NEXT: [[TMP141:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 3, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.11*)* @.omp_task_entry..12 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP142:%.*]] = bitcast i8* [[TMP141]] to %struct.kmp_task_t_with_privates.11*
// CHECK1-NEXT: [[TMP143:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_11:%.*]], %struct.kmp_task_t_with_privates.11* [[TMP142]], i32 0, i32 0
// CHECK1-NEXT: [[TMP144:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP141]])
// CHECK1-NEXT: store i8 0, i8* [[FLAG]], align 1
// CHECK1-NEXT: [[TMP145:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.13*)* @.omp_task_entry..14 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP146:%.*]] = bitcast i8* [[TMP145]] to %struct.kmp_task_t_with_privates.13*
// CHECK1-NEXT: [[TMP147:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_13:%.*]], %struct.kmp_task_t_with_privates.13* [[TMP146]], i32 0, i32 0
// CHECK1-NEXT: [[TMP148:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP145]])
// CHECK1-NEXT: [[TMP149:%.*]] = getelementptr inbounds [[STRUCT_ANON_14]], %struct.anon.14* [[AGG_CAPTURED30]], i32 0, i32 0
// CHECK1-NEXT: store i32* [[C]], i32** [[TMP149]], align 8
// CHECK1-NEXT: [[TMP150:%.*]] = load i8, i8* [[B]], align 1
// CHECK1-NEXT: [[TOBOOL:%.*]] = icmp ne i8 [[TMP150]], 0
// CHECK1-NEXT: [[TMP151:%.*]] = select i1 [[TOBOOL]], i32 2, i32 0
// CHECK1-NEXT: [[TMP152:%.*]] = or i32 [[TMP151]], 1
// CHECK1-NEXT: [[TMP153:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP152]], i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.15*)* @.omp_task_entry..16 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP154:%.*]] = bitcast i8* [[TMP153]] to %struct.kmp_task_t_with_privates.15*
// CHECK1-NEXT: [[TMP155:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_15:%.*]], %struct.kmp_task_t_with_privates.15* [[TMP154]], i32 0, i32 0
// CHECK1-NEXT: [[TMP156:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP155]], i32 0, i32 0
// CHECK1-NEXT: [[TMP157:%.*]] = load i8*, i8** [[TMP156]], align 8
// CHECK1-NEXT: [[TMP158:%.*]] = bitcast %struct.anon.14* [[AGG_CAPTURED30]] to i8*
// CHECK1-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP157]], i8* align 8 [[TMP158]], i64 8, i1 false)
// CHECK1-NEXT: [[TMP159:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP153]])
// CHECK1-NEXT: [[TMP160:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 0, i64 256, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.19*)* @.omp_task_entry..21 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP161:%.*]] = bitcast i8* [[TMP160]] to %struct.kmp_task_t_with_privates.19*
// CHECK1-NEXT: [[TMP162:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19:%.*]], %struct.kmp_task_t_with_privates.19* [[TMP161]], i32 0, i32 0
// CHECK1-NEXT: [[TMP163:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19]], %struct.kmp_task_t_with_privates.19* [[TMP161]], i32 0, i32 2
// CHECK1-NEXT: [[TMP164:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20:%.*]], %struct..kmp_privates.t.20* [[TMP163]], i32 0, i32 0
// CHECK1-NEXT: [[TMP165:%.*]] = load i32, i32* [[C]], align 128
// CHECK1-NEXT: store i32 [[TMP165]], i32* [[TMP164]], align 128
// CHECK1-NEXT: [[TMP166:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP162]], i32 0, i32 2
// CHECK1-NEXT: store i32 0, i32* [[TMP166]], align 16
// CHECK1-NEXT: [[TMP167:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP160]])
// CHECK1-NEXT: [[TMP168:%.*]] = load i32, i32* @a, align 4
// CHECK1-NEXT: store i32 [[TMP168]], i32* [[RETVAL]], align 4
// CHECK1-NEXT: [[TMP169:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
// CHECK1-NEXT: call void @llvm.stackrestore(i8* [[TMP169]])
// CHECK1-NEXT: [[ARRAY_BEGIN32:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i32 0, i32 0
// CHECK1-NEXT: [[TMP170:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN32]], i64 2
// CHECK1-NEXT: br label [[ARRAYDESTROY_BODY:%.*]]
// CHECK1: arraydestroy.body:
// CHECK1-NEXT: [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP170]], [[ARRAYCTOR_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
// CHECK1-NEXT: [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4:[0-9]+]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN32]]
// CHECK1-NEXT: br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE33:%.*]], label [[ARRAYDESTROY_BODY]]
// CHECK1: arraydestroy.done33:
// CHECK1-NEXT: [[TMP171:%.*]] = load i32, i32* [[RETVAL]], align 4
// CHECK1-NEXT: ret i32 [[TMP171]]
//
//
// CHECK1-LABEL: define {{[^@]+}}@_ZN1SC1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1:[0-9]+]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK1-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK1-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SC2Ev(%struct.S* noundef [[THIS1]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry.
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias noundef [[TMP1:%.*]]) #[[ATTR3:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META3:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META6:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META8:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META10:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !12
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !12
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !12
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !12
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !12
// CHECK1-NEXT: store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !12
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !12
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 15, i32* @a, align 4
// CHECK1-NEXT: [[TMP11:%.*]] = load i32, i32* @a, align 4
// CHECK1-NEXT: [[CONV_I:%.*]] = trunc i32 [[TMP11]] to i8
// CHECK1-NEXT: [[TMP12:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP10]], i32 0, i32 0
// CHECK1-NEXT: [[TMP13:%.*]] = load i8*, i8** [[TMP12]], align 8
// CHECK1-NEXT: store i8 [[CONV_I]], i8* [[TMP13]], align 1
// CHECK1-NEXT: [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP10]], i32 0, i32 1
// CHECK1-NEXT: [[TMP15:%.*]] = load [2 x %struct.S]*, [2 x %struct.S]** [[TMP14]], align 8
// CHECK1-NEXT: [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[TMP15]], i64 0, i64 0
// CHECK1-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAYIDX_I]], i32 0, i32 0
// CHECK1-NEXT: store i32 10, i32* [[A_I]], align 4
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..2
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.1* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.0*, align 8
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.1*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.1* [[TMP1]], %struct.kmp_task_t_with_privates.1** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.1*, %struct.kmp_task_t_with_privates.1** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_1:%.*]], %struct.kmp_task_t_with_privates.1* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.0*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.1* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META13:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META16:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META18:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META20:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !22
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !22
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !22
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !22
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !22
// CHECK1-NEXT: store %struct.anon.0* [[TMP8]], %struct.anon.0** [[__CONTEXT_ADDR_I]], align 8, !noalias !22
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR_I]], align 8, !noalias !22
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 15, i32* @a, align 4
// CHECK1-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP10]], i32 0, i32 0
// CHECK1-NEXT: [[TMP12:%.*]] = load [2 x %struct.S]*, [2 x %struct.S]** [[TMP11]], align 8
// CHECK1-NEXT: [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[TMP12]], i64 0, i64 1
// CHECK1-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAYIDX_I]], i32 0, i32 0
// CHECK1-NEXT: store i32 10, i32* [[A_I]], align 4
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..4
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.3* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.2*, align 8
// CHECK1-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.3*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.3* [[TMP1]], %struct.kmp_task_t_with_privates.3** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.3*, %struct.kmp_task_t_with_privates.3** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_3:%.*]], %struct.kmp_task_t_with_privates.3* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.2*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.3* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META23:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META26:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META28:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META30:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !32
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !32
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !32
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !32
// CHECK1-NEXT: store %struct.anon.2* [[TMP8]], %struct.anon.2** [[__CONTEXT_ADDR_I]], align 8, !noalias !32
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon.2*, %struct.anon.2** [[__CONTEXT_ADDR_I]], align 8, !noalias !32
// CHECK1-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK1-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK1-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK1-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK1-NEXT: ]
// CHECK1: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK1: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !32
// CHECK1-NEXT: [[TMP15:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !32
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP16:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP14]], i8* [[TMP15]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
// CHECK1: .untied.jmp.1.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP17:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !32
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: call void @__kmpc_critical(%struct.ident_t* @[[GLOB1]], i32 [[TMP17]], [8 x i32]* @.gomp_critical_user_.var) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: call void @__kmpc_end_critical(%struct.ident_t* @[[GLOB1]], i32 [[TMP17]], [8 x i32]* @.gomp_critical_user_.var) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[CLEANUP_I]]
// CHECK1: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__3_EXIT]]
// CHECK1: .omp_outlined..3.exit:
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..6
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.5* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.4*, align 8
// CHECK1-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.5*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.5* [[TMP1]], %struct.kmp_task_t_with_privates.5** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.5*, %struct.kmp_task_t_with_privates.5** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_5:%.*]], %struct.kmp_task_t_with_privates.5* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.4*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.5* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META33:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META36:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META38:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META40:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !42
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !42
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !42
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !42
// CHECK1-NEXT: store %struct.anon.4* [[TMP8]], %struct.anon.4** [[__CONTEXT_ADDR_I]], align 8, !noalias !42
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon.4*, %struct.anon.4** [[__CONTEXT_ADDR_I]], align 8, !noalias !42
// CHECK1-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK1-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK1-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK1-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK1-NEXT: ]
// CHECK1: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK1: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !42
// CHECK1-NEXT: [[TMP15:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !42
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP16:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP14]], i8* [[TMP15]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__5_EXIT:%.*]]
// CHECK1: .untied.jmp.1.i:
// CHECK1-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[CLEANUP_I]]
// CHECK1: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__5_EXIT]]
// CHECK1: .omp_outlined..5.exit:
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.7* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.6*, align 8
// CHECK1-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.7*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.7* [[TMP1]], %struct.kmp_task_t_with_privates.7** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.7*, %struct.kmp_task_t_with_privates.7** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_7:%.*]], %struct.kmp_task_t_with_privates.7* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.6*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.7* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META43:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META46:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META48:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META50:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !52
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !52
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !52
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !52
// CHECK1-NEXT: store %struct.anon.6* [[TMP8]], %struct.anon.6** [[__CONTEXT_ADDR_I]], align 8, !noalias !52
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon.6*, %struct.anon.6** [[__CONTEXT_ADDR_I]], align 8, !noalias !52
// CHECK1-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK1-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK1-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK1-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK1-NEXT: ]
// CHECK1: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK1: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !52
// CHECK1-NEXT: [[TMP15:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !52
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP16:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP14]], i8* [[TMP15]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__7_EXIT:%.*]]
// CHECK1: .untied.jmp.1.i:
// CHECK1-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[CLEANUP_I]]
// CHECK1: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__7_EXIT]]
// CHECK1: .omp_outlined..7.exit:
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..10
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.9* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.8*, align 8
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.9*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.9* [[TMP1]], %struct.kmp_task_t_with_privates.9** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.9*, %struct.kmp_task_t_with_privates.9** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_9:%.*]], %struct.kmp_task_t_with_privates.9* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.8*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.9* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META53:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META56:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META58:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META60:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !62
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !62
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !62
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !62
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !62
// CHECK1-NEXT: store %struct.anon.8* [[TMP8]], %struct.anon.8** [[__CONTEXT_ADDR_I]], align 8, !noalias !62
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon.8*, %struct.anon.8** [[__CONTEXT_ADDR_I]], align 8, !noalias !62
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 2, i32* @a, align 4
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..12
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.11* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.10*, align 8
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.11*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.11* [[TMP1]], %struct.kmp_task_t_with_privates.11** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.11*, %struct.kmp_task_t_with_privates.11** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_11:%.*]], %struct.kmp_task_t_with_privates.11* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.10*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.11* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META63:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META66:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META68:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META70:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !72
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !72
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !72
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !72
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !72
// CHECK1-NEXT: store %struct.anon.10* [[TMP8]], %struct.anon.10** [[__CONTEXT_ADDR_I]], align 8, !noalias !72
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon.10*, %struct.anon.10** [[__CONTEXT_ADDR_I]], align 8, !noalias !72
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 2, i32* @a, align 4
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..14
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.13* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.12*, align 8
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.13*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.13* [[TMP1]], %struct.kmp_task_t_with_privates.13** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.13*, %struct.kmp_task_t_with_privates.13** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_13:%.*]], %struct.kmp_task_t_with_privates.13* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.12*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.13* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META73:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META76:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META78:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META80:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !82
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !82
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !82
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !82
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !82
// CHECK1-NEXT: store %struct.anon.12* [[TMP8]], %struct.anon.12** [[__CONTEXT_ADDR_I]], align 8, !noalias !82
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon.12*, %struct.anon.12** [[__CONTEXT_ADDR_I]], align 8, !noalias !82
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 3, i32* @a, align 4
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..16
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.15* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.14*, align 8
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.15*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.15* [[TMP1]], %struct.kmp_task_t_with_privates.15** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.15*, %struct.kmp_task_t_with_privates.15** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_15:%.*]], %struct.kmp_task_t_with_privates.15* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.14*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.15* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META83:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META86:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META88:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META90:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !92
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !92
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !92
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !92
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !92
// CHECK1-NEXT: store %struct.anon.14* [[TMP8]], %struct.anon.14** [[__CONTEXT_ADDR_I]], align 8, !noalias !92
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon.14*, %struct.anon.14** [[__CONTEXT_ADDR_I]], align 8, !noalias !92
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 4, i32* @a, align 4
// CHECK1-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_14:%.*]], %struct.anon.14* [[TMP10]], i32 0, i32 0
// CHECK1-NEXT: [[TMP12:%.*]] = load i32*, i32** [[TMP11]], align 8
// CHECK1-NEXT: store i32 5, i32* [[TMP12]], align 128
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_privates_map.
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct..kmp_privates.t* noalias noundef [[TMP0:%.*]], i32** noalias noundef [[TMP1:%.*]]) #[[ATTR7:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 8
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca i32**, align 8
// CHECK1-NEXT: store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 8
// CHECK1-NEXT: store i32** [[TMP1]], i32*** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP2]], i32 0, i32 0
// CHECK1-NEXT: [[TMP4:%.*]] = load i32**, i32*** [[DOTADDR1]], align 8
// CHECK1-NEXT: store i32* [[TMP3]], i32** [[TMP4]], align 8
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..19
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.18* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.17*, align 8
// CHECK1-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.18*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.18* [[TMP1]], %struct.kmp_task_t_with_privates.18** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.18*, %struct.kmp_task_t_with_privates.18** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18:%.*]], %struct.kmp_task_t_with_privates.18* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 128
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.17*
// CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18]], %struct.kmp_task_t_with_privates.18* [[TMP3]], i32 0, i32 2
// CHECK1-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
// CHECK1-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.18* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META93:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META96:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META98:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META100:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !102
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !102
// CHECK1-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !102
// CHECK1-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i32**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !102
// CHECK1-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !102
// CHECK1-NEXT: store %struct.anon.17* [[TMP8]], %struct.anon.17** [[__CONTEXT_ADDR_I]], align 8, !noalias !102
// CHECK1-NEXT: [[TMP12:%.*]] = load %struct.anon.17*, %struct.anon.17** [[__CONTEXT_ADDR_I]], align 8, !noalias !102
// CHECK1-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !102
// CHECK1-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !102
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i32**)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: call void [[TMP15]](i8* [[TMP14]], i32** [[DOTFIRSTPRIV_PTR_ADDR_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP16:%.*]] = load i32*, i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !102
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 4, i32* [[TMP16]], align 128
// CHECK1-NEXT: store i32 4, i32* @a, align 4
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@_ZN1SD1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK1-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK1-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SD2Ev(%struct.S* noundef [[THIS1]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_privates_map..20
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct..kmp_privates.t.20* noalias noundef [[TMP0:%.*]], i32** noalias noundef [[TMP1:%.*]], %struct.S** noalias noundef [[TMP2:%.*]], %struct.S*** noalias noundef [[TMP3:%.*]]) #[[ATTR7]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t.20*, align 8
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca i32**, align 8
// CHECK1-NEXT: [[DOTADDR2:%.*]] = alloca %struct.S**, align 8
// CHECK1-NEXT: [[DOTADDR3:%.*]] = alloca %struct.S***, align 8
// CHECK1-NEXT: store %struct..kmp_privates.t.20* [[TMP0]], %struct..kmp_privates.t.20** [[DOTADDR]], align 8
// CHECK1-NEXT: store i32** [[TMP1]], i32*** [[DOTADDR1]], align 8
// CHECK1-NEXT: store %struct.S** [[TMP2]], %struct.S*** [[DOTADDR2]], align 8
// CHECK1-NEXT: store %struct.S*** [[TMP3]], %struct.S**** [[DOTADDR3]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = load %struct..kmp_privates.t.20*, %struct..kmp_privates.t.20** [[DOTADDR]], align 8
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20:%.*]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP6:%.*]] = load i32**, i32*** [[DOTADDR1]], align 8
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP7:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP8:%.*]] = load %struct.S***, %struct.S**** [[DOTADDR3]], align 8
// CHECK1-NEXT: store %struct.S** [[TMP7]], %struct.S*** [[TMP8]], align 8
// CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 3
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.S**, %struct.S*** [[DOTADDR2]], align 8
// CHECK1-NEXT: store %struct.S* [[TMP9]], %struct.S** [[TMP10]], align 8
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..21
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.19* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.16*, align 8
// CHECK1-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTLOCAL_PTR_ADDR_I:%.*]] = alloca %struct.S*, align 8
// CHECK1-NEXT: [[DOTLOCAL_PTR_ADDR1_I:%.*]] = alloca %struct.S**, align 8
// CHECK1-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[REF_TMP_I:%.*]] = alloca [[STRUCT_S:%.*]], align 4
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.19*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.19* [[TMP1]], %struct.kmp_task_t_with_privates.19** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.19*, %struct.kmp_task_t_with_privates.19** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19:%.*]], %struct.kmp_task_t_with_privates.19* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 128
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.16*
// CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19]], %struct.kmp_task_t_with_privates.19* [[TMP3]], i32 0, i32 2
// CHECK1-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t.20* [[TMP9]] to i8*
// CHECK1-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.19* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META103:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META106:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META108:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META110:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
// CHECK1-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !112
// CHECK1-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t.20*, i32**, %struct.S**, %struct.S***)* @.omp_task_privates_map..20 to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !112
// CHECK1-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
// CHECK1-NEXT: store %struct.anon.16* [[TMP8]], %struct.anon.16** [[__CONTEXT_ADDR_I]], align 8, !noalias !112
// CHECK1-NEXT: [[TMP12:%.*]] = load %struct.anon.16*, %struct.anon.16** [[__CONTEXT_ADDR_I]], align 8, !noalias !112
// CHECK1-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !112
// CHECK1-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i32**, %struct.S**, %struct.S***)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: call void [[TMP15]](i8* [[TMP14]], i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], %struct.S** [[DOTLOCAL_PTR_ADDR_I]], %struct.S*** [[DOTLOCAL_PTR_ADDR1_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP16:%.*]] = load i32*, i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !112
// CHECK1-NEXT: [[TMP17:%.*]] = load %struct.S*, %struct.S** [[DOTLOCAL_PTR_ADDR_I]], align 8, !noalias !112
// CHECK1-NEXT: [[TMP18:%.*]] = load %struct.S**, %struct.S*** [[DOTLOCAL_PTR_ADDR1_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP19:%.*]] = load %struct.S*, %struct.S** [[TMP18]], align 8
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP20:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP21:%.*]] = load i32, i32* [[TMP20]], align 4
// CHECK1-NEXT: switch i32 [[TMP21]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK1-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK1-NEXT: i32 1, label [[DOTUNTIED_JMP_2_I:%.*]]
// CHECK1-NEXT: i32 2, label [[DOTUNTIED_JMP_3_I:%.*]]
// CHECK1-NEXT: i32 3, label [[DOTUNTIED_JMP_5_I:%.*]]
// CHECK1-NEXT: i32 4, label [[DOTUNTIED_JMP_7_I:%.*]]
// CHECK1-NEXT: i32 5, label [[DOTUNTIED_JMP_10_I:%.*]]
// CHECK1-NEXT: ]
// CHECK1: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK1: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP22:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 1, i32* [[TMP22]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP23:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK1-NEXT: [[TMP24:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP25:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP23]], i8* [[TMP24]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT:%.*]]
// CHECK1: .untied.jmp.2.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[TMP17]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP26:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[DOTS2__VOID_ADDR_I:%.*]] = call i8* @__kmpc_alloc(i32 [[TMP26]], i64 4, i8* inttoptr (i64 7 to i8*)) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[DOTS2__ADDR_I:%.*]] = bitcast i8* [[DOTS2__VOID_ADDR_I]] to %struct.S*
// CHECK1-NEXT: store %struct.S* [[DOTS2__ADDR_I]], %struct.S** [[TMP18]], align 8
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP27:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 2, i32* [[TMP27]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP28:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK1-NEXT: [[TMP29:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP30:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP28]], i8* [[TMP29]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK1: .untied.jmp.3.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[TMP19]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[TMP19]], i32 0, i32 0
// CHECK1-NEXT: store i32 0, i32* [[A_I]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP31:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP32:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP31]], i32 1, i64 256, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.18*)* @.omp_task_entry..19 to i32 (i32, i8*)*)) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP33:%.*]] = bitcast i8* [[TMP32]] to %struct.kmp_task_t_with_privates.18*
// CHECK1-NEXT: [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18:%.*]], %struct.kmp_task_t_with_privates.18* [[TMP33]], i32 0, i32 0
// CHECK1-NEXT: [[TMP35:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18]], %struct.kmp_task_t_with_privates.18* [[TMP33]], i32 0, i32 2
// CHECK1-NEXT: [[TMP36:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP35]], i32 0, i32 0
// CHECK1-NEXT: [[TMP37:%.*]] = load i32, i32* [[TMP16]], align 128
// CHECK1-NEXT: store i32 [[TMP37]], i32* [[TMP36]], align 128
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP38:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP39:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP38]], i8* [[TMP32]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP40:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 3, i32* [[TMP40]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP41:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK1-NEXT: [[TMP42:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP43:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP41]], i8* [[TMP42]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK1: .untied.jmp.5.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP44:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP45:%.*]] = call i32 @__kmpc_omp_taskyield(%struct.ident_t* @[[GLOB1]], i32 [[TMP44]], i32 0) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP46:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 4, i32* [[TMP46]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP47:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK1-NEXT: [[TMP48:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP49:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP47]], i8* [[TMP48]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK1: .untied.jmp.7.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[REF_TMP_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP50:%.*]] = bitcast %struct.S* [[TMP17]] to i8*
// CHECK1-NEXT: [[TMP51:%.*]] = bitcast %struct.S* [[REF_TMP_I]] to i8*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP50]], i8* align 4 [[TMP51]], i64 4, i1 false) #[[ATTR4]]
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[REF_TMP_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[A9_I:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[TMP19]], i32 0, i32 0
// CHECK1-NEXT: store i32 10, i32* [[A9_I]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP52:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP53:%.*]] = call i32 @__kmpc_omp_taskwait(%struct.ident_t* @[[GLOB1]], i32 [[TMP52]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP54:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 5, i32* [[TMP54]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP55:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK1-NEXT: [[TMP56:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: [[TMP57:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP55]], i8* [[TMP56]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK1: .untied.jmp.10.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[TMP19]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP58:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP59:%.*]] = bitcast %struct.S* [[TMP19]] to i8*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: call void @__kmpc_free(i32 [[TMP58]], i8* [[TMP59]], i8* inttoptr (i64 7 to i8*)) #[[ATTR4]]
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[TMP17]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[CLEANUP_I]]
// CHECK1: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK1: .omp_outlined..17.exit:
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@_ZN1SC2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK1-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK1-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
// CHECK1-NEXT: [[A:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[THIS1]], i32 0, i32 0
// CHECK1-NEXT: store i32 0, i32* [[A]], align 4
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@_ZN1SD2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK1-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK1-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
// CHECK1-NEXT: ret void
//
//
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-LABEL: define {{[^@]+}}@__cxx_global_var_init
// CHECK1-SAME: () #[[ATTR7]] section "__TEXT,__StaticInit,regular,pure_instructions" {
// CHECK1-NEXT: entry:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN2S1C1Ev(%struct.S1* noundef @s1)
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@_ZN2S1C1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct.S1* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK1-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK1-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN2S1C2Ev(%struct.S1* noundef [[THIS1]])
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: ret void
//
//
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-LABEL: define {{[^@]+}}@_ZN2S1C2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct.S1* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK1-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK1-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-NEXT: call void @_ZN2S18taskinitEv(%struct.S1* noundef [[THIS1]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@_ZN2S18taskinitEv
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct.S1* noundef [[THIS:%.*]]) #[[ATTR8:[0-9]+]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK1-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON_21:%.*]], align 8
// CHECK1-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
// CHECK1-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK1-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
// CHECK1-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_21]], %struct.anon.21* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK1-NEXT: store %struct.S1* [[THIS1]], %struct.S1** [[TMP1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.22*)* @.omp_task_entry..23 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP3:%.*]] = bitcast i8* [[TMP2]] to %struct.kmp_task_t_with_privates.22*
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_22:%.*]], %struct.kmp_task_t_with_privates.22* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP6:%.*]] = load i8*, i8** [[TMP5]], align 8
// CHECK1-NEXT: [[TMP7:%.*]] = bitcast %struct.anon.21* [[AGG_CAPTURED]] to i8*
// CHECK1-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP6]], i8* align 8 [[TMP7]], i64 8, i1 false)
// CHECK1-NEXT: [[TMP8:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP2]])
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..23
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.22* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.21*, align 8
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.22*, align 8
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.22* [[TMP1]], %struct.kmp_task_t_with_privates.22** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.22*, %struct.kmp_task_t_with_privates.22** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_22:%.*]], %struct.kmp_task_t_with_privates.22* [[TMP3]], i32 0, i32 0
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.21*
// CHECK1-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.22* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META113:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META116:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META118:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META120:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !122
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !122
// CHECK1-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !122
// CHECK1-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !122
// CHECK1-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !122
// CHECK1-NEXT: store %struct.anon.21* [[TMP8]], %struct.anon.21** [[__CONTEXT_ADDR_I]], align 8, !noalias !122
// CHECK1-NEXT: [[TMP10:%.*]] = load %struct.anon.21*, %struct.anon.21** [[__CONTEXT_ADDR_I]], align 8, !noalias !122
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_21:%.*]], %struct.anon.21* [[TMP10]], i32 0, i32 0
// CHECK1-NEXT: [[TMP12:%.*]] = load %struct.S1*, %struct.S1** [[TMP11]], align 8
// CHECK1-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP12]], i32 0, i32 0
// CHECK1-NEXT: store i32 0, i32* [[A_I]], align 4
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@_Z4xxxxv
// CHECK1-SAME: () #[[ATTR8]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: call void @_Z6foobarIvEvv()
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@_Z6foobarIvEvv
// CHECK1-SAME: () #[[ATTR8]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[A:%.*]] = alloca float, align 4
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, float*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), float* [[A]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_outlined..24
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32* noalias noundef [[DOTGLOBAL_TID_:%.*]], i32* noalias noundef [[DOTBOUND_TID_:%.*]], float* noundef nonnull align 4 dereferenceable(4) [[A:%.*]]) #[[ATTR9:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[A_ADDR:%.*]] = alloca float*, align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[B:%.*]] = alloca double, align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON_23:%.*]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK1-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: store float* [[A]], float** [[A_ADDR]], align 8
// CHECK1-NEXT: [[TMP0:%.*]] = load float*, float** [[A_ADDR]], align 8
// CHECK1-NEXT: [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
// CHECK1-NEXT: [[TMP3:%.*]] = call i32 @__kmpc_single(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
// CHECK1-NEXT: [[TMP4:%.*]] = icmp ne i32 [[TMP3]], 0
// CHECK1-NEXT: br i1 [[TMP4]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_END:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1: omp_if.then:
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_ANON_23]], %struct.anon.23* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK1-NEXT: store float* [[TMP0]], float** [[TMP5]], align 8
// CHECK1-NEXT: [[TMP6:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 1, i64 48, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.24*)* @.omp_task_entry..27 to i32 (i32, i8*)*))
// CHECK1-NEXT: [[TMP7:%.*]] = bitcast i8* [[TMP6]] to %struct.kmp_task_t_with_privates.24*
// CHECK1-NEXT: [[TMP8:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_24:%.*]], %struct.kmp_task_t_with_privates.24* [[TMP7]], i32 0, i32 0
// CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP8]], i32 0, i32 0
// CHECK1-NEXT: [[TMP10:%.*]] = load i8*, i8** [[TMP9]], align 8
// CHECK1-NEXT: [[TMP11:%.*]] = bitcast %struct.anon.23* [[AGG_CAPTURED]] to i8*
// CHECK1-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP10]], i8* align 8 [[TMP11]], i64 8, i1 false)
// CHECK1-NEXT: [[TMP12:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_24]], %struct.kmp_task_t_with_privates.24* [[TMP7]], i32 0, i32 1
// CHECK1-NEXT: [[TMP13:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_25:%.*]], %struct..kmp_privates.t.25* [[TMP12]], i32 0, i32 0
// CHECK1-NEXT: [[TMP14:%.*]] = load double, double* [[B]], align 8
// CHECK1-NEXT: store double [[TMP14]], double* [[TMP13]], align 8
// CHECK1-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i8* [[TMP6]])
// CHECK1-NEXT: call void @__kmpc_end_single(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: br label [[OMP_IF_END]]
// CHECK1: omp_if.end:
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP2]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_privates_map..26
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (%struct..kmp_privates.t.25* noalias noundef [[TMP0:%.*]], double** noalias noundef [[TMP1:%.*]]) #[[ATTR7]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t.25*, align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca double**, align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: store %struct..kmp_privates.t.25* [[TMP0]], %struct..kmp_privates.t.25** [[DOTADDR]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store double** [[TMP1]], double*** [[DOTADDR1]], align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[TMP2:%.*]] = load %struct..kmp_privates.t.25*, %struct..kmp_privates.t.25** [[DOTADDR]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_25:%.*]], %struct..kmp_privates.t.25* [[TMP2]], i32 0, i32 0
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP4:%.*]] = load double**, double*** [[DOTADDR1]], align 8
// CHECK1-NEXT: store double* [[TMP3]], double** [[TMP4]], align 8
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_task_entry..27
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK1-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.24* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK1-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK1-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.23*, align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca double*, align 8
// CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.24*, align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: store %struct.kmp_task_t_with_privates.24* [[TMP1]], %struct.kmp_task_t_with_privates.24** [[DOTADDR1]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.24*, %struct.kmp_task_t_with_privates.24** [[DOTADDR1]], align 8
// CHECK1-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_24:%.*]], %struct.kmp_task_t_with_privates.24* [[TMP3]], i32 0, i32 0
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK1-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.23*
// CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_24]], %struct.kmp_task_t_with_privates.24* [[TMP3]], i32 0, i32 1
// CHECK1-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t.25* [[TMP9]] to i8*
// CHECK1-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.24* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META123:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META126:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META128:![0-9]+]])
// CHECK1-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META130:![0-9]+]])
// CHECK1-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !132
// CHECK1-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !132
// CHECK1-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !132
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t.25*, double**)* @.omp_task_privates_map..26 to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !132
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !132
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: store %struct.anon.23* [[TMP8]], %struct.anon.23** [[__CONTEXT_ADDR_I]], align 8, !noalias !132
// CHECK1-NEXT: [[TMP12:%.*]] = load %struct.anon.23*, %struct.anon.23** [[__CONTEXT_ADDR_I]], align 8, !noalias !132
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !132
// CHECK1-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !132
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, double**)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-NEXT: call void [[TMP15]](i8* [[TMP14]], double** [[DOTFIRSTPRIV_PTR_ADDR_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK1-NEXT: [[TMP16:%.*]] = load double*, double** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !132
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP17:%.*]] = load double, double* [[TMP16]], align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK1-NEXT: [[TMP18:%.*]] = getelementptr inbounds [[STRUCT_ANON_23:%.*]], %struct.anon.23* [[TMP12]], i32 0, i32 0
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: [[TMP19:%.*]] = load float*, float** [[TMP18]], align 8
// CHECK1-NEXT: [[TMP20:%.*]] = load float, float* [[TMP19]], align 4
// CHECK1-NEXT: [[CONV_I:%.*]] = fpext float [[TMP20]] to double
// CHECK1-NEXT: [[ADD_I:%.*]] = fadd double [[CONV_I]], [[TMP17]]
// CHECK1-NEXT: [[CONV1_I:%.*]] = fptrunc double [[ADD_I]] to float
// CHECK1-NEXT: store float [[CONV1_I]], float* [[TMP19]], align 4
// CHECK1-NEXT: ret i32 0
//
//
// CHECK1-LABEL: define {{[^@]+}}@_GLOBAL__sub_I_task_codegen.cpp
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK1-SAME: () #[[ATTR7]] section "__TEXT,__StaticInit,regular,pure_instructions" {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: call void @__cxx_global_var_init()
// CHECK1-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@main
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-SAME: () #[[ATTR0:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[RETVAL:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[B:%.*]] = alloca i8, align 1
// CHECK2-NEXT: [[S:%.*]] = alloca [2 x %struct.S], align 4
// CHECK2-NEXT: [[SAVED_STACK:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__VLA_EXPR0:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 8
// CHECK2-NEXT: [[AGG_CAPTURED1:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 8
// CHECK2-NEXT: [[DOTDEP_ARR_ADDR:%.*]] = alloca [4 x %struct.kmp_depend_info], align 8
// CHECK2-NEXT: [[DEP_COUNTER_ADDR:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[AGG_CAPTURED3:%.*]] = alloca [[STRUCT_ANON_2:%.*]], align 1
// CHECK2-NEXT: [[AGG_CAPTURED4:%.*]] = alloca [[STRUCT_ANON_4:%.*]], align 1
// CHECK2-NEXT: [[DOTDEP_ARR_ADDR5:%.*]] = alloca [2 x %struct.kmp_depend_info], align 8
// CHECK2-NEXT: [[DEP_COUNTER_ADDR11:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[AGG_CAPTURED12:%.*]] = alloca [[STRUCT_ANON_6:%.*]], align 1
// CHECK2-NEXT: [[DOTDEP_ARR_ADDR13:%.*]] = alloca [2 x %struct.kmp_depend_info], align 8
// CHECK2-NEXT: [[DEP_COUNTER_ADDR19:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[AGG_CAPTURED20:%.*]] = alloca [[STRUCT_ANON_8:%.*]], align 1
// CHECK2-NEXT: [[DOTDEP_ARR_ADDR21:%.*]] = alloca [3 x %struct.kmp_depend_info], align 8
// CHECK2-NEXT: [[DEP_COUNTER_ADDR27:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[AGG_CAPTURED28:%.*]] = alloca [[STRUCT_ANON_10:%.*]], align 1
// CHECK2-NEXT: [[FLAG:%.*]] = alloca i8, align 1
// CHECK2-NEXT: [[AGG_CAPTURED29:%.*]] = alloca [[STRUCT_ANON_12:%.*]], align 1
// CHECK2-NEXT: [[C:%.*]] = alloca i32, align 128
// CHECK2-NEXT: [[AGG_CAPTURED30:%.*]] = alloca [[STRUCT_ANON_14:%.*]], align 8
// CHECK2-NEXT: [[AGG_CAPTURED31:%.*]] = alloca [[STRUCT_ANON_16:%.*]], align 1
// CHECK2-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]])
// CHECK2-NEXT: store i32 0, i32* [[RETVAL]], align 4
// CHECK2-NEXT: [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i32 0, i32 0
// CHECK2-NEXT: [[ARRAYCTOR_END:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAY_BEGIN]], i64 2
// CHECK2-NEXT: br label [[ARRAYCTOR_LOOP:%.*]]
// CHECK2: arrayctor.loop:
// CHECK2-NEXT: [[ARRAYCTOR_CUR:%.*]] = phi %struct.S* [ [[ARRAY_BEGIN]], [[ENTRY:%.*]] ], [ [[ARRAYCTOR_NEXT:%.*]], [[ARRAYCTOR_LOOP]] ]
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[ARRAYCTOR_CUR]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[ARRAYCTOR_NEXT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYCTOR_CUR]], i64 1
// CHECK2-NEXT: [[ARRAYCTOR_DONE:%.*]] = icmp eq %struct.S* [[ARRAYCTOR_NEXT]], [[ARRAYCTOR_END]]
// CHECK2-NEXT: br i1 [[ARRAYCTOR_DONE]], label [[ARRAYCTOR_CONT:%.*]], label [[ARRAYCTOR_LOOP]]
// CHECK2: arrayctor.cont:
// CHECK2-NEXT: [[TMP1:%.*]] = load i32, i32* @a, align 4
// CHECK2-NEXT: [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
// CHECK2-NEXT: [[TMP3:%.*]] = call i8* @llvm.stacksave()
// CHECK2-NEXT: store i8* [[TMP3]], i8** [[SAVED_STACK]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = mul nuw i64 10, [[TMP2]]
// CHECK2-NEXT: [[VLA:%.*]] = alloca i32, i64 [[TMP4]], align 16
// CHECK2-NEXT: store i64 [[TMP2]], i64* [[__VLA_EXPR0]], align 8
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK2-NEXT: store i8* [[B]], i8** [[TMP5]], align 8
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
// CHECK2-NEXT: store [2 x %struct.S]* [[S]], [2 x %struct.S]** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP7:%.*]] = load i8, i8* [[B]], align 1
// CHECK2-NEXT: [[CONV:%.*]] = sext i8 [[TMP7]] to i32
// CHECK2-NEXT: [[TMP8:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 33, i64 40, i64 16, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast i8* [[TMP8]] to %struct.kmp_task_t_with_privates*
// CHECK2-NEXT: [[TMP10:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP9]], i32 0, i32 0
// CHECK2-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP10]], i32 0, i32 0
// CHECK2-NEXT: [[TMP12:%.*]] = load i8*, i8** [[TMP11]], align 8
// CHECK2-NEXT: [[TMP13:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
// CHECK2-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP12]], i8* align 8 [[TMP13]], i64 16, i1 false)
// CHECK2-NEXT: [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP10]], i32 0, i32 4
// CHECK2-NEXT: [[TMP15:%.*]] = bitcast %union.kmp_cmplrdata_t* [[TMP14]] to i32*
// CHECK2-NEXT: store i32 [[CONV]], i32* [[TMP15]], align 8
// CHECK2-NEXT: [[TMP16:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP8]])
// CHECK2-NEXT: [[TMP17:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[AGG_CAPTURED1]], i32 0, i32 0
// CHECK2-NEXT: store [2 x %struct.S]* [[S]], [2 x %struct.S]** [[TMP17]], align 8
// CHECK2-NEXT: [[TMP18:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.1*)* @.omp_task_entry..2 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP19:%.*]] = bitcast i8* [[TMP18]] to %struct.kmp_task_t_with_privates.1*
// CHECK2-NEXT: [[TMP20:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_1:%.*]], %struct.kmp_task_t_with_privates.1* [[TMP19]], i32 0, i32 0
// CHECK2-NEXT: [[TMP21:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP20]], i32 0, i32 0
// CHECK2-NEXT: [[TMP22:%.*]] = load i8*, i8** [[TMP21]], align 8
// CHECK2-NEXT: [[TMP23:%.*]] = bitcast %struct.anon.0* [[AGG_CAPTURED1]] to i8*
// CHECK2-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP22]], i8* align 8 [[TMP23]], i64 8, i1 false)
// CHECK2-NEXT: [[TMP24:%.*]] = getelementptr inbounds [4 x %struct.kmp_depend_info], [4 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR]], i64 0, i64 0
// CHECK2-NEXT: [[TMP25:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO:%.*]], %struct.kmp_depend_info* [[TMP24]], i64 0
// CHECK2-NEXT: [[TMP26:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP25]], i32 0, i32 0
// CHECK2-NEXT: store i64 ptrtoint (i32* @a to i64), i64* [[TMP26]], align 8
// CHECK2-NEXT: [[TMP27:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP25]], i32 0, i32 1
// CHECK2-NEXT: store i64 4, i64* [[TMP27]], align 8
// CHECK2-NEXT: [[TMP28:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP25]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 1, i8* [[TMP28]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP29:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i64 1
// CHECK2-NEXT: [[TMP30:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP29]], i32 0, i32 0
// CHECK2-NEXT: [[TMP31:%.*]] = ptrtoint i8* [[B]] to i64
// CHECK2-NEXT: store i64 [[TMP31]], i64* [[TMP30]], align 8
// CHECK2-NEXT: [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP29]], i32 0, i32 1
// CHECK2-NEXT: store i64 1, i64* [[TMP32]], align 8
// CHECK2-NEXT: [[TMP33:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP29]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 1, i8* [[TMP33]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP34:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i64 2
// CHECK2-NEXT: [[TMP35:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP34]], i32 0, i32 0
// CHECK2-NEXT: [[TMP36:%.*]] = ptrtoint [2 x %struct.S]* [[S]] to i64
// CHECK2-NEXT: store i64 [[TMP36]], i64* [[TMP35]], align 8
// CHECK2-NEXT: [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP34]], i32 0, i32 1
// CHECK2-NEXT: store i64 8, i64* [[TMP37]], align 8
// CHECK2-NEXT: [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP34]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 1, i8* [[TMP38]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP39:%.*]] = mul nsw i64 0, [[TMP2]]
// CHECK2-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP39]]
// CHECK2-NEXT: [[TMP40:%.*]] = mul nsw i64 9, [[TMP2]]
// CHECK2-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP40]]
// CHECK2-NEXT: [[TMP41:%.*]] = getelementptr i32, i32* [[ARRAYIDX2]], i32 1
// CHECK2-NEXT: [[TMP42:%.*]] = ptrtoint i32* [[ARRAYIDX]] to i64
// CHECK2-NEXT: [[TMP43:%.*]] = ptrtoint i32* [[TMP41]] to i64
// CHECK2-NEXT: [[TMP44:%.*]] = sub nuw i64 [[TMP43]], [[TMP42]]
// CHECK2-NEXT: [[TMP45:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i64 3
// CHECK2-NEXT: [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP45]], i32 0, i32 0
// CHECK2-NEXT: [[TMP47:%.*]] = ptrtoint i32* [[ARRAYIDX]] to i64
// CHECK2-NEXT: store i64 [[TMP47]], i64* [[TMP46]], align 8
// CHECK2-NEXT: [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP45]], i32 0, i32 1
// CHECK2-NEXT: store i64 [[TMP44]], i64* [[TMP48]], align 8
// CHECK2-NEXT: [[TMP49:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP45]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 1, i8* [[TMP49]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i64 4, i64* [[DEP_COUNTER_ADDR]], align 8
// CHECK2-NEXT: [[TMP50:%.*]] = bitcast %struct.kmp_depend_info* [[TMP24]] to i8*
// CHECK2-NEXT: [[TMP51:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP18]], i32 4, i8* [[TMP50]], i32 0, i8* null)
// CHECK2-NEXT: [[TMP52:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.3*)* @.omp_task_entry..4 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP53:%.*]] = bitcast i8* [[TMP52]] to %struct.kmp_task_t_with_privates.3*
// CHECK2-NEXT: [[TMP54:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_3:%.*]], %struct.kmp_task_t_with_privates.3* [[TMP53]], i32 0, i32 0
// CHECK2-NEXT: [[TMP55:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP54]], i32 0, i32 2
// CHECK2-NEXT: store i32 0, i32* [[TMP55]], align 8
// CHECK2-NEXT: [[TMP56:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP52]])
// CHECK2-NEXT: [[TMP57:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.5*)* @.omp_task_entry..6 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP58:%.*]] = bitcast i8* [[TMP57]] to %struct.kmp_task_t_with_privates.5*
// CHECK2-NEXT: [[TMP59:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_5:%.*]], %struct.kmp_task_t_with_privates.5* [[TMP58]], i32 0, i32 0
// CHECK2-NEXT: [[TMP60:%.*]] = getelementptr inbounds [2 x %struct.kmp_depend_info], [2 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR5]], i64 0, i64 0
// CHECK2-NEXT: [[ARRAYIDX6:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 0
// CHECK2-NEXT: [[TMP61:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i64 0
// CHECK2-NEXT: [[TMP62:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP61]], i32 0, i32 0
// CHECK2-NEXT: [[TMP63:%.*]] = ptrtoint %struct.S* [[ARRAYIDX6]] to i64
// CHECK2-NEXT: store i64 [[TMP63]], i64* [[TMP62]], align 8
// CHECK2-NEXT: [[TMP64:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP61]], i32 0, i32 1
// CHECK2-NEXT: store i64 4, i64* [[TMP64]], align 8
// CHECK2-NEXT: [[TMP65:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP61]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 3, i8* [[TMP65]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP66:%.*]] = load i8, i8* [[B]], align 1
// CHECK2-NEXT: [[TMP67:%.*]] = sext i8 [[TMP66]] to i64
// CHECK2-NEXT: [[TMP68:%.*]] = mul nsw i64 4, [[TMP2]]
// CHECK2-NEXT: [[ARRAYIDX7:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP68]]
// CHECK2-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX7]], i64 [[TMP67]]
// CHECK2-NEXT: [[TMP69:%.*]] = load i8, i8* [[B]], align 1
// CHECK2-NEXT: [[TMP70:%.*]] = sext i8 [[TMP69]] to i64
// CHECK2-NEXT: [[TMP71:%.*]] = mul nsw i64 9, [[TMP2]]
// CHECK2-NEXT: [[ARRAYIDX9:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP71]]
// CHECK2-NEXT: [[ARRAYIDX10:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX9]], i64 [[TMP70]]
// CHECK2-NEXT: [[TMP72:%.*]] = getelementptr i32, i32* [[ARRAYIDX10]], i32 1
// CHECK2-NEXT: [[TMP73:%.*]] = ptrtoint i32* [[ARRAYIDX8]] to i64
// CHECK2-NEXT: [[TMP74:%.*]] = ptrtoint i32* [[TMP72]] to i64
// CHECK2-NEXT: [[TMP75:%.*]] = sub nuw i64 [[TMP74]], [[TMP73]]
// CHECK2-NEXT: [[TMP76:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i64 1
// CHECK2-NEXT: [[TMP77:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP76]], i32 0, i32 0
// CHECK2-NEXT: [[TMP78:%.*]] = ptrtoint i32* [[ARRAYIDX8]] to i64
// CHECK2-NEXT: store i64 [[TMP78]], i64* [[TMP77]], align 8
// CHECK2-NEXT: [[TMP79:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP76]], i32 0, i32 1
// CHECK2-NEXT: store i64 [[TMP75]], i64* [[TMP79]], align 8
// CHECK2-NEXT: [[TMP80:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP76]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 3, i8* [[TMP80]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i64 2, i64* [[DEP_COUNTER_ADDR11]], align 8
// CHECK2-NEXT: [[TMP81:%.*]] = bitcast %struct.kmp_depend_info* [[TMP60]] to i8*
// CHECK2-NEXT: [[TMP82:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP59]], i32 0, i32 2
// CHECK2-NEXT: store i32 0, i32* [[TMP82]], align 8
// CHECK2-NEXT: [[TMP83:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP57]], i32 2, i8* [[TMP81]], i32 0, i8* null)
// CHECK2-NEXT: [[TMP84:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.7*)* @.omp_task_entry..8 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP85:%.*]] = bitcast i8* [[TMP84]] to %struct.kmp_task_t_with_privates.7*
// CHECK2-NEXT: [[TMP86:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_7:%.*]], %struct.kmp_task_t_with_privates.7* [[TMP85]], i32 0, i32 0
// CHECK2-NEXT: [[TMP87:%.*]] = getelementptr inbounds [2 x %struct.kmp_depend_info], [2 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR13]], i64 0, i64 0
// CHECK2-NEXT: [[ARRAYIDX14:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 0
// CHECK2-NEXT: [[TMP88:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i64 0
// CHECK2-NEXT: [[TMP89:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP88]], i32 0, i32 0
// CHECK2-NEXT: [[TMP90:%.*]] = ptrtoint %struct.S* [[ARRAYIDX14]] to i64
// CHECK2-NEXT: store i64 [[TMP90]], i64* [[TMP89]], align 8
// CHECK2-NEXT: [[TMP91:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP88]], i32 0, i32 1
// CHECK2-NEXT: store i64 4, i64* [[TMP91]], align 8
// CHECK2-NEXT: [[TMP92:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP88]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 4, i8* [[TMP92]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP93:%.*]] = load i8, i8* [[B]], align 1
// CHECK2-NEXT: [[TMP94:%.*]] = sext i8 [[TMP93]] to i64
// CHECK2-NEXT: [[TMP95:%.*]] = mul nsw i64 4, [[TMP2]]
// CHECK2-NEXT: [[ARRAYIDX15:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP95]]
// CHECK2-NEXT: [[ARRAYIDX16:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX15]], i64 [[TMP94]]
// CHECK2-NEXT: [[TMP96:%.*]] = load i8, i8* [[B]], align 1
// CHECK2-NEXT: [[TMP97:%.*]] = sext i8 [[TMP96]] to i64
// CHECK2-NEXT: [[TMP98:%.*]] = mul nsw i64 9, [[TMP2]]
// CHECK2-NEXT: [[ARRAYIDX17:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP98]]
// CHECK2-NEXT: [[ARRAYIDX18:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX17]], i64 [[TMP97]]
// CHECK2-NEXT: [[TMP99:%.*]] = getelementptr i32, i32* [[ARRAYIDX18]], i32 1
// CHECK2-NEXT: [[TMP100:%.*]] = ptrtoint i32* [[ARRAYIDX16]] to i64
// CHECK2-NEXT: [[TMP101:%.*]] = ptrtoint i32* [[TMP99]] to i64
// CHECK2-NEXT: [[TMP102:%.*]] = sub nuw i64 [[TMP101]], [[TMP100]]
// CHECK2-NEXT: [[TMP103:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i64 1
// CHECK2-NEXT: [[TMP104:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP103]], i32 0, i32 0
// CHECK2-NEXT: [[TMP105:%.*]] = ptrtoint i32* [[ARRAYIDX16]] to i64
// CHECK2-NEXT: store i64 [[TMP105]], i64* [[TMP104]], align 8
// CHECK2-NEXT: [[TMP106:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP103]], i32 0, i32 1
// CHECK2-NEXT: store i64 [[TMP102]], i64* [[TMP106]], align 8
// CHECK2-NEXT: [[TMP107:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP103]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 4, i8* [[TMP107]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i64 2, i64* [[DEP_COUNTER_ADDR19]], align 8
// CHECK2-NEXT: [[TMP108:%.*]] = bitcast %struct.kmp_depend_info* [[TMP87]] to i8*
// CHECK2-NEXT: [[TMP109:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP86]], i32 0, i32 2
// CHECK2-NEXT: store i32 0, i32* [[TMP109]], align 8
// CHECK2-NEXT: [[TMP110:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP84]], i32 2, i8* [[TMP108]], i32 0, i8* null)
// CHECK2-NEXT: [[TMP111:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 3, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.9*)* @.omp_task_entry..10 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP112:%.*]] = bitcast i8* [[TMP111]] to %struct.kmp_task_t_with_privates.9*
// CHECK2-NEXT: [[TMP113:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_9:%.*]], %struct.kmp_task_t_with_privates.9* [[TMP112]], i32 0, i32 0
// CHECK2-NEXT: [[TMP114:%.*]] = getelementptr inbounds [3 x %struct.kmp_depend_info], [3 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR21]], i64 0, i64 0
// CHECK2-NEXT: [[TMP115:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i64 0
// CHECK2-NEXT: [[TMP116:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP115]], i32 0, i32 0
// CHECK2-NEXT: store i64 ptrtoint (i32* @a to i64), i64* [[TMP116]], align 8
// CHECK2-NEXT: [[TMP117:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP115]], i32 0, i32 1
// CHECK2-NEXT: store i64 4, i64* [[TMP117]], align 8
// CHECK2-NEXT: [[TMP118:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP115]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 3, i8* [[TMP118]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[ARRAYIDX22:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 1
// CHECK2-NEXT: [[TMP119:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i64 1
// CHECK2-NEXT: [[TMP120:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP119]], i32 0, i32 0
// CHECK2-NEXT: [[TMP121:%.*]] = ptrtoint %struct.S* [[ARRAYIDX22]] to i64
// CHECK2-NEXT: store i64 [[TMP121]], i64* [[TMP120]], align 8
// CHECK2-NEXT: [[TMP122:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP119]], i32 0, i32 1
// CHECK2-NEXT: store i64 4, i64* [[TMP122]], align 8
// CHECK2-NEXT: [[TMP123:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP119]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 3, i8* [[TMP123]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP124:%.*]] = mul nsw i64 0, [[TMP2]]
// CHECK2-NEXT: [[ARRAYIDX23:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP124]]
// CHECK2-NEXT: [[ARRAYIDX24:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX23]], i64 3
// CHECK2-NEXT: [[TMP125:%.*]] = load i32, i32* @a, align 4
// CHECK2-NEXT: [[TMP126:%.*]] = sext i32 [[TMP125]] to i64
// CHECK2-NEXT: [[LEN_SUB_1:%.*]] = sub nsw i64 [[TMP126]], 1
// CHECK2-NEXT: [[TMP127:%.*]] = load i32, i32* @a, align 4
// CHECK2-NEXT: [[TMP128:%.*]] = sext i32 [[TMP127]] to i64
// CHECK2-NEXT: [[LB_ADD_LEN:%.*]] = add nsw i64 -1, [[TMP128]]
// CHECK2-NEXT: [[TMP129:%.*]] = mul nsw i64 [[LB_ADD_LEN]], [[TMP2]]
// CHECK2-NEXT: [[ARRAYIDX25:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP129]]
// CHECK2-NEXT: [[ARRAYIDX26:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX25]], i64 [[LEN_SUB_1]]
// CHECK2-NEXT: [[TMP130:%.*]] = getelementptr i32, i32* [[ARRAYIDX26]], i32 1
// CHECK2-NEXT: [[TMP131:%.*]] = ptrtoint i32* [[ARRAYIDX24]] to i64
// CHECK2-NEXT: [[TMP132:%.*]] = ptrtoint i32* [[TMP130]] to i64
// CHECK2-NEXT: [[TMP133:%.*]] = sub nuw i64 [[TMP132]], [[TMP131]]
// CHECK2-NEXT: [[TMP134:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i64 2
// CHECK2-NEXT: [[TMP135:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP134]], i32 0, i32 0
// CHECK2-NEXT: [[TMP136:%.*]] = ptrtoint i32* [[ARRAYIDX24]] to i64
// CHECK2-NEXT: store i64 [[TMP136]], i64* [[TMP135]], align 8
// CHECK2-NEXT: [[TMP137:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP134]], i32 0, i32 1
// CHECK2-NEXT: store i64 [[TMP133]], i64* [[TMP137]], align 8
// CHECK2-NEXT: [[TMP138:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP134]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK2-NEXT: store i8 3, i8* [[TMP138]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i64 3, i64* [[DEP_COUNTER_ADDR27]], align 8
// CHECK2-NEXT: [[TMP139:%.*]] = bitcast %struct.kmp_depend_info* [[TMP114]] to i8*
// CHECK2-NEXT: [[TMP140:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP111]], i32 3, i8* [[TMP139]], i32 0, i8* null)
// CHECK2-NEXT: [[TMP141:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 3, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.11*)* @.omp_task_entry..12 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP142:%.*]] = bitcast i8* [[TMP141]] to %struct.kmp_task_t_with_privates.11*
// CHECK2-NEXT: [[TMP143:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_11:%.*]], %struct.kmp_task_t_with_privates.11* [[TMP142]], i32 0, i32 0
// CHECK2-NEXT: [[TMP144:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP141]])
// CHECK2-NEXT: store i8 0, i8* [[FLAG]], align 1
// CHECK2-NEXT: [[TMP145:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.13*)* @.omp_task_entry..14 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP146:%.*]] = bitcast i8* [[TMP145]] to %struct.kmp_task_t_with_privates.13*
// CHECK2-NEXT: [[TMP147:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_13:%.*]], %struct.kmp_task_t_with_privates.13* [[TMP146]], i32 0, i32 0
// CHECK2-NEXT: [[TMP148:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP145]])
// CHECK2-NEXT: [[TMP149:%.*]] = getelementptr inbounds [[STRUCT_ANON_14]], %struct.anon.14* [[AGG_CAPTURED30]], i32 0, i32 0
// CHECK2-NEXT: store i32* [[C]], i32** [[TMP149]], align 8
// CHECK2-NEXT: [[TMP150:%.*]] = load i8, i8* [[B]], align 1
// CHECK2-NEXT: [[TOBOOL:%.*]] = icmp ne i8 [[TMP150]], 0
// CHECK2-NEXT: [[TMP151:%.*]] = select i1 [[TOBOOL]], i32 2, i32 0
// CHECK2-NEXT: [[TMP152:%.*]] = or i32 [[TMP151]], 1
// CHECK2-NEXT: [[TMP153:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 [[TMP152]], i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.15*)* @.omp_task_entry..16 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP154:%.*]] = bitcast i8* [[TMP153]] to %struct.kmp_task_t_with_privates.15*
// CHECK2-NEXT: [[TMP155:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_15:%.*]], %struct.kmp_task_t_with_privates.15* [[TMP154]], i32 0, i32 0
// CHECK2-NEXT: [[TMP156:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP155]], i32 0, i32 0
// CHECK2-NEXT: [[TMP157:%.*]] = load i8*, i8** [[TMP156]], align 8
// CHECK2-NEXT: [[TMP158:%.*]] = bitcast %struct.anon.14* [[AGG_CAPTURED30]] to i8*
// CHECK2-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP157]], i8* align 8 [[TMP158]], i64 8, i1 false)
// CHECK2-NEXT: [[TMP159:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP153]])
// CHECK2-NEXT: [[TMP160:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 0, i64 256, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.19*)* @.omp_task_entry..21 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP161:%.*]] = bitcast i8* [[TMP160]] to %struct.kmp_task_t_with_privates.19*
// CHECK2-NEXT: [[TMP162:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19:%.*]], %struct.kmp_task_t_with_privates.19* [[TMP161]], i32 0, i32 0
// CHECK2-NEXT: [[TMP163:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19]], %struct.kmp_task_t_with_privates.19* [[TMP161]], i32 0, i32 2
// CHECK2-NEXT: [[TMP164:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20:%.*]], %struct..kmp_privates.t.20* [[TMP163]], i32 0, i32 0
// CHECK2-NEXT: [[TMP165:%.*]] = load i32, i32* [[C]], align 128
// CHECK2-NEXT: store i32 [[TMP165]], i32* [[TMP164]], align 128
// CHECK2-NEXT: [[TMP166:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP162]], i32 0, i32 2
// CHECK2-NEXT: store i32 0, i32* [[TMP166]], align 16
// CHECK2-NEXT: [[TMP167:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP160]])
// CHECK2-NEXT: [[TMP168:%.*]] = load i32, i32* @a, align 4
// CHECK2-NEXT: store i32 [[TMP168]], i32* [[RETVAL]], align 4
// CHECK2-NEXT: [[TMP169:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
// CHECK2-NEXT: call void @llvm.stackrestore(i8* [[TMP169]])
// CHECK2-NEXT: [[ARRAY_BEGIN32:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i32 0, i32 0
// CHECK2-NEXT: [[TMP170:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN32]], i64 2
// CHECK2-NEXT: br label [[ARRAYDESTROY_BODY:%.*]]
// CHECK2: arraydestroy.body:
// CHECK2-NEXT: [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP170]], [[ARRAYCTOR_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
// CHECK2-NEXT: [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4:[0-9]+]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN32]]
// CHECK2-NEXT: br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE33:%.*]], label [[ARRAYDESTROY_BODY]]
// CHECK2: arraydestroy.done33:
// CHECK2-NEXT: [[TMP171:%.*]] = load i32, i32* [[RETVAL]], align 4
// CHECK2-NEXT: ret i32 [[TMP171]]
//
//
// CHECK2-LABEL: define {{[^@]+}}@_ZN1SC1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1:[0-9]+]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK2-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK2-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SC2Ev(%struct.S* noundef [[THIS1]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry.
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias noundef [[TMP1:%.*]]) #[[ATTR3:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META3:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META6:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META8:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META10:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !12
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !12
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !12
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !12
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !12
// CHECK2-NEXT: store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !12
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !12
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 15, i32* @a, align 4
// CHECK2-NEXT: [[TMP11:%.*]] = load i32, i32* @a, align 4
// CHECK2-NEXT: [[CONV_I:%.*]] = trunc i32 [[TMP11]] to i8
// CHECK2-NEXT: [[TMP12:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP10]], i32 0, i32 0
// CHECK2-NEXT: [[TMP13:%.*]] = load i8*, i8** [[TMP12]], align 8
// CHECK2-NEXT: store i8 [[CONV_I]], i8* [[TMP13]], align 1
// CHECK2-NEXT: [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP10]], i32 0, i32 1
// CHECK2-NEXT: [[TMP15:%.*]] = load [2 x %struct.S]*, [2 x %struct.S]** [[TMP14]], align 8
// CHECK2-NEXT: [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[TMP15]], i64 0, i64 0
// CHECK2-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAYIDX_I]], i32 0, i32 0
// CHECK2-NEXT: store i32 10, i32* [[A_I]], align 4
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..2
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.1* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.0*, align 8
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.1*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.1* [[TMP1]], %struct.kmp_task_t_with_privates.1** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.1*, %struct.kmp_task_t_with_privates.1** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_1:%.*]], %struct.kmp_task_t_with_privates.1* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.0*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.1* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META13:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META16:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META18:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META20:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !22
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !22
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !22
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !22
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !22
// CHECK2-NEXT: store %struct.anon.0* [[TMP8]], %struct.anon.0** [[__CONTEXT_ADDR_I]], align 8, !noalias !22
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR_I]], align 8, !noalias !22
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 15, i32* @a, align 4
// CHECK2-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP10]], i32 0, i32 0
// CHECK2-NEXT: [[TMP12:%.*]] = load [2 x %struct.S]*, [2 x %struct.S]** [[TMP11]], align 8
// CHECK2-NEXT: [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[TMP12]], i64 0, i64 1
// CHECK2-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAYIDX_I]], i32 0, i32 0
// CHECK2-NEXT: store i32 10, i32* [[A_I]], align 4
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..4
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.3* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.2*, align 8
// CHECK2-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.3*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.3* [[TMP1]], %struct.kmp_task_t_with_privates.3** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.3*, %struct.kmp_task_t_with_privates.3** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_3:%.*]], %struct.kmp_task_t_with_privates.3* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.2*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.3* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META23:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META26:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META28:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META30:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !32
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !32
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !32
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !32
// CHECK2-NEXT: store %struct.anon.2* [[TMP8]], %struct.anon.2** [[__CONTEXT_ADDR_I]], align 8, !noalias !32
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon.2*, %struct.anon.2** [[__CONTEXT_ADDR_I]], align 8, !noalias !32
// CHECK2-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK2-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK2-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK2-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK2-NEXT: ]
// CHECK2: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK2: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !32
// CHECK2-NEXT: [[TMP15:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !32
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP16:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP14]], i8* [[TMP15]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
// CHECK2: .untied.jmp.1.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP17:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !32
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: call void @__kmpc_critical(%struct.ident_t* @[[GLOB1]], i32 [[TMP17]], [8 x i32]* @.gomp_critical_user_.var) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: call void @__kmpc_end_critical(%struct.ident_t* @[[GLOB1]], i32 [[TMP17]], [8 x i32]* @.gomp_critical_user_.var) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[CLEANUP_I]]
// CHECK2: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__3_EXIT]]
// CHECK2: .omp_outlined..3.exit:
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..6
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.5* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.4*, align 8
// CHECK2-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.5*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.5* [[TMP1]], %struct.kmp_task_t_with_privates.5** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.5*, %struct.kmp_task_t_with_privates.5** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_5:%.*]], %struct.kmp_task_t_with_privates.5* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.4*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.5* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META33:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META36:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META38:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META40:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !42
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !42
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !42
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !42
// CHECK2-NEXT: store %struct.anon.4* [[TMP8]], %struct.anon.4** [[__CONTEXT_ADDR_I]], align 8, !noalias !42
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon.4*, %struct.anon.4** [[__CONTEXT_ADDR_I]], align 8, !noalias !42
// CHECK2-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK2-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK2-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK2-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK2-NEXT: ]
// CHECK2: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK2: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !42
// CHECK2-NEXT: [[TMP15:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !42
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP16:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP14]], i8* [[TMP15]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__5_EXIT:%.*]]
// CHECK2: .untied.jmp.1.i:
// CHECK2-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[CLEANUP_I]]
// CHECK2: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__5_EXIT]]
// CHECK2: .omp_outlined..5.exit:
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.7* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.6*, align 8
// CHECK2-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.7*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.7* [[TMP1]], %struct.kmp_task_t_with_privates.7** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.7*, %struct.kmp_task_t_with_privates.7** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_7:%.*]], %struct.kmp_task_t_with_privates.7* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.6*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.7* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META43:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META46:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META48:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META50:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !52
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !52
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !52
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !52
// CHECK2-NEXT: store %struct.anon.6* [[TMP8]], %struct.anon.6** [[__CONTEXT_ADDR_I]], align 8, !noalias !52
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon.6*, %struct.anon.6** [[__CONTEXT_ADDR_I]], align 8, !noalias !52
// CHECK2-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK2-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK2-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK2-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK2-NEXT: ]
// CHECK2: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK2: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !52
// CHECK2-NEXT: [[TMP15:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !52
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP16:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP14]], i8* [[TMP15]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__7_EXIT:%.*]]
// CHECK2: .untied.jmp.1.i:
// CHECK2-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[CLEANUP_I]]
// CHECK2: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__7_EXIT]]
// CHECK2: .omp_outlined..7.exit:
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..10
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.9* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.8*, align 8
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.9*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.9* [[TMP1]], %struct.kmp_task_t_with_privates.9** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.9*, %struct.kmp_task_t_with_privates.9** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_9:%.*]], %struct.kmp_task_t_with_privates.9* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.8*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.9* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META53:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META56:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META58:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META60:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !62
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !62
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !62
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !62
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !62
// CHECK2-NEXT: store %struct.anon.8* [[TMP8]], %struct.anon.8** [[__CONTEXT_ADDR_I]], align 8, !noalias !62
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon.8*, %struct.anon.8** [[__CONTEXT_ADDR_I]], align 8, !noalias !62
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 2, i32* @a, align 4
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..12
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.11* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.10*, align 8
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.11*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.11* [[TMP1]], %struct.kmp_task_t_with_privates.11** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.11*, %struct.kmp_task_t_with_privates.11** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_11:%.*]], %struct.kmp_task_t_with_privates.11* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.10*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.11* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META63:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META66:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META68:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META70:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !72
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !72
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !72
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !72
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !72
// CHECK2-NEXT: store %struct.anon.10* [[TMP8]], %struct.anon.10** [[__CONTEXT_ADDR_I]], align 8, !noalias !72
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon.10*, %struct.anon.10** [[__CONTEXT_ADDR_I]], align 8, !noalias !72
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 2, i32* @a, align 4
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..14
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.13* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.12*, align 8
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.13*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.13* [[TMP1]], %struct.kmp_task_t_with_privates.13** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.13*, %struct.kmp_task_t_with_privates.13** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_13:%.*]], %struct.kmp_task_t_with_privates.13* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.12*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.13* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META73:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META76:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META78:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META80:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !82
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !82
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !82
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !82
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !82
// CHECK2-NEXT: store %struct.anon.12* [[TMP8]], %struct.anon.12** [[__CONTEXT_ADDR_I]], align 8, !noalias !82
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon.12*, %struct.anon.12** [[__CONTEXT_ADDR_I]], align 8, !noalias !82
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 3, i32* @a, align 4
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..16
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.15* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.14*, align 8
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.15*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.15* [[TMP1]], %struct.kmp_task_t_with_privates.15** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.15*, %struct.kmp_task_t_with_privates.15** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_15:%.*]], %struct.kmp_task_t_with_privates.15* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.14*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.15* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META83:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META86:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META88:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META90:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !92
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !92
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !92
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !92
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !92
// CHECK2-NEXT: store %struct.anon.14* [[TMP8]], %struct.anon.14** [[__CONTEXT_ADDR_I]], align 8, !noalias !92
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon.14*, %struct.anon.14** [[__CONTEXT_ADDR_I]], align 8, !noalias !92
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 4, i32* @a, align 4
// CHECK2-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_14:%.*]], %struct.anon.14* [[TMP10]], i32 0, i32 0
// CHECK2-NEXT: [[TMP12:%.*]] = load i32*, i32** [[TMP11]], align 8
// CHECK2-NEXT: store i32 5, i32* [[TMP12]], align 128
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_privates_map.
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct..kmp_privates.t* noalias noundef [[TMP0:%.*]], i32** noalias noundef [[TMP1:%.*]]) #[[ATTR7:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 8
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca i32**, align 8
// CHECK2-NEXT: store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 8
// CHECK2-NEXT: store i32** [[TMP1]], i32*** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP2]], i32 0, i32 0
// CHECK2-NEXT: [[TMP4:%.*]] = load i32**, i32*** [[DOTADDR1]], align 8
// CHECK2-NEXT: store i32* [[TMP3]], i32** [[TMP4]], align 8
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..19
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.18* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.17*, align 8
// CHECK2-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.18*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.18* [[TMP1]], %struct.kmp_task_t_with_privates.18** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.18*, %struct.kmp_task_t_with_privates.18** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18:%.*]], %struct.kmp_task_t_with_privates.18* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 128
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.17*
// CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18]], %struct.kmp_task_t_with_privates.18* [[TMP3]], i32 0, i32 2
// CHECK2-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
// CHECK2-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.18* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META93:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META96:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META98:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META100:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !102
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !102
// CHECK2-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !102
// CHECK2-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i32**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !102
// CHECK2-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !102
// CHECK2-NEXT: store %struct.anon.17* [[TMP8]], %struct.anon.17** [[__CONTEXT_ADDR_I]], align 8, !noalias !102
// CHECK2-NEXT: [[TMP12:%.*]] = load %struct.anon.17*, %struct.anon.17** [[__CONTEXT_ADDR_I]], align 8, !noalias !102
// CHECK2-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !102
// CHECK2-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !102
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i32**)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: call void [[TMP15]](i8* [[TMP14]], i32** [[DOTFIRSTPRIV_PTR_ADDR_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP16:%.*]] = load i32*, i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !102
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 4, i32* [[TMP16]], align 128
// CHECK2-NEXT: store i32 4, i32* @a, align 4
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@_ZN1SD1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK2-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK2-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SD2Ev(%struct.S* noundef [[THIS1]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_privates_map..20
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct..kmp_privates.t.20* noalias noundef [[TMP0:%.*]], i32** noalias noundef [[TMP1:%.*]], %struct.S** noalias noundef [[TMP2:%.*]], %struct.S*** noalias noundef [[TMP3:%.*]]) #[[ATTR7]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t.20*, align 8
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca i32**, align 8
// CHECK2-NEXT: [[DOTADDR2:%.*]] = alloca %struct.S**, align 8
// CHECK2-NEXT: [[DOTADDR3:%.*]] = alloca %struct.S***, align 8
// CHECK2-NEXT: store %struct..kmp_privates.t.20* [[TMP0]], %struct..kmp_privates.t.20** [[DOTADDR]], align 8
// CHECK2-NEXT: store i32** [[TMP1]], i32*** [[DOTADDR1]], align 8
// CHECK2-NEXT: store %struct.S** [[TMP2]], %struct.S*** [[DOTADDR2]], align 8
// CHECK2-NEXT: store %struct.S*** [[TMP3]], %struct.S**** [[DOTADDR3]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = load %struct..kmp_privates.t.20*, %struct..kmp_privates.t.20** [[DOTADDR]], align 8
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20:%.*]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP6:%.*]] = load i32**, i32*** [[DOTADDR1]], align 8
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP7:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP8:%.*]] = load %struct.S***, %struct.S**** [[DOTADDR3]], align 8
// CHECK2-NEXT: store %struct.S** [[TMP7]], %struct.S*** [[TMP8]], align 8
// CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 3
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.S**, %struct.S*** [[DOTADDR2]], align 8
// CHECK2-NEXT: store %struct.S* [[TMP9]], %struct.S** [[TMP10]], align 8
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..21
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.19* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.16*, align 8
// CHECK2-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTLOCAL_PTR_ADDR_I:%.*]] = alloca %struct.S*, align 8
// CHECK2-NEXT: [[DOTLOCAL_PTR_ADDR1_I:%.*]] = alloca %struct.S**, align 8
// CHECK2-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[REF_TMP_I:%.*]] = alloca [[STRUCT_S:%.*]], align 4
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.19*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.19* [[TMP1]], %struct.kmp_task_t_with_privates.19** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.19*, %struct.kmp_task_t_with_privates.19** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19:%.*]], %struct.kmp_task_t_with_privates.19* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 128
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.16*
// CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19]], %struct.kmp_task_t_with_privates.19* [[TMP3]], i32 0, i32 2
// CHECK2-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t.20* [[TMP9]] to i8*
// CHECK2-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.19* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META103:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META106:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META108:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META110:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
// CHECK2-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !112
// CHECK2-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t.20*, i32**, %struct.S**, %struct.S***)* @.omp_task_privates_map..20 to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !112
// CHECK2-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
// CHECK2-NEXT: store %struct.anon.16* [[TMP8]], %struct.anon.16** [[__CONTEXT_ADDR_I]], align 8, !noalias !112
// CHECK2-NEXT: [[TMP12:%.*]] = load %struct.anon.16*, %struct.anon.16** [[__CONTEXT_ADDR_I]], align 8, !noalias !112
// CHECK2-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !112
// CHECK2-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i32**, %struct.S**, %struct.S***)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: call void [[TMP15]](i8* [[TMP14]], i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], %struct.S** [[DOTLOCAL_PTR_ADDR_I]], %struct.S*** [[DOTLOCAL_PTR_ADDR1_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP16:%.*]] = load i32*, i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !112
// CHECK2-NEXT: [[TMP17:%.*]] = load %struct.S*, %struct.S** [[DOTLOCAL_PTR_ADDR_I]], align 8, !noalias !112
// CHECK2-NEXT: [[TMP18:%.*]] = load %struct.S**, %struct.S*** [[DOTLOCAL_PTR_ADDR1_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP19:%.*]] = load %struct.S*, %struct.S** [[TMP18]], align 8
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP20:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP21:%.*]] = load i32, i32* [[TMP20]], align 4
// CHECK2-NEXT: switch i32 [[TMP21]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK2-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK2-NEXT: i32 1, label [[DOTUNTIED_JMP_2_I:%.*]]
// CHECK2-NEXT: i32 2, label [[DOTUNTIED_JMP_3_I:%.*]]
// CHECK2-NEXT: i32 3, label [[DOTUNTIED_JMP_5_I:%.*]]
// CHECK2-NEXT: i32 4, label [[DOTUNTIED_JMP_7_I:%.*]]
// CHECK2-NEXT: i32 5, label [[DOTUNTIED_JMP_10_I:%.*]]
// CHECK2-NEXT: ]
// CHECK2: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK2: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP22:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 1, i32* [[TMP22]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP23:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK2-NEXT: [[TMP24:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP25:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP23]], i8* [[TMP24]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT:%.*]]
// CHECK2: .untied.jmp.2.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[TMP17]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP26:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[DOTS2__VOID_ADDR_I:%.*]] = call i8* @__kmpc_alloc(i32 [[TMP26]], i64 4, i8* inttoptr (i64 7 to i8*)) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[DOTS2__ADDR_I:%.*]] = bitcast i8* [[DOTS2__VOID_ADDR_I]] to %struct.S*
// CHECK2-NEXT: store %struct.S* [[DOTS2__ADDR_I]], %struct.S** [[TMP18]], align 8
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP27:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 2, i32* [[TMP27]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP28:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK2-NEXT: [[TMP29:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP30:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP28]], i8* [[TMP29]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK2: .untied.jmp.3.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[TMP19]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[TMP19]], i32 0, i32 0
// CHECK2-NEXT: store i32 0, i32* [[A_I]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP31:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP32:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP31]], i32 1, i64 256, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.18*)* @.omp_task_entry..19 to i32 (i32, i8*)*)) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP33:%.*]] = bitcast i8* [[TMP32]] to %struct.kmp_task_t_with_privates.18*
// CHECK2-NEXT: [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18:%.*]], %struct.kmp_task_t_with_privates.18* [[TMP33]], i32 0, i32 0
// CHECK2-NEXT: [[TMP35:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18]], %struct.kmp_task_t_with_privates.18* [[TMP33]], i32 0, i32 2
// CHECK2-NEXT: [[TMP36:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP35]], i32 0, i32 0
// CHECK2-NEXT: [[TMP37:%.*]] = load i32, i32* [[TMP16]], align 128
// CHECK2-NEXT: store i32 [[TMP37]], i32* [[TMP36]], align 128
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP38:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP39:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP38]], i8* [[TMP32]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP40:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 3, i32* [[TMP40]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP41:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK2-NEXT: [[TMP42:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP43:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP41]], i8* [[TMP42]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK2: .untied.jmp.5.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP44:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP45:%.*]] = call i32 @__kmpc_omp_taskyield(%struct.ident_t* @[[GLOB1]], i32 [[TMP44]], i32 0) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP46:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 4, i32* [[TMP46]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP47:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK2-NEXT: [[TMP48:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP49:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP47]], i8* [[TMP48]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK2: .untied.jmp.7.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[REF_TMP_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP50:%.*]] = bitcast %struct.S* [[TMP17]] to i8*
// CHECK2-NEXT: [[TMP51:%.*]] = bitcast %struct.S* [[REF_TMP_I]] to i8*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP50]], i8* align 4 [[TMP51]], i64 4, i1 false) #[[ATTR4]]
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[REF_TMP_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[A9_I:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[TMP19]], i32 0, i32 0
// CHECK2-NEXT: store i32 10, i32* [[A9_I]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP52:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP53:%.*]] = call i32 @__kmpc_omp_taskwait(%struct.ident_t* @[[GLOB1]], i32 [[TMP52]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP54:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 5, i32* [[TMP54]], align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP55:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK2-NEXT: [[TMP56:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: [[TMP57:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP55]], i8* [[TMP56]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK2: .untied.jmp.10.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[TMP19]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP58:%.*]] = load i32, i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP59:%.*]] = bitcast %struct.S* [[TMP19]] to i8*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: call void @__kmpc_free(i32 [[TMP58]], i8* [[TMP59]], i8* inttoptr (i64 7 to i8*)) #[[ATTR4]]
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[TMP17]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[CLEANUP_I]]
// CHECK2: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK2: .omp_outlined..17.exit:
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@_ZN1SC2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK2-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK2-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
// CHECK2-NEXT: [[A:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[THIS1]], i32 0, i32 0
// CHECK2-NEXT: store i32 0, i32* [[A]], align 4
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@_ZN1SD2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK2-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK2-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
// CHECK2-NEXT: ret void
//
//
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-LABEL: define {{[^@]+}}@__cxx_global_var_init
// CHECK2-SAME: () #[[ATTR7]] section "__TEXT,__StaticInit,regular,pure_instructions" {
// CHECK2-NEXT: entry:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN2S1C1Ev(%struct.S1* noundef @s1)
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@_ZN2S1C1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct.S1* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK2-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK2-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN2S1C2Ev(%struct.S1* noundef [[THIS1]])
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: ret void
//
//
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-LABEL: define {{[^@]+}}@_ZN2S1C2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct.S1* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK2-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK2-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-NEXT: call void @_ZN2S18taskinitEv(%struct.S1* noundef [[THIS1]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@_ZN2S18taskinitEv
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct.S1* noundef [[THIS:%.*]]) #[[ATTR8:[0-9]+]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK2-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON_21:%.*]], align 8
// CHECK2-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]])
// CHECK2-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK2-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
// CHECK2-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_21]], %struct.anon.21* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK2-NEXT: store %struct.S1* [[THIS1]], %struct.S1** [[TMP1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i32 1, i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.22*)* @.omp_task_entry..23 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP3:%.*]] = bitcast i8* [[TMP2]] to %struct.kmp_task_t_with_privates.22*
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_22:%.*]], %struct.kmp_task_t_with_privates.22* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP6:%.*]] = load i8*, i8** [[TMP5]], align 8
// CHECK2-NEXT: [[TMP7:%.*]] = bitcast %struct.anon.21* [[AGG_CAPTURED]] to i8*
// CHECK2-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP6]], i8* align 8 [[TMP7]], i64 8, i1 false)
// CHECK2-NEXT: [[TMP8:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP0]], i8* [[TMP2]])
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..23
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.22* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.21*, align 8
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.22*, align 8
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.22* [[TMP1]], %struct.kmp_task_t_with_privates.22** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.22*, %struct.kmp_task_t_with_privates.22** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_22:%.*]], %struct.kmp_task_t_with_privates.22* [[TMP3]], i32 0, i32 0
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.21*
// CHECK2-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.22* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META113:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META116:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META118:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META120:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !122
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !122
// CHECK2-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !122
// CHECK2-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !122
// CHECK2-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !122
// CHECK2-NEXT: store %struct.anon.21* [[TMP8]], %struct.anon.21** [[__CONTEXT_ADDR_I]], align 8, !noalias !122
// CHECK2-NEXT: [[TMP10:%.*]] = load %struct.anon.21*, %struct.anon.21** [[__CONTEXT_ADDR_I]], align 8, !noalias !122
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_21:%.*]], %struct.anon.21* [[TMP10]], i32 0, i32 0
// CHECK2-NEXT: [[TMP12:%.*]] = load %struct.S1*, %struct.S1** [[TMP11]], align 8
// CHECK2-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP12]], i32 0, i32 0
// CHECK2-NEXT: store i32 0, i32* [[A_I]], align 4
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@_Z4xxxxv
// CHECK2-SAME: () #[[ATTR8]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: call void @_Z6foobarIvEvv()
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@_Z6foobarIvEvv
// CHECK2-SAME: () #[[ATTR8]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[A:%.*]] = alloca float, align 4
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, float*)* @.omp_outlined..24 to void (i32*, i32*, ...)*), float* [[A]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_outlined..24
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32* noalias noundef [[DOTGLOBAL_TID_:%.*]], i32* noalias noundef [[DOTBOUND_TID_:%.*]], float* noundef nonnull align 4 dereferenceable(4) [[A:%.*]]) #[[ATTR9:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[A_ADDR:%.*]] = alloca float*, align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[B:%.*]] = alloca double, align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON_23:%.*]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK2-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: store float* [[A]], float** [[A_ADDR]], align 8
// CHECK2-NEXT: [[TMP0:%.*]] = load float*, float** [[A_ADDR]], align 8
// CHECK2-NEXT: [[TMP1:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[TMP1]], align 4
// CHECK2-NEXT: [[TMP3:%.*]] = call i32 @__kmpc_single(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
// CHECK2-NEXT: [[TMP4:%.*]] = icmp ne i32 [[TMP3]], 0
// CHECK2-NEXT: br i1 [[TMP4]], label [[OMP_IF_THEN:%.*]], label [[OMP_IF_END:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2: omp_if.then:
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_ANON_23]], %struct.anon.23* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK2-NEXT: store float* [[TMP0]], float** [[TMP5]], align 8
// CHECK2-NEXT: [[TMP6:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i32 1, i64 48, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.24*)* @.omp_task_entry..27 to i32 (i32, i8*)*))
// CHECK2-NEXT: [[TMP7:%.*]] = bitcast i8* [[TMP6]] to %struct.kmp_task_t_with_privates.24*
// CHECK2-NEXT: [[TMP8:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_24:%.*]], %struct.kmp_task_t_with_privates.24* [[TMP7]], i32 0, i32 0
// CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP8]], i32 0, i32 0
// CHECK2-NEXT: [[TMP10:%.*]] = load i8*, i8** [[TMP9]], align 8
// CHECK2-NEXT: [[TMP11:%.*]] = bitcast %struct.anon.23* [[AGG_CAPTURED]] to i8*
// CHECK2-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP10]], i8* align 8 [[TMP11]], i64 8, i1 false)
// CHECK2-NEXT: [[TMP12:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_24]], %struct.kmp_task_t_with_privates.24* [[TMP7]], i32 0, i32 1
// CHECK2-NEXT: [[TMP13:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_25:%.*]], %struct..kmp_privates.t.25* [[TMP12]], i32 0, i32 0
// CHECK2-NEXT: [[TMP14:%.*]] = load double, double* [[B]], align 8
// CHECK2-NEXT: store double [[TMP14]], double* [[TMP13]], align 8
// CHECK2-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]], i8* [[TMP6]])
// CHECK2-NEXT: call void @__kmpc_end_single(%struct.ident_t* @[[GLOB1]], i32 [[TMP2]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: br label [[OMP_IF_END]]
// CHECK2: omp_if.end:
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP2]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_privates_map..26
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (%struct..kmp_privates.t.25* noalias noundef [[TMP0:%.*]], double** noalias noundef [[TMP1:%.*]]) #[[ATTR7]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t.25*, align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca double**, align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: store %struct..kmp_privates.t.25* [[TMP0]], %struct..kmp_privates.t.25** [[DOTADDR]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store double** [[TMP1]], double*** [[DOTADDR1]], align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[TMP2:%.*]] = load %struct..kmp_privates.t.25*, %struct..kmp_privates.t.25** [[DOTADDR]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_25:%.*]], %struct..kmp_privates.t.25* [[TMP2]], i32 0, i32 0
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP4:%.*]] = load double**, double*** [[DOTADDR1]], align 8
// CHECK2-NEXT: store double* [[TMP3]], double** [[TMP4]], align 8
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_task_entry..27
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK2-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.24* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK2-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK2-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.23*, align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca double*, align 8
// CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.24*, align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: store %struct.kmp_task_t_with_privates.24* [[TMP1]], %struct.kmp_task_t_with_privates.24** [[DOTADDR1]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.24*, %struct.kmp_task_t_with_privates.24** [[DOTADDR1]], align 8
// CHECK2-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_24:%.*]], %struct.kmp_task_t_with_privates.24* [[TMP3]], i32 0, i32 0
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK2-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.23*
// CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_24]], %struct.kmp_task_t_with_privates.24* [[TMP3]], i32 0, i32 1
// CHECK2-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t.25* [[TMP9]] to i8*
// CHECK2-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.24* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META123:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META126:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META128:![0-9]+]])
// CHECK2-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META130:![0-9]+]])
// CHECK2-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !132
// CHECK2-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !132
// CHECK2-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !132
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t.25*, double**)* @.omp_task_privates_map..26 to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !132
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !132
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: store %struct.anon.23* [[TMP8]], %struct.anon.23** [[__CONTEXT_ADDR_I]], align 8, !noalias !132
// CHECK2-NEXT: [[TMP12:%.*]] = load %struct.anon.23*, %struct.anon.23** [[__CONTEXT_ADDR_I]], align 8, !noalias !132
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !132
// CHECK2-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !132
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, double**)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-NEXT: call void [[TMP15]](i8* [[TMP14]], double** [[DOTFIRSTPRIV_PTR_ADDR_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK2-NEXT: [[TMP16:%.*]] = load double*, double** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !132
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP17:%.*]] = load double, double* [[TMP16]], align 8
Revert "[OpenMP] Codegen aggregate for outlined function captures" This reverts commit 1d66649adf28d48ae1731516d87fb899426e3349. Revert to fix AMG GPU issue.
2021-09-22 04:20:39 +08:00
// CHECK2-NEXT: [[TMP18:%.*]] = getelementptr inbounds [[STRUCT_ANON_23:%.*]], %struct.anon.23* [[TMP12]], i32 0, i32 0
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: [[TMP19:%.*]] = load float*, float** [[TMP18]], align 8
// CHECK2-NEXT: [[TMP20:%.*]] = load float, float* [[TMP19]], align 4
// CHECK2-NEXT: [[CONV_I:%.*]] = fpext float [[TMP20]] to double
// CHECK2-NEXT: [[ADD_I:%.*]] = fadd double [[CONV_I]], [[TMP17]]
// CHECK2-NEXT: [[CONV1_I:%.*]] = fptrunc double [[ADD_I]] to float
// CHECK2-NEXT: store float [[CONV1_I]], float* [[TMP19]], align 4
// CHECK2-NEXT: ret i32 0
//
//
// CHECK2-LABEL: define {{[^@]+}}@_GLOBAL__sub_I_task_codegen.cpp
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK2-SAME: () #[[ATTR7]] section "__TEXT,__StaticInit,regular,pure_instructions" {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: call void @__cxx_global_var_init()
// CHECK2-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@main
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-SAME: () #[[ATTR0:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[RETVAL:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[B:%.*]] = alloca i8, align 1
// CHECK3-NEXT: [[S:%.*]] = alloca [2 x %struct.S], align 4
// CHECK3-NEXT: [[SAVED_STACK:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__VLA_EXPR0:%.*]] = alloca i64, align 8
// CHECK3-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 8
// CHECK3-NEXT: [[AGG_CAPTURED2:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 8
// CHECK3-NEXT: [[DOTDEP_ARR_ADDR:%.*]] = alloca [4 x %struct.kmp_depend_info], align 8
// CHECK3-NEXT: [[DEP_COUNTER_ADDR:%.*]] = alloca i64, align 8
// CHECK3-NEXT: [[AGG_CAPTURED6:%.*]] = alloca [[STRUCT_ANON_2:%.*]], align 1
// CHECK3-NEXT: [[AGG_CAPTURED9:%.*]] = alloca [[STRUCT_ANON_4:%.*]], align 1
// CHECK3-NEXT: [[DOTDEP_ARR_ADDR11:%.*]] = alloca [2 x %struct.kmp_depend_info], align 8
// CHECK3-NEXT: [[DEP_COUNTER_ADDR17:%.*]] = alloca i64, align 8
// CHECK3-NEXT: [[AGG_CAPTURED19:%.*]] = alloca [[STRUCT_ANON_6:%.*]], align 1
// CHECK3-NEXT: [[DOTDEP_ARR_ADDR21:%.*]] = alloca [2 x %struct.kmp_depend_info], align 8
// CHECK3-NEXT: [[DEP_COUNTER_ADDR27:%.*]] = alloca i64, align 8
// CHECK3-NEXT: [[AGG_CAPTURED29:%.*]] = alloca [[STRUCT_ANON_8:%.*]], align 1
// CHECK3-NEXT: [[DOTDEP_ARR_ADDR31:%.*]] = alloca [3 x %struct.kmp_depend_info], align 8
// CHECK3-NEXT: [[DEP_COUNTER_ADDR37:%.*]] = alloca i64, align 8
// CHECK3-NEXT: [[AGG_CAPTURED39:%.*]] = alloca [[STRUCT_ANON_10:%.*]], align 1
// CHECK3-NEXT: [[FLAG:%.*]] = alloca i8, align 1
// CHECK3-NEXT: [[AGG_CAPTURED42:%.*]] = alloca [[STRUCT_ANON_12:%.*]], align 1
// CHECK3-NEXT: [[C:%.*]] = alloca i32, align 128
// CHECK3-NEXT: [[AGG_CAPTURED45:%.*]] = alloca [[STRUCT_ANON_14:%.*]], align 8
// CHECK3-NEXT: [[AGG_CAPTURED48:%.*]] = alloca [[STRUCT_ANON_16:%.*]], align 1
// CHECK3-NEXT: store i32 0, i32* [[RETVAL]], align 4
// CHECK3-NEXT: [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i32 0, i32 0
// CHECK3-NEXT: [[ARRAYCTOR_END:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAY_BEGIN]], i64 2
// CHECK3-NEXT: br label [[ARRAYCTOR_LOOP:%.*]]
// CHECK3: arrayctor.loop:
// CHECK3-NEXT: [[ARRAYCTOR_CUR:%.*]] = phi %struct.S* [ [[ARRAY_BEGIN]], [[ENTRY:%.*]] ], [ [[ARRAYCTOR_NEXT:%.*]], [[ARRAYCTOR_LOOP]] ]
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[ARRAYCTOR_CUR]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[ARRAYCTOR_NEXT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYCTOR_CUR]], i64 1
// CHECK3-NEXT: [[ARRAYCTOR_DONE:%.*]] = icmp eq %struct.S* [[ARRAYCTOR_NEXT]], [[ARRAYCTOR_END]]
// CHECK3-NEXT: br i1 [[ARRAYCTOR_DONE]], label [[ARRAYCTOR_CONT:%.*]], label [[ARRAYCTOR_LOOP]]
// CHECK3: arrayctor.cont:
// CHECK3-NEXT: [[TMP0:%.*]] = load i32, i32* @a, align 4
// CHECK3-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
// CHECK3-NEXT: [[TMP2:%.*]] = call i8* @llvm.stacksave()
// CHECK3-NEXT: store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
// CHECK3-NEXT: [[TMP3:%.*]] = mul nuw i64 10, [[TMP1]]
// CHECK3-NEXT: [[VLA:%.*]] = alloca i32, i64 [[TMP3]], align 16
// CHECK3-NEXT: store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK3-NEXT: store i8* [[B]], i8** [[TMP4]], align 8
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
// CHECK3-NEXT: store [2 x %struct.S]* [[S]], [2 x %struct.S]** [[TMP5]], align 8
// CHECK3-NEXT: [[TMP6:%.*]] = load i8, i8* [[B]], align 1
// CHECK3-NEXT: [[CONV:%.*]] = sext i8 [[TMP6]] to i32
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB3:[0-9]+]])
// CHECK3-NEXT: [[TMP7:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[OMP_GLOBAL_THREAD_NUM]], i32 33, i64 40, i64 16, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.kmp_task_t_with_privates*
// CHECK3-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP8]], i32 0, i32 0
// CHECK3-NEXT: [[TMP10:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP9]], i32 0, i32 0
// CHECK3-NEXT: [[TMP11:%.*]] = load i8*, i8** [[TMP10]], align 8
// CHECK3-NEXT: [[TMP12:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
// CHECK3-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP11]], i8* align 8 [[TMP12]], i64 16, i1 false)
// CHECK3-NEXT: [[TMP13:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP9]], i32 0, i32 4
// CHECK3-NEXT: [[TMP14:%.*]] = bitcast %union.kmp_cmplrdata_t* [[TMP13]] to i32*
// CHECK3-NEXT: store i32 [[CONV]], i32* [[TMP14]], align 8
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB3]])
// CHECK3-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM1]], i8* [[TMP7]])
// CHECK3-NEXT: [[TMP16:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[AGG_CAPTURED2]], i32 0, i32 0
// CHECK3-NEXT: store [2 x %struct.S]* [[S]], [2 x %struct.S]** [[TMP16]], align 8
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM3:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB5:[0-9]+]])
// CHECK3-NEXT: [[TMP17:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM3]], i32 1, i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.1*)* @.omp_task_entry..2 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP18:%.*]] = bitcast i8* [[TMP17]] to %struct.kmp_task_t_with_privates.1*
// CHECK3-NEXT: [[TMP19:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_1:%.*]], %struct.kmp_task_t_with_privates.1* [[TMP18]], i32 0, i32 0
// CHECK3-NEXT: [[TMP20:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP19]], i32 0, i32 0
// CHECK3-NEXT: [[TMP21:%.*]] = load i8*, i8** [[TMP20]], align 8
// CHECK3-NEXT: [[TMP22:%.*]] = bitcast %struct.anon.0* [[AGG_CAPTURED2]] to i8*
// CHECK3-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP21]], i8* align 8 [[TMP22]], i64 8, i1 false)
// CHECK3-NEXT: [[TMP23:%.*]] = getelementptr inbounds [4 x %struct.kmp_depend_info], [4 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR]], i64 0, i64 0
// CHECK3-NEXT: [[TMP24:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO:%.*]], %struct.kmp_depend_info* [[TMP23]], i64 0
// CHECK3-NEXT: [[TMP25:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i32 0, i32 0
// CHECK3-NEXT: store i64 ptrtoint (i32* @a to i64), i64* [[TMP25]], align 8
// CHECK3-NEXT: [[TMP26:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i32 0, i32 1
// CHECK3-NEXT: store i64 4, i64* [[TMP26]], align 8
// CHECK3-NEXT: [[TMP27:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 1, i8* [[TMP27]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP28:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP23]], i64 1
// CHECK3-NEXT: [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP28]], i32 0, i32 0
// CHECK3-NEXT: [[TMP30:%.*]] = ptrtoint i8* [[B]] to i64
// CHECK3-NEXT: store i64 [[TMP30]], i64* [[TMP29]], align 8
// CHECK3-NEXT: [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP28]], i32 0, i32 1
// CHECK3-NEXT: store i64 1, i64* [[TMP31]], align 8
// CHECK3-NEXT: [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP28]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 1, i8* [[TMP32]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP33:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP23]], i64 2
// CHECK3-NEXT: [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP33]], i32 0, i32 0
// CHECK3-NEXT: [[TMP35:%.*]] = ptrtoint [2 x %struct.S]* [[S]] to i64
// CHECK3-NEXT: store i64 [[TMP35]], i64* [[TMP34]], align 8
// CHECK3-NEXT: [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP33]], i32 0, i32 1
// CHECK3-NEXT: store i64 8, i64* [[TMP36]], align 8
// CHECK3-NEXT: [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP33]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 1, i8* [[TMP37]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP38:%.*]] = mul nsw i64 0, [[TMP1]]
// CHECK3-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP38]]
// CHECK3-NEXT: [[TMP39:%.*]] = mul nsw i64 9, [[TMP1]]
// CHECK3-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP39]]
// CHECK3-NEXT: [[TMP40:%.*]] = getelementptr i32, i32* [[ARRAYIDX4]], i32 1
// CHECK3-NEXT: [[TMP41:%.*]] = ptrtoint i32* [[ARRAYIDX]] to i64
// CHECK3-NEXT: [[TMP42:%.*]] = ptrtoint i32* [[TMP40]] to i64
// CHECK3-NEXT: [[TMP43:%.*]] = sub nuw i64 [[TMP42]], [[TMP41]]
// CHECK3-NEXT: [[TMP44:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP23]], i64 3
// CHECK3-NEXT: [[TMP45:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP44]], i32 0, i32 0
// CHECK3-NEXT: [[TMP46:%.*]] = ptrtoint i32* [[ARRAYIDX]] to i64
// CHECK3-NEXT: store i64 [[TMP46]], i64* [[TMP45]], align 8
// CHECK3-NEXT: [[TMP47:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP44]], i32 0, i32 1
// CHECK3-NEXT: store i64 [[TMP43]], i64* [[TMP47]], align 8
// CHECK3-NEXT: [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP44]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 1, i8* [[TMP48]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i64 4, i64* [[DEP_COUNTER_ADDR]], align 8
// CHECK3-NEXT: [[TMP49:%.*]] = bitcast %struct.kmp_depend_info* [[TMP23]] to i8*
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM5:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB5]])
// CHECK3-NEXT: [[TMP50:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM5]], i8* [[TMP17]], i32 4, i8* [[TMP49]], i32 0, i8* null)
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM7:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB7:[0-9]+]])
// CHECK3-NEXT: [[TMP51:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM7]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.3*)* @.omp_task_entry..4 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP52:%.*]] = bitcast i8* [[TMP51]] to %struct.kmp_task_t_with_privates.3*
// CHECK3-NEXT: [[TMP53:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_3:%.*]], %struct.kmp_task_t_with_privates.3* [[TMP52]], i32 0, i32 0
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM8:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB7]])
// CHECK3-NEXT: [[TMP54:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP53]], i32 0, i32 2
// CHECK3-NEXT: store i32 0, i32* [[TMP54]], align 8
// CHECK3-NEXT: [[TMP55:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM8]], i8* [[TMP51]])
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM10:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB9:[0-9]+]])
// CHECK3-NEXT: [[TMP56:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM10]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.5*)* @.omp_task_entry..6 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP57:%.*]] = bitcast i8* [[TMP56]] to %struct.kmp_task_t_with_privates.5*
// CHECK3-NEXT: [[TMP58:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_5:%.*]], %struct.kmp_task_t_with_privates.5* [[TMP57]], i32 0, i32 0
// CHECK3-NEXT: [[TMP59:%.*]] = getelementptr inbounds [2 x %struct.kmp_depend_info], [2 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR11]], i64 0, i64 0
// CHECK3-NEXT: [[ARRAYIDX12:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 0
// CHECK3-NEXT: [[TMP60:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP59]], i64 0
// CHECK3-NEXT: [[TMP61:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i32 0, i32 0
// CHECK3-NEXT: [[TMP62:%.*]] = ptrtoint %struct.S* [[ARRAYIDX12]] to i64
// CHECK3-NEXT: store i64 [[TMP62]], i64* [[TMP61]], align 8
// CHECK3-NEXT: [[TMP63:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i32 0, i32 1
// CHECK3-NEXT: store i64 4, i64* [[TMP63]], align 8
// CHECK3-NEXT: [[TMP64:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 3, i8* [[TMP64]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP65:%.*]] = load i8, i8* [[B]], align 1
// CHECK3-NEXT: [[TMP66:%.*]] = sext i8 [[TMP65]] to i64
// CHECK3-NEXT: [[TMP67:%.*]] = mul nsw i64 4, [[TMP1]]
// CHECK3-NEXT: [[ARRAYIDX13:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP67]]
// CHECK3-NEXT: [[ARRAYIDX14:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX13]], i64 [[TMP66]]
// CHECK3-NEXT: [[TMP68:%.*]] = load i8, i8* [[B]], align 1
// CHECK3-NEXT: [[TMP69:%.*]] = sext i8 [[TMP68]] to i64
// CHECK3-NEXT: [[TMP70:%.*]] = mul nsw i64 9, [[TMP1]]
// CHECK3-NEXT: [[ARRAYIDX15:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP70]]
// CHECK3-NEXT: [[ARRAYIDX16:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX15]], i64 [[TMP69]]
// CHECK3-NEXT: [[TMP71:%.*]] = getelementptr i32, i32* [[ARRAYIDX16]], i32 1
// CHECK3-NEXT: [[TMP72:%.*]] = ptrtoint i32* [[ARRAYIDX14]] to i64
// CHECK3-NEXT: [[TMP73:%.*]] = ptrtoint i32* [[TMP71]] to i64
// CHECK3-NEXT: [[TMP74:%.*]] = sub nuw i64 [[TMP73]], [[TMP72]]
// CHECK3-NEXT: [[TMP75:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP59]], i64 1
// CHECK3-NEXT: [[TMP76:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP75]], i32 0, i32 0
// CHECK3-NEXT: [[TMP77:%.*]] = ptrtoint i32* [[ARRAYIDX14]] to i64
// CHECK3-NEXT: store i64 [[TMP77]], i64* [[TMP76]], align 8
// CHECK3-NEXT: [[TMP78:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP75]], i32 0, i32 1
// CHECK3-NEXT: store i64 [[TMP74]], i64* [[TMP78]], align 8
// CHECK3-NEXT: [[TMP79:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP75]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 3, i8* [[TMP79]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i64 2, i64* [[DEP_COUNTER_ADDR17]], align 8
// CHECK3-NEXT: [[TMP80:%.*]] = bitcast %struct.kmp_depend_info* [[TMP59]] to i8*
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM18:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB9]])
// CHECK3-NEXT: [[TMP81:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP58]], i32 0, i32 2
// CHECK3-NEXT: store i32 0, i32* [[TMP81]], align 8
// CHECK3-NEXT: [[TMP82:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM18]], i8* [[TMP56]], i32 2, i8* [[TMP80]], i32 0, i8* null)
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM20:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB11:[0-9]+]])
// CHECK3-NEXT: [[TMP83:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM20]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.7*)* @.omp_task_entry..8 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP84:%.*]] = bitcast i8* [[TMP83]] to %struct.kmp_task_t_with_privates.7*
// CHECK3-NEXT: [[TMP85:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_7:%.*]], %struct.kmp_task_t_with_privates.7* [[TMP84]], i32 0, i32 0
// CHECK3-NEXT: [[TMP86:%.*]] = getelementptr inbounds [2 x %struct.kmp_depend_info], [2 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR21]], i64 0, i64 0
// CHECK3-NEXT: [[ARRAYIDX22:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 0
// CHECK3-NEXT: [[TMP87:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP86]], i64 0
// CHECK3-NEXT: [[TMP88:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i32 0, i32 0
// CHECK3-NEXT: [[TMP89:%.*]] = ptrtoint %struct.S* [[ARRAYIDX22]] to i64
// CHECK3-NEXT: store i64 [[TMP89]], i64* [[TMP88]], align 8
// CHECK3-NEXT: [[TMP90:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i32 0, i32 1
// CHECK3-NEXT: store i64 4, i64* [[TMP90]], align 8
// CHECK3-NEXT: [[TMP91:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 4, i8* [[TMP91]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP92:%.*]] = load i8, i8* [[B]], align 1
// CHECK3-NEXT: [[TMP93:%.*]] = sext i8 [[TMP92]] to i64
// CHECK3-NEXT: [[TMP94:%.*]] = mul nsw i64 4, [[TMP1]]
// CHECK3-NEXT: [[ARRAYIDX23:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP94]]
// CHECK3-NEXT: [[ARRAYIDX24:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX23]], i64 [[TMP93]]
// CHECK3-NEXT: [[TMP95:%.*]] = load i8, i8* [[B]], align 1
// CHECK3-NEXT: [[TMP96:%.*]] = sext i8 [[TMP95]] to i64
// CHECK3-NEXT: [[TMP97:%.*]] = mul nsw i64 9, [[TMP1]]
// CHECK3-NEXT: [[ARRAYIDX25:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP97]]
// CHECK3-NEXT: [[ARRAYIDX26:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX25]], i64 [[TMP96]]
// CHECK3-NEXT: [[TMP98:%.*]] = getelementptr i32, i32* [[ARRAYIDX26]], i32 1
// CHECK3-NEXT: [[TMP99:%.*]] = ptrtoint i32* [[ARRAYIDX24]] to i64
// CHECK3-NEXT: [[TMP100:%.*]] = ptrtoint i32* [[TMP98]] to i64
// CHECK3-NEXT: [[TMP101:%.*]] = sub nuw i64 [[TMP100]], [[TMP99]]
// CHECK3-NEXT: [[TMP102:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP86]], i64 1
// CHECK3-NEXT: [[TMP103:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP102]], i32 0, i32 0
// CHECK3-NEXT: [[TMP104:%.*]] = ptrtoint i32* [[ARRAYIDX24]] to i64
// CHECK3-NEXT: store i64 [[TMP104]], i64* [[TMP103]], align 8
// CHECK3-NEXT: [[TMP105:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP102]], i32 0, i32 1
// CHECK3-NEXT: store i64 [[TMP101]], i64* [[TMP105]], align 8
// CHECK3-NEXT: [[TMP106:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP102]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 4, i8* [[TMP106]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i64 2, i64* [[DEP_COUNTER_ADDR27]], align 8
// CHECK3-NEXT: [[TMP107:%.*]] = bitcast %struct.kmp_depend_info* [[TMP86]] to i8*
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM28:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB11]])
// CHECK3-NEXT: [[TMP108:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP85]], i32 0, i32 2
// CHECK3-NEXT: store i32 0, i32* [[TMP108]], align 8
// CHECK3-NEXT: [[TMP109:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM28]], i8* [[TMP83]], i32 2, i8* [[TMP107]], i32 0, i8* null)
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM30:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB13:[0-9]+]])
// CHECK3-NEXT: [[TMP110:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM30]], i32 3, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.9*)* @.omp_task_entry..10 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP111:%.*]] = bitcast i8* [[TMP110]] to %struct.kmp_task_t_with_privates.9*
// CHECK3-NEXT: [[TMP112:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_9:%.*]], %struct.kmp_task_t_with_privates.9* [[TMP111]], i32 0, i32 0
// CHECK3-NEXT: [[TMP113:%.*]] = getelementptr inbounds [3 x %struct.kmp_depend_info], [3 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR31]], i64 0, i64 0
// CHECK3-NEXT: [[TMP114:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP113]], i64 0
// CHECK3-NEXT: [[TMP115:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i32 0, i32 0
// CHECK3-NEXT: store i64 ptrtoint (i32* @a to i64), i64* [[TMP115]], align 8
// CHECK3-NEXT: [[TMP116:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i32 0, i32 1
// CHECK3-NEXT: store i64 4, i64* [[TMP116]], align 8
// CHECK3-NEXT: [[TMP117:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 3, i8* [[TMP117]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[ARRAYIDX32:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 1
// CHECK3-NEXT: [[TMP118:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP113]], i64 1
// CHECK3-NEXT: [[TMP119:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP118]], i32 0, i32 0
// CHECK3-NEXT: [[TMP120:%.*]] = ptrtoint %struct.S* [[ARRAYIDX32]] to i64
// CHECK3-NEXT: store i64 [[TMP120]], i64* [[TMP119]], align 8
// CHECK3-NEXT: [[TMP121:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP118]], i32 0, i32 1
// CHECK3-NEXT: store i64 4, i64* [[TMP121]], align 8
// CHECK3-NEXT: [[TMP122:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP118]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 3, i8* [[TMP122]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP123:%.*]] = mul nsw i64 0, [[TMP1]]
// CHECK3-NEXT: [[ARRAYIDX33:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP123]]
// CHECK3-NEXT: [[ARRAYIDX34:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX33]], i64 3
// CHECK3-NEXT: [[TMP124:%.*]] = load i32, i32* @a, align 4
// CHECK3-NEXT: [[TMP125:%.*]] = sext i32 [[TMP124]] to i64
// CHECK3-NEXT: [[LEN_SUB_1:%.*]] = sub nsw i64 [[TMP125]], 1
// CHECK3-NEXT: [[TMP126:%.*]] = load i32, i32* @a, align 4
// CHECK3-NEXT: [[TMP127:%.*]] = sext i32 [[TMP126]] to i64
// CHECK3-NEXT: [[LB_ADD_LEN:%.*]] = add nsw i64 -1, [[TMP127]]
// CHECK3-NEXT: [[TMP128:%.*]] = mul nsw i64 [[LB_ADD_LEN]], [[TMP1]]
// CHECK3-NEXT: [[ARRAYIDX35:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP128]]
// CHECK3-NEXT: [[ARRAYIDX36:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX35]], i64 [[LEN_SUB_1]]
// CHECK3-NEXT: [[TMP129:%.*]] = getelementptr i32, i32* [[ARRAYIDX36]], i32 1
// CHECK3-NEXT: [[TMP130:%.*]] = ptrtoint i32* [[ARRAYIDX34]] to i64
// CHECK3-NEXT: [[TMP131:%.*]] = ptrtoint i32* [[TMP129]] to i64
// CHECK3-NEXT: [[TMP132:%.*]] = sub nuw i64 [[TMP131]], [[TMP130]]
// CHECK3-NEXT: [[TMP133:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP113]], i64 2
// CHECK3-NEXT: [[TMP134:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP133]], i32 0, i32 0
// CHECK3-NEXT: [[TMP135:%.*]] = ptrtoint i32* [[ARRAYIDX34]] to i64
// CHECK3-NEXT: store i64 [[TMP135]], i64* [[TMP134]], align 8
// CHECK3-NEXT: [[TMP136:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP133]], i32 0, i32 1
// CHECK3-NEXT: store i64 [[TMP132]], i64* [[TMP136]], align 8
// CHECK3-NEXT: [[TMP137:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP133]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK3-NEXT: store i8 3, i8* [[TMP137]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i64 3, i64* [[DEP_COUNTER_ADDR37]], align 8
// CHECK3-NEXT: [[TMP138:%.*]] = bitcast %struct.kmp_depend_info* [[TMP113]] to i8*
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM38:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB13]])
// CHECK3-NEXT: [[TMP139:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM38]], i8* [[TMP110]], i32 3, i8* [[TMP138]], i32 0, i8* null)
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM40:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB15:[0-9]+]])
// CHECK3-NEXT: [[TMP140:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM40]], i32 3, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.11*)* @.omp_task_entry..12 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP141:%.*]] = bitcast i8* [[TMP140]] to %struct.kmp_task_t_with_privates.11*
// CHECK3-NEXT: [[TMP142:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_11:%.*]], %struct.kmp_task_t_with_privates.11* [[TMP141]], i32 0, i32 0
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM41:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB15]])
// CHECK3-NEXT: [[TMP143:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM41]], i8* [[TMP140]])
// CHECK3-NEXT: store i8 0, i8* [[FLAG]], align 1
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM43:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB17:[0-9]+]])
// CHECK3-NEXT: [[TMP144:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM43]], i32 1, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.13*)* @.omp_task_entry..14 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP145:%.*]] = bitcast i8* [[TMP144]] to %struct.kmp_task_t_with_privates.13*
// CHECK3-NEXT: [[TMP146:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_13:%.*]], %struct.kmp_task_t_with_privates.13* [[TMP145]], i32 0, i32 0
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM44:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB17]])
// CHECK3-NEXT: [[TMP147:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM44]], i8* [[TMP144]])
// CHECK3-NEXT: [[TMP148:%.*]] = getelementptr inbounds [[STRUCT_ANON_14]], %struct.anon.14* [[AGG_CAPTURED45]], i32 0, i32 0
// CHECK3-NEXT: store i32* [[C]], i32** [[TMP148]], align 8
// CHECK3-NEXT: [[TMP149:%.*]] = load i8, i8* [[B]], align 1
// CHECK3-NEXT: [[TOBOOL:%.*]] = icmp ne i8 [[TMP149]], 0
// CHECK3-NEXT: [[TMP150:%.*]] = select i1 [[TOBOOL]], i32 2, i32 0
// CHECK3-NEXT: [[TMP151:%.*]] = or i32 [[TMP150]], 1
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM46:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB19:[0-9]+]])
// CHECK3-NEXT: [[TMP152:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM46]], i32 [[TMP151]], i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.15*)* @.omp_task_entry..16 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP153:%.*]] = bitcast i8* [[TMP152]] to %struct.kmp_task_t_with_privates.15*
// CHECK3-NEXT: [[TMP154:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_15:%.*]], %struct.kmp_task_t_with_privates.15* [[TMP153]], i32 0, i32 0
// CHECK3-NEXT: [[TMP155:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP154]], i32 0, i32 0
// CHECK3-NEXT: [[TMP156:%.*]] = load i8*, i8** [[TMP155]], align 8
// CHECK3-NEXT: [[TMP157:%.*]] = bitcast %struct.anon.14* [[AGG_CAPTURED45]] to i8*
// CHECK3-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP156]], i8* align 8 [[TMP157]], i64 8, i1 false)
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM47:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB19]])
// CHECK3-NEXT: [[TMP158:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM47]], i8* [[TMP152]])
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM49:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21:[0-9]+]])
// CHECK3-NEXT: [[TMP159:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM49]], i32 0, i64 256, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.19*)* @.omp_task_entry..21 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP160:%.*]] = bitcast i8* [[TMP159]] to %struct.kmp_task_t_with_privates.19*
// CHECK3-NEXT: [[TMP161:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19:%.*]], %struct.kmp_task_t_with_privates.19* [[TMP160]], i32 0, i32 0
// CHECK3-NEXT: [[TMP162:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19]], %struct.kmp_task_t_with_privates.19* [[TMP160]], i32 0, i32 2
// CHECK3-NEXT: [[TMP163:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20:%.*]], %struct..kmp_privates.t.20* [[TMP162]], i32 0, i32 0
// CHECK3-NEXT: [[TMP164:%.*]] = load i32, i32* [[C]], align 128
// CHECK3-NEXT: store i32 [[TMP164]], i32* [[TMP163]], align 128
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM50:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]])
// CHECK3-NEXT: [[TMP165:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP161]], i32 0, i32 2
// CHECK3-NEXT: store i32 0, i32* [[TMP165]], align 16
// CHECK3-NEXT: [[TMP166:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM50]], i8* [[TMP159]])
// CHECK3-NEXT: [[TMP167:%.*]] = load i32, i32* @a, align 4
// CHECK3-NEXT: store i32 [[TMP167]], i32* [[RETVAL]], align 4
// CHECK3-NEXT: [[TMP168:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
// CHECK3-NEXT: call void @llvm.stackrestore(i8* [[TMP168]])
// CHECK3-NEXT: [[ARRAY_BEGIN51:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i32 0, i32 0
// CHECK3-NEXT: [[TMP169:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN51]], i64 2
// CHECK3-NEXT: br label [[ARRAYDESTROY_BODY:%.*]]
// CHECK3: arraydestroy.body:
// CHECK3-NEXT: [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP169]], [[ARRAYCTOR_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
// CHECK3-NEXT: [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4:[0-9]+]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN51]]
// CHECK3-NEXT: br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE52:%.*]], label [[ARRAYDESTROY_BODY]]
// CHECK3: arraydestroy.done52:
// CHECK3-NEXT: [[TMP170:%.*]] = load i32, i32* [[RETVAL]], align 4
// CHECK3-NEXT: ret i32 [[TMP170]]
//
//
// CHECK3-LABEL: define {{[^@]+}}@_ZN1SC1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1:[0-9]+]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK3-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK3-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN1SC2Ev(%struct.S* noundef [[THIS1]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry.
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias noundef [[TMP1:%.*]]) #[[ATTR3:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META3:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META6:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META8:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META10:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !12
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !12
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !12
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !12
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !12
// CHECK3-NEXT: store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !12
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !12
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 15, i32* @a, align 4
// CHECK3-NEXT: [[TMP11:%.*]] = load i32, i32* @a, align 4
// CHECK3-NEXT: [[CONV_I:%.*]] = trunc i32 [[TMP11]] to i8
// CHECK3-NEXT: [[TMP12:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP10]], i32 0, i32 0
// CHECK3-NEXT: [[TMP13:%.*]] = load i8*, i8** [[TMP12]], align 8
// CHECK3-NEXT: store i8 [[CONV_I]], i8* [[TMP13]], align 1
// CHECK3-NEXT: [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP10]], i32 0, i32 1
// CHECK3-NEXT: [[TMP15:%.*]] = load [2 x %struct.S]*, [2 x %struct.S]** [[TMP14]], align 8
// CHECK3-NEXT: [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[TMP15]], i64 0, i64 0
// CHECK3-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAYIDX_I]], i32 0, i32 0
// CHECK3-NEXT: store i32 10, i32* [[A_I]], align 4
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..2
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.1* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.0*, align 8
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.1*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.1* [[TMP1]], %struct.kmp_task_t_with_privates.1** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.1*, %struct.kmp_task_t_with_privates.1** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_1:%.*]], %struct.kmp_task_t_with_privates.1* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.0*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.1* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META13:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META16:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META18:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META20:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !22
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !22
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !22
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !22
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !22
// CHECK3-NEXT: store %struct.anon.0* [[TMP8]], %struct.anon.0** [[__CONTEXT_ADDR_I]], align 8, !noalias !22
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR_I]], align 8, !noalias !22
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 15, i32* @a, align 4
// CHECK3-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP10]], i32 0, i32 0
// CHECK3-NEXT: [[TMP12:%.*]] = load [2 x %struct.S]*, [2 x %struct.S]** [[TMP11]], align 8
// CHECK3-NEXT: [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[TMP12]], i64 0, i64 1
// CHECK3-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAYIDX_I]], i32 0, i32 0
// CHECK3-NEXT: store i32 10, i32* [[A_I]], align 4
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..4
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.3* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.2*, align 8
// CHECK3-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.3*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.3* [[TMP1]], %struct.kmp_task_t_with_privates.3** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.3*, %struct.kmp_task_t_with_privates.3** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_3:%.*]], %struct.kmp_task_t_with_privates.3* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.2*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.3* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META23:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META26:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META28:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META30:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !32
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !32
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !32
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !32
// CHECK3-NEXT: store %struct.anon.2* [[TMP8]], %struct.anon.2** [[__CONTEXT_ADDR_I]], align 8, !noalias !32
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon.2*, %struct.anon.2** [[__CONTEXT_ADDR_I]], align 8, !noalias !32
// CHECK3-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK3-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK3-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK3-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK3-NEXT: ]
// CHECK3: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK3: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB7]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !32
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM_I]], i8* [[TMP14]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
// CHECK3: .untied.jmp.1.i:
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM2_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) #[[ATTR4]]
// CHECK3-NEXT: call void @__kmpc_critical(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM2_I]], [8 x i32]* @.gomp_critical_user_.var) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: call void @__kmpc_end_critical(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM2_I]], [8 x i32]* @.gomp_critical_user_.var) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[CLEANUP_I]]
// CHECK3: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__3_EXIT]]
// CHECK3: .omp_outlined..3.exit:
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..6
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.5* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.4*, align 8
// CHECK3-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.5*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.5* [[TMP1]], %struct.kmp_task_t_with_privates.5** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.5*, %struct.kmp_task_t_with_privates.5** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_5:%.*]], %struct.kmp_task_t_with_privates.5* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.4*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.5* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META33:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META36:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META38:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META40:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !42
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !42
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !42
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !42
// CHECK3-NEXT: store %struct.anon.4* [[TMP8]], %struct.anon.4** [[__CONTEXT_ADDR_I]], align 8, !noalias !42
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon.4*, %struct.anon.4** [[__CONTEXT_ADDR_I]], align 8, !noalias !42
// CHECK3-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK3-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK3-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK3-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK3-NEXT: ]
// CHECK3: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK3: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB9]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !42
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM_I]], i8* [[TMP14]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__5_EXIT:%.*]]
// CHECK3: .untied.jmp.1.i:
// CHECK3-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[CLEANUP_I]]
// CHECK3: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__5_EXIT]]
// CHECK3: .omp_outlined..5.exit:
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.7* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.6*, align 8
// CHECK3-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.7*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.7* [[TMP1]], %struct.kmp_task_t_with_privates.7** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.7*, %struct.kmp_task_t_with_privates.7** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_7:%.*]], %struct.kmp_task_t_with_privates.7* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.6*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.7* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META43:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META46:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META48:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META50:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !52
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !52
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !52
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !52
// CHECK3-NEXT: store %struct.anon.6* [[TMP8]], %struct.anon.6** [[__CONTEXT_ADDR_I]], align 8, !noalias !52
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon.6*, %struct.anon.6** [[__CONTEXT_ADDR_I]], align 8, !noalias !52
// CHECK3-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK3-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK3-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK3-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK3-NEXT: ]
// CHECK3: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK3: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB11]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !52
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM_I]], i8* [[TMP14]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__7_EXIT:%.*]]
// CHECK3: .untied.jmp.1.i:
// CHECK3-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[CLEANUP_I]]
// CHECK3: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__7_EXIT]]
// CHECK3: .omp_outlined..7.exit:
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..10
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.9* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.8*, align 8
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.9*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.9* [[TMP1]], %struct.kmp_task_t_with_privates.9** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.9*, %struct.kmp_task_t_with_privates.9** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_9:%.*]], %struct.kmp_task_t_with_privates.9* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.8*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.9* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META53:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META56:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META58:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META60:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !62
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !62
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !62
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !62
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !62
// CHECK3-NEXT: store %struct.anon.8* [[TMP8]], %struct.anon.8** [[__CONTEXT_ADDR_I]], align 8, !noalias !62
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon.8*, %struct.anon.8** [[__CONTEXT_ADDR_I]], align 8, !noalias !62
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 2, i32* @a, align 4
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..12
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.11* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.10*, align 8
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.11*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.11* [[TMP1]], %struct.kmp_task_t_with_privates.11** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.11*, %struct.kmp_task_t_with_privates.11** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_11:%.*]], %struct.kmp_task_t_with_privates.11* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.10*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.11* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META63:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META66:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META68:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META70:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !72
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !72
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !72
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !72
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !72
// CHECK3-NEXT: store %struct.anon.10* [[TMP8]], %struct.anon.10** [[__CONTEXT_ADDR_I]], align 8, !noalias !72
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon.10*, %struct.anon.10** [[__CONTEXT_ADDR_I]], align 8, !noalias !72
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 2, i32* @a, align 4
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..14
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.13* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.12*, align 8
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.13*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.13* [[TMP1]], %struct.kmp_task_t_with_privates.13** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.13*, %struct.kmp_task_t_with_privates.13** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_13:%.*]], %struct.kmp_task_t_with_privates.13* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.12*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.13* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META73:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META76:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META78:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META80:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !82
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !82
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !82
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !82
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !82
// CHECK3-NEXT: store %struct.anon.12* [[TMP8]], %struct.anon.12** [[__CONTEXT_ADDR_I]], align 8, !noalias !82
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon.12*, %struct.anon.12** [[__CONTEXT_ADDR_I]], align 8, !noalias !82
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 3, i32* @a, align 4
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..16
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.15* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.14*, align 8
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.15*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.15* [[TMP1]], %struct.kmp_task_t_with_privates.15** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.15*, %struct.kmp_task_t_with_privates.15** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_15:%.*]], %struct.kmp_task_t_with_privates.15* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.14*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.15* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META83:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META86:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META88:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META90:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !92
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !92
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !92
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !92
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !92
// CHECK3-NEXT: store %struct.anon.14* [[TMP8]], %struct.anon.14** [[__CONTEXT_ADDR_I]], align 8, !noalias !92
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon.14*, %struct.anon.14** [[__CONTEXT_ADDR_I]], align 8, !noalias !92
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 4, i32* @a, align 4
// CHECK3-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_14:%.*]], %struct.anon.14* [[TMP10]], i32 0, i32 0
// CHECK3-NEXT: [[TMP12:%.*]] = load i32*, i32** [[TMP11]], align 8
// CHECK3-NEXT: store i32 5, i32* [[TMP12]], align 128
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_privates_map.
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (%struct..kmp_privates.t* noalias noundef [[TMP0:%.*]], i32** noalias noundef [[TMP1:%.*]]) #[[ATTR7:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 8
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca i32**, align 8
// CHECK3-NEXT: store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 8
// CHECK3-NEXT: store i32** [[TMP1]], i32*** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 8
// CHECK3-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP2]], i32 0, i32 0
// CHECK3-NEXT: [[TMP4:%.*]] = load i32**, i32*** [[DOTADDR1]], align 8
// CHECK3-NEXT: store i32* [[TMP3]], i32** [[TMP4]], align 8
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..19
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.18* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.17*, align 8
// CHECK3-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.18*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.18* [[TMP1]], %struct.kmp_task_t_with_privates.18** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.18*, %struct.kmp_task_t_with_privates.18** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18:%.*]], %struct.kmp_task_t_with_privates.18* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 128
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.17*
// CHECK3-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18]], %struct.kmp_task_t_with_privates.18* [[TMP3]], i32 0, i32 2
// CHECK3-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
// CHECK3-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.18* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META93:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META96:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META98:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META100:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !102
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !102
// CHECK3-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !102
// CHECK3-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i32**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !102
// CHECK3-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !102
// CHECK3-NEXT: store %struct.anon.17* [[TMP8]], %struct.anon.17** [[__CONTEXT_ADDR_I]], align 8, !noalias !102
// CHECK3-NEXT: [[TMP12:%.*]] = load %struct.anon.17*, %struct.anon.17** [[__CONTEXT_ADDR_I]], align 8, !noalias !102
// CHECK3-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !102
// CHECK3-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !102
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i32**)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: call void [[TMP15]](i8* [[TMP14]], i32** [[DOTFIRSTPRIV_PTR_ADDR_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP16:%.*]] = load i32*, i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !102
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 4, i32* [[TMP16]], align 128
// CHECK3-NEXT: store i32 4, i32* @a, align 4
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@_ZN1SD1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK3-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK3-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN1SD2Ev(%struct.S* noundef [[THIS1]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_privates_map..20
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (%struct..kmp_privates.t.20* noalias noundef [[TMP0:%.*]], i32** noalias noundef [[TMP1:%.*]], %struct.S** noalias noundef [[TMP2:%.*]], %struct.S** noalias noundef [[TMP3:%.*]]) #[[ATTR7]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t.20*, align 8
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca i32**, align 8
// CHECK3-NEXT: [[DOTADDR2:%.*]] = alloca %struct.S**, align 8
// CHECK3-NEXT: [[DOTADDR3:%.*]] = alloca %struct.S**, align 8
// CHECK3-NEXT: store %struct..kmp_privates.t.20* [[TMP0]], %struct..kmp_privates.t.20** [[DOTADDR]], align 8
// CHECK3-NEXT: store i32** [[TMP1]], i32*** [[DOTADDR1]], align 8
// CHECK3-NEXT: store %struct.S** [[TMP2]], %struct.S*** [[DOTADDR2]], align 8
// CHECK3-NEXT: store %struct.S** [[TMP3]], %struct.S*** [[DOTADDR3]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = load %struct..kmp_privates.t.20*, %struct..kmp_privates.t.20** [[DOTADDR]], align 8
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20:%.*]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP6:%.*]] = load i32**, i32*** [[DOTADDR1]], align 8
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP7:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 1
// CHECK3-NEXT: [[TMP8:%.*]] = load %struct.S**, %struct.S*** [[DOTADDR2]], align 8
// CHECK3-NEXT: store %struct.S* [[TMP7]], %struct.S** [[TMP8]], align 8
// CHECK3-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.S**, %struct.S*** [[DOTADDR3]], align 8
// CHECK3-NEXT: store %struct.S* [[TMP9]], %struct.S** [[TMP10]], align 8
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..21
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.19* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.16*, align 8
// CHECK3-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTLOCAL_PTR_ADDR_I:%.*]] = alloca %struct.S*, align 8
// CHECK3-NEXT: [[DOTLOCAL_PTR_ADDR1_I:%.*]] = alloca %struct.S*, align 8
// CHECK3-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[S1_I:%.*]] = alloca [[STRUCT_S:%.*]], align 4
// CHECK3-NEXT: [[S2_I:%.*]] = alloca [[STRUCT_S]], align 4
// CHECK3-NEXT: [[REF_TMP_I:%.*]] = alloca [[STRUCT_S]], align 4
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.19*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.19* [[TMP1]], %struct.kmp_task_t_with_privates.19** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.19*, %struct.kmp_task_t_with_privates.19** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19:%.*]], %struct.kmp_task_t_with_privates.19* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 128
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.16*
// CHECK3-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19]], %struct.kmp_task_t_with_privates.19* [[TMP3]], i32 0, i32 2
// CHECK3-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t.20* [[TMP9]] to i8*
// CHECK3-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.19* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META103:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META106:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META108:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META110:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
// CHECK3-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !112
// CHECK3-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t.20*, i32**, %struct.S**, %struct.S**)* @.omp_task_privates_map..20 to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !112
// CHECK3-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
// CHECK3-NEXT: store %struct.anon.16* [[TMP8]], %struct.anon.16** [[__CONTEXT_ADDR_I]], align 8, !noalias !112
// CHECK3-NEXT: [[TMP12:%.*]] = load %struct.anon.16*, %struct.anon.16** [[__CONTEXT_ADDR_I]], align 8, !noalias !112
// CHECK3-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !112
// CHECK3-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i32**, %struct.S**, %struct.S**)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: call void [[TMP15]](i8* [[TMP14]], i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], %struct.S** [[DOTLOCAL_PTR_ADDR_I]], %struct.S** [[DOTLOCAL_PTR_ADDR1_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP16:%.*]] = load i32*, i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !112
// CHECK3-NEXT: [[TMP17:%.*]] = load %struct.S*, %struct.S** [[DOTLOCAL_PTR_ADDR_I]], align 8, !noalias !112
// CHECK3-NEXT: [[TMP18:%.*]] = load %struct.S*, %struct.S** [[DOTLOCAL_PTR_ADDR1_I]], align 8, !noalias !112
// CHECK3-NEXT: [[TMP19:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP20:%.*]] = load i32, i32* [[TMP19]], align 4
// CHECK3-NEXT: switch i32 [[TMP20]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK3-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK3-NEXT: i32 1, label [[DOTUNTIED_JMP_2_I:%.*]]
// CHECK3-NEXT: i32 2, label [[DOTUNTIED_JMP_6_I:%.*]]
// CHECK3-NEXT: i32 3, label [[DOTUNTIED_JMP_10_I:%.*]]
// CHECK3-NEXT: i32 4, label [[DOTUNTIED_JMP_15_I:%.*]]
// CHECK3-NEXT: ]
// CHECK3: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK3: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP21:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 1, i32* [[TMP21]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP22:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[TMP23:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM_I]], i8* [[TMP22]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT:%.*]]
// CHECK3: .untied.jmp.2.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[S1_I]]) #[[ATTR4]]
// CHECK3-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[S2_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[S2_I]], i32 0, i32 0
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 0, i32* [[A_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM3_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB23:[0-9]+]]) #[[ATTR4]]
// CHECK3-NEXT: [[TMP24:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM3_I]], i32 1, i64 256, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.18*)* @.omp_task_entry..19 to i32 (i32, i8*)*)) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP25:%.*]] = bitcast i8* [[TMP24]] to %struct.kmp_task_t_with_privates.18*
// CHECK3-NEXT: [[TMP26:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18:%.*]], %struct.kmp_task_t_with_privates.18* [[TMP25]], i32 0, i32 0
// CHECK3-NEXT: [[TMP27:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18]], %struct.kmp_task_t_with_privates.18* [[TMP25]], i32 0, i32 2
// CHECK3-NEXT: [[TMP28:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP27]], i32 0, i32 0
// CHECK3-NEXT: [[TMP29:%.*]] = load i32, i32* [[TMP16]], align 128
// CHECK3-NEXT: store i32 [[TMP29]], i32* [[TMP28]], align 128
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM4_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB23]]) #[[ATTR4]]
// CHECK3-NEXT: [[TMP30:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM4_I]], i8* [[TMP24]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP31:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 2, i32* [[TMP31]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM5_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP32:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[TMP33:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM5_I]], i8* [[TMP32]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK3: .untied.jmp.6.i:
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM8_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) #[[ATTR4]]
// CHECK3-NEXT: [[TMP34:%.*]] = call i32 @__kmpc_omp_taskyield(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM8_I]], i32 0) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP35:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 3, i32* [[TMP35]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM9_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP36:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[TMP37:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM9_I]], i8* [[TMP36]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK3: .untied.jmp.10.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[REF_TMP_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP38:%.*]] = bitcast %struct.S* [[S1_I]] to i8*
// CHECK3-NEXT: [[TMP39:%.*]] = bitcast %struct.S* [[REF_TMP_I]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP38]], i8* align 4 [[TMP39]], i64 4, i1 false) #[[ATTR4]], !noalias !112
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[REF_TMP_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[A12_I:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[S2_I]], i32 0, i32 0
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 10, i32* [[A12_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM13_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) #[[ATTR4]]
// CHECK3-NEXT: [[TMP40:%.*]] = call i32 @__kmpc_omp_taskwait(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM13_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP41:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: store i32 4, i32* [[TMP41]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM14_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[TMP42:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: [[TMP43:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM14_I]], i8* [[TMP42]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK3: .untied.jmp.15.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[S2_I]]) #[[ATTR4]]
// CHECK3-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[S1_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[CLEANUP_I]]
// CHECK3: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK3: .omp_outlined..17.exit:
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@_ZN1SC2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK3-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK3-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
// CHECK3-NEXT: [[A:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[THIS1]], i32 0, i32 0
// CHECK3-NEXT: store i32 0, i32* [[A]], align 4
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@_ZN1SD2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK3-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK3-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
// CHECK3-NEXT: ret void
//
//
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-LABEL: define {{[^@]+}}@__cxx_global_var_init
// CHECK3-SAME: () #[[ATTR7]] section "__TEXT,__StaticInit,regular,pure_instructions" {
// CHECK3-NEXT: entry:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN2S1C1Ev(%struct.S1* noundef @s1)
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@_ZN2S1C1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (%struct.S1* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK3-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK3-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN2S1C2Ev(%struct.S1* noundef [[THIS1]])
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-NEXT: ret void
//
//
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-LABEL: define {{[^@]+}}@_ZN2S1C2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (%struct.S1* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK3-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK3-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-NEXT: call void @_ZN2S18taskinitEv(%struct.S1* noundef [[THIS1]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@_ZN2S18taskinitEv
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (%struct.S1* noundef [[THIS:%.*]]) #[[ATTR8:[0-9]+]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK3-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON_21:%.*]], align 8
// CHECK3-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK3-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
// CHECK3-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_21]], %struct.anon.21* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK3-NEXT: store %struct.S1* [[THIS1]], %struct.S1** [[TMP0]], align 8
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB25:[0-9]+]])
// CHECK3-NEXT: [[TMP1:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM]], i32 1, i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.22*)* @.omp_task_entry..23 to i32 (i32, i8*)*))
// CHECK3-NEXT: [[TMP2:%.*]] = bitcast i8* [[TMP1]] to %struct.kmp_task_t_with_privates.22*
// CHECK3-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_22:%.*]], %struct.kmp_task_t_with_privates.22* [[TMP2]], i32 0, i32 0
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = load i8*, i8** [[TMP4]], align 8
// CHECK3-NEXT: [[TMP6:%.*]] = bitcast %struct.anon.21* [[AGG_CAPTURED]] to i8*
// CHECK3-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP5]], i8* align 8 [[TMP6]], i64 8, i1 false)
// CHECK3-NEXT: [[OMP_GLOBAL_THREAD_NUM2:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB25]])
// CHECK3-NEXT: [[TMP7:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM2]], i8* [[TMP1]])
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_task_entry..23
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK3-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.22* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK3-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK3-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.21*, align 8
// CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.22*, align 8
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK3-NEXT: store %struct.kmp_task_t_with_privates.22* [[TMP1]], %struct.kmp_task_t_with_privates.22** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.22*, %struct.kmp_task_t_with_privates.22** [[DOTADDR1]], align 8
// CHECK3-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_22:%.*]], %struct.kmp_task_t_with_privates.22* [[TMP3]], i32 0, i32 0
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK3-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.21*
// CHECK3-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.22* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META113:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META116:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META118:![0-9]+]])
// CHECK3-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META120:![0-9]+]])
// CHECK3-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !122
// CHECK3-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !122
// CHECK3-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !122
// CHECK3-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !122
// CHECK3-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !122
// CHECK3-NEXT: store %struct.anon.21* [[TMP8]], %struct.anon.21** [[__CONTEXT_ADDR_I]], align 8, !noalias !122
// CHECK3-NEXT: [[TMP10:%.*]] = load %struct.anon.21*, %struct.anon.21** [[__CONTEXT_ADDR_I]], align 8, !noalias !122
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_21:%.*]], %struct.anon.21* [[TMP10]], i32 0, i32 0
// CHECK3-NEXT: [[TMP12:%.*]] = load %struct.S1*, %struct.S1** [[TMP11]], align 8
// CHECK3-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP12]], i32 0, i32 0
// CHECK3-NEXT: store i32 0, i32* [[A_I]], align 4
// CHECK3-NEXT: ret i32 0
//
//
// CHECK3-LABEL: define {{[^@]+}}@_GLOBAL__sub_I_task_codegen.cpp
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK3-SAME: () #[[ATTR7]] section "__TEXT,__StaticInit,regular,pure_instructions" {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: call void @__cxx_global_var_init()
// CHECK3-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@main
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-SAME: () #[[ATTR0:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[RETVAL:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[B:%.*]] = alloca i8, align 1
// CHECK4-NEXT: [[S:%.*]] = alloca [2 x %struct.S], align 4
// CHECK4-NEXT: [[SAVED_STACK:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__VLA_EXPR0:%.*]] = alloca i64, align 8
// CHECK4-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 8
// CHECK4-NEXT: [[AGG_CAPTURED2:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 8
// CHECK4-NEXT: [[DOTDEP_ARR_ADDR:%.*]] = alloca [4 x %struct.kmp_depend_info], align 8
// CHECK4-NEXT: [[DEP_COUNTER_ADDR:%.*]] = alloca i64, align 8
// CHECK4-NEXT: [[AGG_CAPTURED6:%.*]] = alloca [[STRUCT_ANON_2:%.*]], align 1
// CHECK4-NEXT: [[AGG_CAPTURED9:%.*]] = alloca [[STRUCT_ANON_4:%.*]], align 1
// CHECK4-NEXT: [[DOTDEP_ARR_ADDR11:%.*]] = alloca [2 x %struct.kmp_depend_info], align 8
// CHECK4-NEXT: [[DEP_COUNTER_ADDR17:%.*]] = alloca i64, align 8
// CHECK4-NEXT: [[AGG_CAPTURED19:%.*]] = alloca [[STRUCT_ANON_6:%.*]], align 1
// CHECK4-NEXT: [[DOTDEP_ARR_ADDR21:%.*]] = alloca [2 x %struct.kmp_depend_info], align 8
// CHECK4-NEXT: [[DEP_COUNTER_ADDR27:%.*]] = alloca i64, align 8
// CHECK4-NEXT: [[AGG_CAPTURED29:%.*]] = alloca [[STRUCT_ANON_8:%.*]], align 1
// CHECK4-NEXT: [[DOTDEP_ARR_ADDR31:%.*]] = alloca [3 x %struct.kmp_depend_info], align 8
// CHECK4-NEXT: [[DEP_COUNTER_ADDR37:%.*]] = alloca i64, align 8
// CHECK4-NEXT: [[AGG_CAPTURED39:%.*]] = alloca [[STRUCT_ANON_10:%.*]], align 1
// CHECK4-NEXT: [[FLAG:%.*]] = alloca i8, align 1
// CHECK4-NEXT: [[AGG_CAPTURED42:%.*]] = alloca [[STRUCT_ANON_12:%.*]], align 1
// CHECK4-NEXT: [[C:%.*]] = alloca i32, align 128
// CHECK4-NEXT: [[AGG_CAPTURED45:%.*]] = alloca [[STRUCT_ANON_14:%.*]], align 8
// CHECK4-NEXT: [[AGG_CAPTURED48:%.*]] = alloca [[STRUCT_ANON_16:%.*]], align 1
// CHECK4-NEXT: store i32 0, i32* [[RETVAL]], align 4
// CHECK4-NEXT: [[ARRAY_BEGIN:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i32 0, i32 0
// CHECK4-NEXT: [[ARRAYCTOR_END:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAY_BEGIN]], i64 2
// CHECK4-NEXT: br label [[ARRAYCTOR_LOOP:%.*]]
// CHECK4: arrayctor.loop:
// CHECK4-NEXT: [[ARRAYCTOR_CUR:%.*]] = phi %struct.S* [ [[ARRAY_BEGIN]], [[ENTRY:%.*]] ], [ [[ARRAYCTOR_NEXT:%.*]], [[ARRAYCTOR_LOOP]] ]
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[ARRAYCTOR_CUR]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[ARRAYCTOR_NEXT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYCTOR_CUR]], i64 1
// CHECK4-NEXT: [[ARRAYCTOR_DONE:%.*]] = icmp eq %struct.S* [[ARRAYCTOR_NEXT]], [[ARRAYCTOR_END]]
// CHECK4-NEXT: br i1 [[ARRAYCTOR_DONE]], label [[ARRAYCTOR_CONT:%.*]], label [[ARRAYCTOR_LOOP]]
// CHECK4: arrayctor.cont:
// CHECK4-NEXT: [[TMP0:%.*]] = load i32, i32* @a, align 4
// CHECK4-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
// CHECK4-NEXT: [[TMP2:%.*]] = call i8* @llvm.stacksave()
// CHECK4-NEXT: store i8* [[TMP2]], i8** [[SAVED_STACK]], align 8
// CHECK4-NEXT: [[TMP3:%.*]] = mul nuw i64 10, [[TMP1]]
// CHECK4-NEXT: [[VLA:%.*]] = alloca i32, i64 [[TMP3]], align 16
// CHECK4-NEXT: store i64 [[TMP1]], i64* [[__VLA_EXPR0]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK4-NEXT: store i8* [[B]], i8** [[TMP4]], align 8
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[AGG_CAPTURED]], i32 0, i32 1
// CHECK4-NEXT: store [2 x %struct.S]* [[S]], [2 x %struct.S]** [[TMP5]], align 8
// CHECK4-NEXT: [[TMP6:%.*]] = load i8, i8* [[B]], align 1
// CHECK4-NEXT: [[CONV:%.*]] = sext i8 [[TMP6]] to i32
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB3:[0-9]+]])
// CHECK4-NEXT: [[TMP7:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[OMP_GLOBAL_THREAD_NUM]], i32 33, i64 40, i64 16, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @.omp_task_entry. to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.kmp_task_t_with_privates*
// CHECK4-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP8]], i32 0, i32 0
// CHECK4-NEXT: [[TMP10:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP9]], i32 0, i32 0
// CHECK4-NEXT: [[TMP11:%.*]] = load i8*, i8** [[TMP10]], align 8
// CHECK4-NEXT: [[TMP12:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8*
// CHECK4-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP11]], i8* align 8 [[TMP12]], i64 16, i1 false)
// CHECK4-NEXT: [[TMP13:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP9]], i32 0, i32 4
// CHECK4-NEXT: [[TMP14:%.*]] = bitcast %union.kmp_cmplrdata_t* [[TMP13]] to i32*
// CHECK4-NEXT: store i32 [[CONV]], i32* [[TMP14]], align 8
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB3]])
// CHECK4-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM1]], i8* [[TMP7]])
// CHECK4-NEXT: [[TMP16:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[AGG_CAPTURED2]], i32 0, i32 0
// CHECK4-NEXT: store [2 x %struct.S]* [[S]], [2 x %struct.S]** [[TMP16]], align 8
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM3:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB5:[0-9]+]])
// CHECK4-NEXT: [[TMP17:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM3]], i32 1, i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.1*)* @.omp_task_entry..2 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP18:%.*]] = bitcast i8* [[TMP17]] to %struct.kmp_task_t_with_privates.1*
// CHECK4-NEXT: [[TMP19:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_1:%.*]], %struct.kmp_task_t_with_privates.1* [[TMP18]], i32 0, i32 0
// CHECK4-NEXT: [[TMP20:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP19]], i32 0, i32 0
// CHECK4-NEXT: [[TMP21:%.*]] = load i8*, i8** [[TMP20]], align 8
// CHECK4-NEXT: [[TMP22:%.*]] = bitcast %struct.anon.0* [[AGG_CAPTURED2]] to i8*
// CHECK4-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP21]], i8* align 8 [[TMP22]], i64 8, i1 false)
// CHECK4-NEXT: [[TMP23:%.*]] = getelementptr inbounds [4 x %struct.kmp_depend_info], [4 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR]], i64 0, i64 0
// CHECK4-NEXT: [[TMP24:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO:%.*]], %struct.kmp_depend_info* [[TMP23]], i64 0
// CHECK4-NEXT: [[TMP25:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i32 0, i32 0
// CHECK4-NEXT: store i64 ptrtoint (i32* @a to i64), i64* [[TMP25]], align 8
// CHECK4-NEXT: [[TMP26:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i32 0, i32 1
// CHECK4-NEXT: store i64 4, i64* [[TMP26]], align 8
// CHECK4-NEXT: [[TMP27:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP24]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 1, i8* [[TMP27]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP28:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP23]], i64 1
// CHECK4-NEXT: [[TMP29:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP28]], i32 0, i32 0
// CHECK4-NEXT: [[TMP30:%.*]] = ptrtoint i8* [[B]] to i64
// CHECK4-NEXT: store i64 [[TMP30]], i64* [[TMP29]], align 8
// CHECK4-NEXT: [[TMP31:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP28]], i32 0, i32 1
// CHECK4-NEXT: store i64 1, i64* [[TMP31]], align 8
// CHECK4-NEXT: [[TMP32:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP28]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 1, i8* [[TMP32]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP33:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP23]], i64 2
// CHECK4-NEXT: [[TMP34:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP33]], i32 0, i32 0
// CHECK4-NEXT: [[TMP35:%.*]] = ptrtoint [2 x %struct.S]* [[S]] to i64
// CHECK4-NEXT: store i64 [[TMP35]], i64* [[TMP34]], align 8
// CHECK4-NEXT: [[TMP36:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP33]], i32 0, i32 1
// CHECK4-NEXT: store i64 8, i64* [[TMP36]], align 8
// CHECK4-NEXT: [[TMP37:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP33]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 1, i8* [[TMP37]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP38:%.*]] = mul nsw i64 0, [[TMP1]]
// CHECK4-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP38]]
// CHECK4-NEXT: [[TMP39:%.*]] = mul nsw i64 9, [[TMP1]]
// CHECK4-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP39]]
// CHECK4-NEXT: [[TMP40:%.*]] = getelementptr i32, i32* [[ARRAYIDX4]], i32 1
// CHECK4-NEXT: [[TMP41:%.*]] = ptrtoint i32* [[ARRAYIDX]] to i64
// CHECK4-NEXT: [[TMP42:%.*]] = ptrtoint i32* [[TMP40]] to i64
// CHECK4-NEXT: [[TMP43:%.*]] = sub nuw i64 [[TMP42]], [[TMP41]]
// CHECK4-NEXT: [[TMP44:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP23]], i64 3
// CHECK4-NEXT: [[TMP45:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP44]], i32 0, i32 0
// CHECK4-NEXT: [[TMP46:%.*]] = ptrtoint i32* [[ARRAYIDX]] to i64
// CHECK4-NEXT: store i64 [[TMP46]], i64* [[TMP45]], align 8
// CHECK4-NEXT: [[TMP47:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP44]], i32 0, i32 1
// CHECK4-NEXT: store i64 [[TMP43]], i64* [[TMP47]], align 8
// CHECK4-NEXT: [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP44]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 1, i8* [[TMP48]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i64 4, i64* [[DEP_COUNTER_ADDR]], align 8
// CHECK4-NEXT: [[TMP49:%.*]] = bitcast %struct.kmp_depend_info* [[TMP23]] to i8*
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM5:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB5]])
// CHECK4-NEXT: [[TMP50:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM5]], i8* [[TMP17]], i32 4, i8* [[TMP49]], i32 0, i8* null)
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM7:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB7:[0-9]+]])
// CHECK4-NEXT: [[TMP51:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM7]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.3*)* @.omp_task_entry..4 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP52:%.*]] = bitcast i8* [[TMP51]] to %struct.kmp_task_t_with_privates.3*
// CHECK4-NEXT: [[TMP53:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_3:%.*]], %struct.kmp_task_t_with_privates.3* [[TMP52]], i32 0, i32 0
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM8:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB7]])
// CHECK4-NEXT: [[TMP54:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP53]], i32 0, i32 2
// CHECK4-NEXT: store i32 0, i32* [[TMP54]], align 8
// CHECK4-NEXT: [[TMP55:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM8]], i8* [[TMP51]])
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM10:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB9:[0-9]+]])
// CHECK4-NEXT: [[TMP56:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM10]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.5*)* @.omp_task_entry..6 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP57:%.*]] = bitcast i8* [[TMP56]] to %struct.kmp_task_t_with_privates.5*
// CHECK4-NEXT: [[TMP58:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_5:%.*]], %struct.kmp_task_t_with_privates.5* [[TMP57]], i32 0, i32 0
// CHECK4-NEXT: [[TMP59:%.*]] = getelementptr inbounds [2 x %struct.kmp_depend_info], [2 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR11]], i64 0, i64 0
// CHECK4-NEXT: [[ARRAYIDX12:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 0
// CHECK4-NEXT: [[TMP60:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP59]], i64 0
// CHECK4-NEXT: [[TMP61:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i32 0, i32 0
// CHECK4-NEXT: [[TMP62:%.*]] = ptrtoint %struct.S* [[ARRAYIDX12]] to i64
// CHECK4-NEXT: store i64 [[TMP62]], i64* [[TMP61]], align 8
// CHECK4-NEXT: [[TMP63:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i32 0, i32 1
// CHECK4-NEXT: store i64 4, i64* [[TMP63]], align 8
// CHECK4-NEXT: [[TMP64:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP60]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 3, i8* [[TMP64]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP65:%.*]] = load i8, i8* [[B]], align 1
// CHECK4-NEXT: [[TMP66:%.*]] = sext i8 [[TMP65]] to i64
// CHECK4-NEXT: [[TMP67:%.*]] = mul nsw i64 4, [[TMP1]]
// CHECK4-NEXT: [[ARRAYIDX13:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP67]]
// CHECK4-NEXT: [[ARRAYIDX14:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX13]], i64 [[TMP66]]
// CHECK4-NEXT: [[TMP68:%.*]] = load i8, i8* [[B]], align 1
// CHECK4-NEXT: [[TMP69:%.*]] = sext i8 [[TMP68]] to i64
// CHECK4-NEXT: [[TMP70:%.*]] = mul nsw i64 9, [[TMP1]]
// CHECK4-NEXT: [[ARRAYIDX15:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP70]]
// CHECK4-NEXT: [[ARRAYIDX16:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX15]], i64 [[TMP69]]
// CHECK4-NEXT: [[TMP71:%.*]] = getelementptr i32, i32* [[ARRAYIDX16]], i32 1
// CHECK4-NEXT: [[TMP72:%.*]] = ptrtoint i32* [[ARRAYIDX14]] to i64
// CHECK4-NEXT: [[TMP73:%.*]] = ptrtoint i32* [[TMP71]] to i64
// CHECK4-NEXT: [[TMP74:%.*]] = sub nuw i64 [[TMP73]], [[TMP72]]
// CHECK4-NEXT: [[TMP75:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP59]], i64 1
// CHECK4-NEXT: [[TMP76:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP75]], i32 0, i32 0
// CHECK4-NEXT: [[TMP77:%.*]] = ptrtoint i32* [[ARRAYIDX14]] to i64
// CHECK4-NEXT: store i64 [[TMP77]], i64* [[TMP76]], align 8
// CHECK4-NEXT: [[TMP78:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP75]], i32 0, i32 1
// CHECK4-NEXT: store i64 [[TMP74]], i64* [[TMP78]], align 8
// CHECK4-NEXT: [[TMP79:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP75]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 3, i8* [[TMP79]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i64 2, i64* [[DEP_COUNTER_ADDR17]], align 8
// CHECK4-NEXT: [[TMP80:%.*]] = bitcast %struct.kmp_depend_info* [[TMP59]] to i8*
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM18:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB9]])
// CHECK4-NEXT: [[TMP81:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP58]], i32 0, i32 2
// CHECK4-NEXT: store i32 0, i32* [[TMP81]], align 8
// CHECK4-NEXT: [[TMP82:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM18]], i8* [[TMP56]], i32 2, i8* [[TMP80]], i32 0, i8* null)
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM20:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB11:[0-9]+]])
// CHECK4-NEXT: [[TMP83:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM20]], i32 0, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.7*)* @.omp_task_entry..8 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP84:%.*]] = bitcast i8* [[TMP83]] to %struct.kmp_task_t_with_privates.7*
// CHECK4-NEXT: [[TMP85:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_7:%.*]], %struct.kmp_task_t_with_privates.7* [[TMP84]], i32 0, i32 0
// CHECK4-NEXT: [[TMP86:%.*]] = getelementptr inbounds [2 x %struct.kmp_depend_info], [2 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR21]], i64 0, i64 0
// CHECK4-NEXT: [[ARRAYIDX22:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 0
// CHECK4-NEXT: [[TMP87:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP86]], i64 0
// CHECK4-NEXT: [[TMP88:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i32 0, i32 0
// CHECK4-NEXT: [[TMP89:%.*]] = ptrtoint %struct.S* [[ARRAYIDX22]] to i64
// CHECK4-NEXT: store i64 [[TMP89]], i64* [[TMP88]], align 8
// CHECK4-NEXT: [[TMP90:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i32 0, i32 1
// CHECK4-NEXT: store i64 4, i64* [[TMP90]], align 8
// CHECK4-NEXT: [[TMP91:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP87]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 4, i8* [[TMP91]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP92:%.*]] = load i8, i8* [[B]], align 1
// CHECK4-NEXT: [[TMP93:%.*]] = sext i8 [[TMP92]] to i64
// CHECK4-NEXT: [[TMP94:%.*]] = mul nsw i64 4, [[TMP1]]
// CHECK4-NEXT: [[ARRAYIDX23:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP94]]
// CHECK4-NEXT: [[ARRAYIDX24:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX23]], i64 [[TMP93]]
// CHECK4-NEXT: [[TMP95:%.*]] = load i8, i8* [[B]], align 1
// CHECK4-NEXT: [[TMP96:%.*]] = sext i8 [[TMP95]] to i64
// CHECK4-NEXT: [[TMP97:%.*]] = mul nsw i64 9, [[TMP1]]
// CHECK4-NEXT: [[ARRAYIDX25:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP97]]
// CHECK4-NEXT: [[ARRAYIDX26:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX25]], i64 [[TMP96]]
// CHECK4-NEXT: [[TMP98:%.*]] = getelementptr i32, i32* [[ARRAYIDX26]], i32 1
// CHECK4-NEXT: [[TMP99:%.*]] = ptrtoint i32* [[ARRAYIDX24]] to i64
// CHECK4-NEXT: [[TMP100:%.*]] = ptrtoint i32* [[TMP98]] to i64
// CHECK4-NEXT: [[TMP101:%.*]] = sub nuw i64 [[TMP100]], [[TMP99]]
// CHECK4-NEXT: [[TMP102:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP86]], i64 1
// CHECK4-NEXT: [[TMP103:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP102]], i32 0, i32 0
// CHECK4-NEXT: [[TMP104:%.*]] = ptrtoint i32* [[ARRAYIDX24]] to i64
// CHECK4-NEXT: store i64 [[TMP104]], i64* [[TMP103]], align 8
// CHECK4-NEXT: [[TMP105:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP102]], i32 0, i32 1
// CHECK4-NEXT: store i64 [[TMP101]], i64* [[TMP105]], align 8
// CHECK4-NEXT: [[TMP106:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP102]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 4, i8* [[TMP106]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i64 2, i64* [[DEP_COUNTER_ADDR27]], align 8
// CHECK4-NEXT: [[TMP107:%.*]] = bitcast %struct.kmp_depend_info* [[TMP86]] to i8*
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM28:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB11]])
// CHECK4-NEXT: [[TMP108:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP85]], i32 0, i32 2
// CHECK4-NEXT: store i32 0, i32* [[TMP108]], align 8
// CHECK4-NEXT: [[TMP109:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM28]], i8* [[TMP83]], i32 2, i8* [[TMP107]], i32 0, i8* null)
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM30:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB13:[0-9]+]])
// CHECK4-NEXT: [[TMP110:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM30]], i32 3, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.9*)* @.omp_task_entry..10 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP111:%.*]] = bitcast i8* [[TMP110]] to %struct.kmp_task_t_with_privates.9*
// CHECK4-NEXT: [[TMP112:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_9:%.*]], %struct.kmp_task_t_with_privates.9* [[TMP111]], i32 0, i32 0
// CHECK4-NEXT: [[TMP113:%.*]] = getelementptr inbounds [3 x %struct.kmp_depend_info], [3 x %struct.kmp_depend_info]* [[DOTDEP_ARR_ADDR31]], i64 0, i64 0
// CHECK4-NEXT: [[TMP114:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP113]], i64 0
// CHECK4-NEXT: [[TMP115:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i32 0, i32 0
// CHECK4-NEXT: store i64 ptrtoint (i32* @a to i64), i64* [[TMP115]], align 8
// CHECK4-NEXT: [[TMP116:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i32 0, i32 1
// CHECK4-NEXT: store i64 4, i64* [[TMP116]], align 8
// CHECK4-NEXT: [[TMP117:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP114]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 3, i8* [[TMP117]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[ARRAYIDX32:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i64 0, i64 1
// CHECK4-NEXT: [[TMP118:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP113]], i64 1
// CHECK4-NEXT: [[TMP119:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP118]], i32 0, i32 0
// CHECK4-NEXT: [[TMP120:%.*]] = ptrtoint %struct.S* [[ARRAYIDX32]] to i64
// CHECK4-NEXT: store i64 [[TMP120]], i64* [[TMP119]], align 8
// CHECK4-NEXT: [[TMP121:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP118]], i32 0, i32 1
// CHECK4-NEXT: store i64 4, i64* [[TMP121]], align 8
// CHECK4-NEXT: [[TMP122:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP118]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 3, i8* [[TMP122]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP123:%.*]] = mul nsw i64 0, [[TMP1]]
// CHECK4-NEXT: [[ARRAYIDX33:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP123]]
// CHECK4-NEXT: [[ARRAYIDX34:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX33]], i64 3
// CHECK4-NEXT: [[TMP124:%.*]] = load i32, i32* @a, align 4
// CHECK4-NEXT: [[TMP125:%.*]] = sext i32 [[TMP124]] to i64
// CHECK4-NEXT: [[LEN_SUB_1:%.*]] = sub nsw i64 [[TMP125]], 1
// CHECK4-NEXT: [[TMP126:%.*]] = load i32, i32* @a, align 4
// CHECK4-NEXT: [[TMP127:%.*]] = sext i32 [[TMP126]] to i64
// CHECK4-NEXT: [[LB_ADD_LEN:%.*]] = add nsw i64 -1, [[TMP127]]
// CHECK4-NEXT: [[TMP128:%.*]] = mul nsw i64 [[LB_ADD_LEN]], [[TMP1]]
// CHECK4-NEXT: [[ARRAYIDX35:%.*]] = getelementptr inbounds i32, i32* [[VLA]], i64 [[TMP128]]
// CHECK4-NEXT: [[ARRAYIDX36:%.*]] = getelementptr inbounds i32, i32* [[ARRAYIDX35]], i64 [[LEN_SUB_1]]
// CHECK4-NEXT: [[TMP129:%.*]] = getelementptr i32, i32* [[ARRAYIDX36]], i32 1
// CHECK4-NEXT: [[TMP130:%.*]] = ptrtoint i32* [[ARRAYIDX34]] to i64
// CHECK4-NEXT: [[TMP131:%.*]] = ptrtoint i32* [[TMP129]] to i64
// CHECK4-NEXT: [[TMP132:%.*]] = sub nuw i64 [[TMP131]], [[TMP130]]
// CHECK4-NEXT: [[TMP133:%.*]] = getelementptr [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP113]], i64 2
// CHECK4-NEXT: [[TMP134:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP133]], i32 0, i32 0
// CHECK4-NEXT: [[TMP135:%.*]] = ptrtoint i32* [[ARRAYIDX34]] to i64
// CHECK4-NEXT: store i64 [[TMP135]], i64* [[TMP134]], align 8
// CHECK4-NEXT: [[TMP136:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP133]], i32 0, i32 1
// CHECK4-NEXT: store i64 [[TMP132]], i64* [[TMP136]], align 8
// CHECK4-NEXT: [[TMP137:%.*]] = getelementptr inbounds [[STRUCT_KMP_DEPEND_INFO]], %struct.kmp_depend_info* [[TMP133]], i32 0, i32 2
Revert "[OpenMP] libomp: implement OpenMP 5.1 inoutset task dependence type" This reverts commit a1f550e052543f75acac9089b760cbc61729131f. Revert in order to fix backwards compatibility breakage caused by type size change for task dependence flag.
2021-06-09 22:35:50 +08:00
// CHECK4-NEXT: store i8 3, i8* [[TMP137]], align 8
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i64 3, i64* [[DEP_COUNTER_ADDR37]], align 8
// CHECK4-NEXT: [[TMP138:%.*]] = bitcast %struct.kmp_depend_info* [[TMP113]] to i8*
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM38:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB13]])
// CHECK4-NEXT: [[TMP139:%.*]] = call i32 @__kmpc_omp_task_with_deps(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM38]], i8* [[TMP110]], i32 3, i8* [[TMP138]], i32 0, i8* null)
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM40:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB15:[0-9]+]])
// CHECK4-NEXT: [[TMP140:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM40]], i32 3, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.11*)* @.omp_task_entry..12 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP141:%.*]] = bitcast i8* [[TMP140]] to %struct.kmp_task_t_with_privates.11*
// CHECK4-NEXT: [[TMP142:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_11:%.*]], %struct.kmp_task_t_with_privates.11* [[TMP141]], i32 0, i32 0
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM41:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB15]])
// CHECK4-NEXT: [[TMP143:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM41]], i8* [[TMP140]])
// CHECK4-NEXT: store i8 0, i8* [[FLAG]], align 1
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM43:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB17:[0-9]+]])
// CHECK4-NEXT: [[TMP144:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM43]], i32 1, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.13*)* @.omp_task_entry..14 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP145:%.*]] = bitcast i8* [[TMP144]] to %struct.kmp_task_t_with_privates.13*
// CHECK4-NEXT: [[TMP146:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_13:%.*]], %struct.kmp_task_t_with_privates.13* [[TMP145]], i32 0, i32 0
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM44:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB17]])
// CHECK4-NEXT: [[TMP147:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM44]], i8* [[TMP144]])
// CHECK4-NEXT: [[TMP148:%.*]] = getelementptr inbounds [[STRUCT_ANON_14]], %struct.anon.14* [[AGG_CAPTURED45]], i32 0, i32 0
// CHECK4-NEXT: store i32* [[C]], i32** [[TMP148]], align 8
// CHECK4-NEXT: [[TMP149:%.*]] = load i8, i8* [[B]], align 1
// CHECK4-NEXT: [[TOBOOL:%.*]] = icmp ne i8 [[TMP149]], 0
// CHECK4-NEXT: [[TMP150:%.*]] = select i1 [[TOBOOL]], i32 2, i32 0
// CHECK4-NEXT: [[TMP151:%.*]] = or i32 [[TMP150]], 1
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM46:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB19:[0-9]+]])
// CHECK4-NEXT: [[TMP152:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM46]], i32 [[TMP151]], i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.15*)* @.omp_task_entry..16 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP153:%.*]] = bitcast i8* [[TMP152]] to %struct.kmp_task_t_with_privates.15*
// CHECK4-NEXT: [[TMP154:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_15:%.*]], %struct.kmp_task_t_with_privates.15* [[TMP153]], i32 0, i32 0
// CHECK4-NEXT: [[TMP155:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP154]], i32 0, i32 0
// CHECK4-NEXT: [[TMP156:%.*]] = load i8*, i8** [[TMP155]], align 8
// CHECK4-NEXT: [[TMP157:%.*]] = bitcast %struct.anon.14* [[AGG_CAPTURED45]] to i8*
// CHECK4-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP156]], i8* align 8 [[TMP157]], i64 8, i1 false)
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM47:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB19]])
// CHECK4-NEXT: [[TMP158:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM47]], i8* [[TMP152]])
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM49:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21:[0-9]+]])
// CHECK4-NEXT: [[TMP159:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM49]], i32 0, i64 256, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.19*)* @.omp_task_entry..21 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP160:%.*]] = bitcast i8* [[TMP159]] to %struct.kmp_task_t_with_privates.19*
// CHECK4-NEXT: [[TMP161:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19:%.*]], %struct.kmp_task_t_with_privates.19* [[TMP160]], i32 0, i32 0
// CHECK4-NEXT: [[TMP162:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19]], %struct.kmp_task_t_with_privates.19* [[TMP160]], i32 0, i32 2
// CHECK4-NEXT: [[TMP163:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20:%.*]], %struct..kmp_privates.t.20* [[TMP162]], i32 0, i32 0
// CHECK4-NEXT: [[TMP164:%.*]] = load i32, i32* [[C]], align 128
// CHECK4-NEXT: store i32 [[TMP164]], i32* [[TMP163]], align 128
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM50:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]])
// CHECK4-NEXT: [[TMP165:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP161]], i32 0, i32 2
// CHECK4-NEXT: store i32 0, i32* [[TMP165]], align 16
// CHECK4-NEXT: [[TMP166:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM50]], i8* [[TMP159]])
// CHECK4-NEXT: [[TMP167:%.*]] = load i32, i32* @a, align 4
// CHECK4-NEXT: store i32 [[TMP167]], i32* [[RETVAL]], align 4
// CHECK4-NEXT: [[TMP168:%.*]] = load i8*, i8** [[SAVED_STACK]], align 8
// CHECK4-NEXT: call void @llvm.stackrestore(i8* [[TMP168]])
// CHECK4-NEXT: [[ARRAY_BEGIN51:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[S]], i32 0, i32 0
// CHECK4-NEXT: [[TMP169:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAY_BEGIN51]], i64 2
// CHECK4-NEXT: br label [[ARRAYDESTROY_BODY:%.*]]
// CHECK4: arraydestroy.body:
// CHECK4-NEXT: [[ARRAYDESTROY_ELEMENTPAST:%.*]] = phi %struct.S* [ [[TMP169]], [[ARRAYCTOR_CONT]] ], [ [[ARRAYDESTROY_ELEMENT:%.*]], [[ARRAYDESTROY_BODY]] ]
// CHECK4-NEXT: [[ARRAYDESTROY_ELEMENT]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[ARRAYDESTROY_ELEMENTPAST]], i64 -1
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[ARRAYDESTROY_ELEMENT]]) #[[ATTR4:[0-9]+]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[ARRAYDESTROY_DONE:%.*]] = icmp eq %struct.S* [[ARRAYDESTROY_ELEMENT]], [[ARRAY_BEGIN51]]
// CHECK4-NEXT: br i1 [[ARRAYDESTROY_DONE]], label [[ARRAYDESTROY_DONE52:%.*]], label [[ARRAYDESTROY_BODY]]
// CHECK4: arraydestroy.done52:
// CHECK4-NEXT: [[TMP170:%.*]] = load i32, i32* [[RETVAL]], align 4
// CHECK4-NEXT: ret i32 [[TMP170]]
//
//
// CHECK4-LABEL: define {{[^@]+}}@_ZN1SC1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1:[0-9]+]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK4-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK4-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN1SC2Ev(%struct.S* noundef [[THIS1]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry.
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates* noalias noundef [[TMP1:%.*]]) #[[ATTR3:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon*, align 8
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates* [[TMP1]], %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates*, %struct.kmp_task_t_with_privates** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META3:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META6:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META8:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META10:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !12
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !12
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !12
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !12
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !12
// CHECK4-NEXT: store %struct.anon* [[TMP8]], %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !12
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR_I]], align 8, !noalias !12
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 15, i32* @a, align 4
// CHECK4-NEXT: [[TMP11:%.*]] = load i32, i32* @a, align 4
// CHECK4-NEXT: [[CONV_I:%.*]] = trunc i32 [[TMP11]] to i8
// CHECK4-NEXT: [[TMP12:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP10]], i32 0, i32 0
// CHECK4-NEXT: [[TMP13:%.*]] = load i8*, i8** [[TMP12]], align 8
// CHECK4-NEXT: store i8 [[CONV_I]], i8* [[TMP13]], align 1
// CHECK4-NEXT: [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP10]], i32 0, i32 1
// CHECK4-NEXT: [[TMP15:%.*]] = load [2 x %struct.S]*, [2 x %struct.S]** [[TMP14]], align 8
// CHECK4-NEXT: [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[TMP15]], i64 0, i64 0
// CHECK4-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAYIDX_I]], i32 0, i32 0
// CHECK4-NEXT: store i32 10, i32* [[A_I]], align 4
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..2
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.1* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.0*, align 8
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.1*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.1* [[TMP1]], %struct.kmp_task_t_with_privates.1** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.1*, %struct.kmp_task_t_with_privates.1** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_1:%.*]], %struct.kmp_task_t_with_privates.1* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.0*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.1* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META13:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META16:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META18:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META20:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !22
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !22
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !22
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !22
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !22
// CHECK4-NEXT: store %struct.anon.0* [[TMP8]], %struct.anon.0** [[__CONTEXT_ADDR_I]], align 8, !noalias !22
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR_I]], align 8, !noalias !22
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 15, i32* @a, align 4
// CHECK4-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP10]], i32 0, i32 0
// CHECK4-NEXT: [[TMP12:%.*]] = load [2 x %struct.S]*, [2 x %struct.S]** [[TMP11]], align 8
// CHECK4-NEXT: [[ARRAYIDX_I:%.*]] = getelementptr inbounds [2 x %struct.S], [2 x %struct.S]* [[TMP12]], i64 0, i64 1
// CHECK4-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[ARRAYIDX_I]], i32 0, i32 0
// CHECK4-NEXT: store i32 10, i32* [[A_I]], align 4
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..4
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.3* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.2*, align 8
// CHECK4-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.3*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.3* [[TMP1]], %struct.kmp_task_t_with_privates.3** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.3*, %struct.kmp_task_t_with_privates.3** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_3:%.*]], %struct.kmp_task_t_with_privates.3* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.2*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.3* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META23:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META26:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META28:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META30:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !32
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !32
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !32
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !32
// CHECK4-NEXT: store %struct.anon.2* [[TMP8]], %struct.anon.2** [[__CONTEXT_ADDR_I]], align 8, !noalias !32
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon.2*, %struct.anon.2** [[__CONTEXT_ADDR_I]], align 8, !noalias !32
// CHECK4-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK4-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK4-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK4-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK4-NEXT: ]
// CHECK4: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK4: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB7]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !32
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM_I]], i8* [[TMP14]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__3_EXIT:%.*]]
// CHECK4: .untied.jmp.1.i:
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM2_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) #[[ATTR4]]
// CHECK4-NEXT: call void @__kmpc_critical(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM2_I]], [8 x i32]* @.gomp_critical_user_.var) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: call void @__kmpc_end_critical(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM2_I]], [8 x i32]* @.gomp_critical_user_.var) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[CLEANUP_I]]
// CHECK4: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !32
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__3_EXIT]]
// CHECK4: .omp_outlined..3.exit:
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..6
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.5* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.4*, align 8
// CHECK4-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.5*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.5* [[TMP1]], %struct.kmp_task_t_with_privates.5** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.5*, %struct.kmp_task_t_with_privates.5** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_5:%.*]], %struct.kmp_task_t_with_privates.5* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.4*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.5* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META33:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META36:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META38:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META40:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !42
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !42
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !42
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !42
// CHECK4-NEXT: store %struct.anon.4* [[TMP8]], %struct.anon.4** [[__CONTEXT_ADDR_I]], align 8, !noalias !42
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon.4*, %struct.anon.4** [[__CONTEXT_ADDR_I]], align 8, !noalias !42
// CHECK4-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK4-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK4-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK4-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK4-NEXT: ]
// CHECK4: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK4: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB9]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !42
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM_I]], i8* [[TMP14]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__5_EXIT:%.*]]
// CHECK4: .untied.jmp.1.i:
// CHECK4-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[CLEANUP_I]]
// CHECK4: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !42
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__5_EXIT]]
// CHECK4: .omp_outlined..5.exit:
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.7* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.6*, align 8
// CHECK4-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.7*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.7* [[TMP1]], %struct.kmp_task_t_with_privates.7** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.7*, %struct.kmp_task_t_with_privates.7** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_7:%.*]], %struct.kmp_task_t_with_privates.7* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.6*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.7* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META43:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META46:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META48:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META50:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !52
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !52
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !52
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !52
// CHECK4-NEXT: store %struct.anon.6* [[TMP8]], %struct.anon.6** [[__CONTEXT_ADDR_I]], align 8, !noalias !52
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon.6*, %struct.anon.6** [[__CONTEXT_ADDR_I]], align 8, !noalias !52
// CHECK4-NEXT: [[TMP11:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP11]], align 4
// CHECK4-NEXT: switch i32 [[TMP12]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK4-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK4-NEXT: i32 1, label [[DOTUNTIED_JMP_1_I:%.*]]
// CHECK4-NEXT: ]
// CHECK4: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK4: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP13:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 1, i32* [[TMP13]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB11]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !52
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[TMP15:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM_I]], i8* [[TMP14]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__7_EXIT:%.*]]
// CHECK4: .untied.jmp.1.i:
// CHECK4-NEXT: store i32 1, i32* @a, align 4
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[CLEANUP_I]]
// CHECK4: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !52
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__7_EXIT]]
// CHECK4: .omp_outlined..7.exit:
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..10
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.9* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.8*, align 8
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.9*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.9* [[TMP1]], %struct.kmp_task_t_with_privates.9** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.9*, %struct.kmp_task_t_with_privates.9** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_9:%.*]], %struct.kmp_task_t_with_privates.9* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.8*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.9* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META53:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META56:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META58:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META60:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !62
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !62
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !62
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !62
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !62
// CHECK4-NEXT: store %struct.anon.8* [[TMP8]], %struct.anon.8** [[__CONTEXT_ADDR_I]], align 8, !noalias !62
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon.8*, %struct.anon.8** [[__CONTEXT_ADDR_I]], align 8, !noalias !62
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 2, i32* @a, align 4
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..12
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.11* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.10*, align 8
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.11*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.11* [[TMP1]], %struct.kmp_task_t_with_privates.11** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.11*, %struct.kmp_task_t_with_privates.11** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_11:%.*]], %struct.kmp_task_t_with_privates.11* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.10*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.11* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META63:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META66:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META68:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META70:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !72
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !72
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !72
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !72
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !72
// CHECK4-NEXT: store %struct.anon.10* [[TMP8]], %struct.anon.10** [[__CONTEXT_ADDR_I]], align 8, !noalias !72
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon.10*, %struct.anon.10** [[__CONTEXT_ADDR_I]], align 8, !noalias !72
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 2, i32* @a, align 4
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..14
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.13* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.12*, align 8
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.13*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.13* [[TMP1]], %struct.kmp_task_t_with_privates.13** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.13*, %struct.kmp_task_t_with_privates.13** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_13:%.*]], %struct.kmp_task_t_with_privates.13* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.12*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.13* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META73:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META76:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META78:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META80:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !82
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !82
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !82
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !82
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !82
// CHECK4-NEXT: store %struct.anon.12* [[TMP8]], %struct.anon.12** [[__CONTEXT_ADDR_I]], align 8, !noalias !82
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon.12*, %struct.anon.12** [[__CONTEXT_ADDR_I]], align 8, !noalias !82
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 3, i32* @a, align 4
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..16
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.15* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.14*, align 8
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.15*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.15* [[TMP1]], %struct.kmp_task_t_with_privates.15** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.15*, %struct.kmp_task_t_with_privates.15** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_15:%.*]], %struct.kmp_task_t_with_privates.15* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.14*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.15* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META83:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META86:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META88:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META90:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !92
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !92
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !92
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !92
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !92
// CHECK4-NEXT: store %struct.anon.14* [[TMP8]], %struct.anon.14** [[__CONTEXT_ADDR_I]], align 8, !noalias !92
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon.14*, %struct.anon.14** [[__CONTEXT_ADDR_I]], align 8, !noalias !92
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 4, i32* @a, align 4
// CHECK4-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_14:%.*]], %struct.anon.14* [[TMP10]], i32 0, i32 0
// CHECK4-NEXT: [[TMP12:%.*]] = load i32*, i32** [[TMP11]], align 8
// CHECK4-NEXT: store i32 5, i32* [[TMP12]], align 128
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_privates_map.
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (%struct..kmp_privates.t* noalias noundef [[TMP0:%.*]], i32** noalias noundef [[TMP1:%.*]]) #[[ATTR7:[0-9]+]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t*, align 8
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca i32**, align 8
// CHECK4-NEXT: store %struct..kmp_privates.t* [[TMP0]], %struct..kmp_privates.t** [[DOTADDR]], align 8
// CHECK4-NEXT: store i32** [[TMP1]], i32*** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load %struct..kmp_privates.t*, %struct..kmp_privates.t** [[DOTADDR]], align 8
// CHECK4-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP2]], i32 0, i32 0
// CHECK4-NEXT: [[TMP4:%.*]] = load i32**, i32*** [[DOTADDR1]], align 8
// CHECK4-NEXT: store i32* [[TMP3]], i32** [[TMP4]], align 8
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..19
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.18* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.17*, align 8
// CHECK4-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.18*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.18* [[TMP1]], %struct.kmp_task_t_with_privates.18** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.18*, %struct.kmp_task_t_with_privates.18** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18:%.*]], %struct.kmp_task_t_with_privates.18* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 128
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.17*
// CHECK4-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18]], %struct.kmp_task_t_with_privates.18* [[TMP3]], i32 0, i32 2
// CHECK4-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t* [[TMP9]] to i8*
// CHECK4-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.18* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META93:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META96:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META98:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META100:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !102
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !102
// CHECK4-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !102
// CHECK4-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t*, i32**)* @.omp_task_privates_map. to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !102
// CHECK4-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !102
// CHECK4-NEXT: store %struct.anon.17* [[TMP8]], %struct.anon.17** [[__CONTEXT_ADDR_I]], align 8, !noalias !102
// CHECK4-NEXT: [[TMP12:%.*]] = load %struct.anon.17*, %struct.anon.17** [[__CONTEXT_ADDR_I]], align 8, !noalias !102
// CHECK4-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !102
// CHECK4-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !102
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i32**)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: call void [[TMP15]](i8* [[TMP14]], i32** [[DOTFIRSTPRIV_PTR_ADDR_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP16:%.*]] = load i32*, i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !102
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 4, i32* [[TMP16]], align 128
// CHECK4-NEXT: store i32 4, i32* @a, align 4
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@_ZN1SD1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK4-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK4-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN1SD2Ev(%struct.S* noundef [[THIS1]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_privates_map..20
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (%struct..kmp_privates.t.20* noalias noundef [[TMP0:%.*]], i32** noalias noundef [[TMP1:%.*]], %struct.S** noalias noundef [[TMP2:%.*]], %struct.S** noalias noundef [[TMP3:%.*]]) #[[ATTR7]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca %struct..kmp_privates.t.20*, align 8
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca i32**, align 8
// CHECK4-NEXT: [[DOTADDR2:%.*]] = alloca %struct.S**, align 8
// CHECK4-NEXT: [[DOTADDR3:%.*]] = alloca %struct.S**, align 8
// CHECK4-NEXT: store %struct..kmp_privates.t.20* [[TMP0]], %struct..kmp_privates.t.20** [[DOTADDR]], align 8
// CHECK4-NEXT: store i32** [[TMP1]], i32*** [[DOTADDR1]], align 8
// CHECK4-NEXT: store %struct.S** [[TMP2]], %struct.S*** [[DOTADDR2]], align 8
// CHECK4-NEXT: store %struct.S** [[TMP3]], %struct.S*** [[DOTADDR3]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = load %struct..kmp_privates.t.20*, %struct..kmp_privates.t.20** [[DOTADDR]], align 8
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20:%.*]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP6:%.*]] = load i32**, i32*** [[DOTADDR1]], align 8
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP7:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 1
// CHECK4-NEXT: [[TMP8:%.*]] = load %struct.S**, %struct.S*** [[DOTADDR2]], align 8
// CHECK4-NEXT: store %struct.S* [[TMP7]], %struct.S** [[TMP8]], align 8
// CHECK4-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T_20]], %struct..kmp_privates.t.20* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.S**, %struct.S*** [[DOTADDR3]], align 8
// CHECK4-NEXT: store %struct.S* [[TMP9]], %struct.S** [[TMP10]], align 8
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..21
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.19* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.16*, align 8
// CHECK4-NEXT: [[DOTFIRSTPRIV_PTR_ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTLOCAL_PTR_ADDR_I:%.*]] = alloca %struct.S*, align 8
// CHECK4-NEXT: [[DOTLOCAL_PTR_ADDR1_I:%.*]] = alloca %struct.S*, align 8
// CHECK4-NEXT: [[CLEANUP_DEST_SLOT_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[S1_I:%.*]] = alloca [[STRUCT_S:%.*]], align 4
// CHECK4-NEXT: [[S2_I:%.*]] = alloca [[STRUCT_S]], align 4
// CHECK4-NEXT: [[REF_TMP_I:%.*]] = alloca [[STRUCT_S]], align 4
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.19*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.19* [[TMP1]], %struct.kmp_task_t_with_privates.19** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.19*, %struct.kmp_task_t_with_privates.19** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19:%.*]], %struct.kmp_task_t_with_privates.19* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 128
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.16*
// CHECK4-NEXT: [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_19]], %struct.kmp_task_t_with_privates.19* [[TMP3]], i32 0, i32 2
// CHECK4-NEXT: [[TMP10:%.*]] = bitcast %struct..kmp_privates.t.20* [[TMP9]] to i8*
// CHECK4-NEXT: [[TMP11:%.*]] = bitcast %struct.kmp_task_t_with_privates.19* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META103:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META106:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META108:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META110:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !112
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
// CHECK4-NEXT: store i8* [[TMP10]], i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !112
// CHECK4-NEXT: store void (i8*, ...)* bitcast (void (%struct..kmp_privates.t.20*, i32**, %struct.S**, %struct.S**)* @.omp_task_privates_map..20 to void (i8*, ...)*), void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !112
// CHECK4-NEXT: store i8* [[TMP11]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
// CHECK4-NEXT: store %struct.anon.16* [[TMP8]], %struct.anon.16** [[__CONTEXT_ADDR_I]], align 8, !noalias !112
// CHECK4-NEXT: [[TMP12:%.*]] = load %struct.anon.16*, %struct.anon.16** [[__CONTEXT_ADDR_I]], align 8, !noalias !112
// CHECK4-NEXT: [[TMP13:%.*]] = load void (i8*, ...)*, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !112
// CHECK4-NEXT: [[TMP14:%.*]] = load i8*, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP15:%.*]] = bitcast void (i8*, ...)* [[TMP13]] to void (i8*, i32**, %struct.S**, %struct.S**)*
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: call void [[TMP15]](i8* [[TMP14]], i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], %struct.S** [[DOTLOCAL_PTR_ADDR_I]], %struct.S** [[DOTLOCAL_PTR_ADDR1_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP16:%.*]] = load i32*, i32** [[DOTFIRSTPRIV_PTR_ADDR_I]], align 8, !noalias !112
// CHECK4-NEXT: [[TMP17:%.*]] = load %struct.S*, %struct.S** [[DOTLOCAL_PTR_ADDR_I]], align 8, !noalias !112
// CHECK4-NEXT: [[TMP18:%.*]] = load %struct.S*, %struct.S** [[DOTLOCAL_PTR_ADDR1_I]], align 8, !noalias !112
// CHECK4-NEXT: [[TMP19:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP20:%.*]] = load i32, i32* [[TMP19]], align 4
// CHECK4-NEXT: switch i32 [[TMP20]], label [[DOTUNTIED_DONE__I:%.*]] [
// CHECK4-NEXT: i32 0, label [[DOTUNTIED_JMP__I:%.*]]
// CHECK4-NEXT: i32 1, label [[DOTUNTIED_JMP_2_I:%.*]]
// CHECK4-NEXT: i32 2, label [[DOTUNTIED_JMP_6_I:%.*]]
// CHECK4-NEXT: i32 3, label [[DOTUNTIED_JMP_10_I:%.*]]
// CHECK4-NEXT: i32 4, label [[DOTUNTIED_JMP_15_I:%.*]]
// CHECK4-NEXT: ]
// CHECK4: .untied.done..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 1, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[CLEANUP_I:%.*]]
// CHECK4: .untied.jmp..i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP21:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 1, i32* [[TMP21]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP22:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[TMP23:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM_I]], i8* [[TMP22]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT:%.*]]
// CHECK4: .untied.jmp.2.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[S1_I]]) #[[ATTR4]]
// CHECK4-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[S2_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[S2_I]], i32 0, i32 0
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 0, i32* [[A_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM3_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB23:[0-9]+]]) #[[ATTR4]]
// CHECK4-NEXT: [[TMP24:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM3_I]], i32 1, i64 256, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.18*)* @.omp_task_entry..19 to i32 (i32, i8*)*)) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP25:%.*]] = bitcast i8* [[TMP24]] to %struct.kmp_task_t_with_privates.18*
// CHECK4-NEXT: [[TMP26:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18:%.*]], %struct.kmp_task_t_with_privates.18* [[TMP25]], i32 0, i32 0
// CHECK4-NEXT: [[TMP27:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_18]], %struct.kmp_task_t_with_privates.18* [[TMP25]], i32 0, i32 2
// CHECK4-NEXT: [[TMP28:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP27]], i32 0, i32 0
// CHECK4-NEXT: [[TMP29:%.*]] = load i32, i32* [[TMP16]], align 128
// CHECK4-NEXT: store i32 [[TMP29]], i32* [[TMP28]], align 128
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM4_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB23]]) #[[ATTR4]]
// CHECK4-NEXT: [[TMP30:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM4_I]], i8* [[TMP24]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP31:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 2, i32* [[TMP31]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM5_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP32:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[TMP33:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM5_I]], i8* [[TMP32]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK4: .untied.jmp.6.i:
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM8_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) #[[ATTR4]]
// CHECK4-NEXT: [[TMP34:%.*]] = call i32 @__kmpc_omp_taskyield(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM8_I]], i32 0) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP35:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 3, i32* [[TMP35]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM9_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP36:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[TMP37:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM9_I]], i8* [[TMP36]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK4: .untied.jmp.10.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN1SC1Ev(%struct.S* noundef [[REF_TMP_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP38:%.*]] = bitcast %struct.S* [[S1_I]] to i8*
// CHECK4-NEXT: [[TMP39:%.*]] = bitcast %struct.S* [[REF_TMP_I]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP38]], i8* align 4 [[TMP39]], i64 4, i1 false) #[[ATTR4]], !noalias !112
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[REF_TMP_I]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[A12_I:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.S* [[S2_I]], i32 0, i32 0
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 10, i32* [[A12_I]], align 4, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM13_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) #[[ATTR4]]
// CHECK4-NEXT: [[TMP40:%.*]] = call i32 @__kmpc_omp_taskwait(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM13_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP41:%.*]] = load i32*, i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: store i32 4, i32* [[TMP41]], align 4
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM14_I:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB21]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[TMP42:%.*]] = load i8*, i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !112
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: [[TMP43:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM14_I]], i8* [[TMP42]]) #[[ATTR4]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK4: .untied.jmp.15.i:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[S2_I]]) #[[ATTR4]]
// CHECK4-NEXT: call void @_ZN1SD1Ev(%struct.S* noundef [[S1_I]]) #[[ATTR4]]
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: store i32 0, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[CLEANUP_I]]
// CHECK4: cleanup.i:
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: [[CLEANUP_DEST_I:%.*]] = load i32, i32* [[CLEANUP_DEST_SLOT_I]], align 4, !noalias !112
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: br label [[DOTOMP_OUTLINED__17_EXIT]]
// CHECK4: .omp_outlined..17.exit:
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@_ZN1SC2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK4-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK4-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
// CHECK4-NEXT: [[A:%.*]] = getelementptr inbounds [[STRUCT_S:%.*]], %struct.S* [[THIS1]], i32 0, i32 0
// CHECK4-NEXT: store i32 0, i32* [[A]], align 4
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@_ZN1SD2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (%struct.S* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S*, align 8
// CHECK4-NEXT: store %struct.S* [[THIS]], %struct.S** [[THIS_ADDR]], align 8
// CHECK4-NEXT: [[THIS1:%.*]] = load %struct.S*, %struct.S** [[THIS_ADDR]], align 8
// CHECK4-NEXT: ret void
//
//
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-LABEL: define {{[^@]+}}@__cxx_global_var_init
// CHECK4-SAME: () #[[ATTR7]] section "__TEXT,__StaticInit,regular,pure_instructions" {
// CHECK4-NEXT: entry:
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN2S1C1Ev(%struct.S1* noundef @s1)
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@_ZN2S1C1Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (%struct.S1* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK4-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK4-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN2S1C2Ev(%struct.S1* noundef [[THIS1]])
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-NEXT: ret void
//
//
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-LABEL: define {{[^@]+}}@_ZN2S1C2Ev
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (%struct.S1* noundef [[THIS:%.*]]) unnamed_addr #[[ATTR1]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK4-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK4-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-NEXT: call void @_ZN2S18taskinitEv(%struct.S1* noundef [[THIS1]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@_ZN2S18taskinitEv
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (%struct.S1* noundef [[THIS:%.*]]) #[[ATTR8:[0-9]+]] align 2 {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[THIS_ADDR:%.*]] = alloca %struct.S1*, align 8
// CHECK4-NEXT: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON_21:%.*]], align 8
// CHECK4-NEXT: store %struct.S1* [[THIS]], %struct.S1** [[THIS_ADDR]], align 8
// CHECK4-NEXT: [[THIS1:%.*]] = load %struct.S1*, %struct.S1** [[THIS_ADDR]], align 8
// CHECK4-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_21]], %struct.anon.21* [[AGG_CAPTURED]], i32 0, i32 0
// CHECK4-NEXT: store %struct.S1* [[THIS1]], %struct.S1** [[TMP0]], align 8
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB25:[0-9]+]])
// CHECK4-NEXT: [[TMP1:%.*]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM]], i32 1, i64 40, i64 8, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates.22*)* @.omp_task_entry..23 to i32 (i32, i8*)*))
// CHECK4-NEXT: [[TMP2:%.*]] = bitcast i8* [[TMP1]] to %struct.kmp_task_t_with_privates.22*
// CHECK4-NEXT: [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_22:%.*]], %struct.kmp_task_t_with_privates.22* [[TMP2]], i32 0, i32 0
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = load i8*, i8** [[TMP4]], align 8
// CHECK4-NEXT: [[TMP6:%.*]] = bitcast %struct.anon.21* [[AGG_CAPTURED]] to i8*
// CHECK4-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP5]], i8* align 8 [[TMP6]], i64 8, i1 false)
// CHECK4-NEXT: [[OMP_GLOBAL_THREAD_NUM2:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB25]])
// CHECK4-NEXT: [[TMP7:%.*]] = call i32 @__kmpc_omp_task(%struct.ident_t* @[[GLOB1]], i32 [[OMP_GLOBAL_THREAD_NUM2]], i8* [[TMP1]])
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_task_entry..23
[Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default Turning on `enable_noundef_analysis` flag allows better codegen by removing freeze instructions. I modified clang by renaming `enable_noundef_analysis` flag to `disable-noundef-analysis` and turning it off by default. Test updates are made as a separate patch: D108453 Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D105169 [Clang/Test]: Rename enable_noundef_analysis to disable-noundef-analysis and turn it off by default (2) This patch updates test files after D105169. Autogenerated test codes are changed by `utils/update_cc_test_checks.py,` and non-autogenerated test codes are changed as follows: (1) I wrote a python script that (partially) updates the tests using regex: {F18594904} The script is not perfect, but I believe it gives hints about which patterns are updated to have `noundef` attached. (2) The remaining tests are updated manually. Reviewed By: eugenis Differential Revision: https://reviews.llvm.org/D108453 Resolve lit failures in clang after 8ca4b3e's land Fix lit test failures in clang-ppc* and clang-x64-windows-msvc Fix missing failures in clang-ppc64be* and retry fixing clang-x64-windows-msvc Fix internal_clone(aarch64) inline assembly
2021-10-15 18:26:07 +08:00
// CHECK4-SAME: (i32 noundef [[TMP0:%.*]], %struct.kmp_task_t_with_privates.22* noalias noundef [[TMP1:%.*]]) #[[ATTR3]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR_I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTPART_ID__ADDR_I:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTPRIVATES__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[DOTCOPY_FN__ADDR_I:%.*]] = alloca void (i8*, ...)*, align 8
// CHECK4-NEXT: [[DOTTASK_T__ADDR_I:%.*]] = alloca i8*, align 8
// CHECK4-NEXT: [[__CONTEXT_ADDR_I:%.*]] = alloca %struct.anon.21*, align 8
// CHECK4-NEXT: [[DOTADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTADDR1:%.*]] = alloca %struct.kmp_task_t_with_privates.22*, align 8
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[DOTADDR]], align 4
// CHECK4-NEXT: store %struct.kmp_task_t_with_privates.22* [[TMP1]], %struct.kmp_task_t_with_privates.22** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP2:%.*]] = load i32, i32* [[DOTADDR]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load %struct.kmp_task_t_with_privates.22*, %struct.kmp_task_t_with_privates.22** [[DOTADDR1]], align 8
// CHECK4-NEXT: [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES_22:%.*]], %struct.kmp_task_t_with_privates.22* [[TMP3]], i32 0, i32 0
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 2
// CHECK4-NEXT: [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP4]], i32 0, i32 0
// CHECK4-NEXT: [[TMP7:%.*]] = load i8*, i8** [[TMP6]], align 8
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8* [[TMP7]] to %struct.anon.21*
// CHECK4-NEXT: [[TMP9:%.*]] = bitcast %struct.kmp_task_t_with_privates.22* [[TMP3]] to i8*
[OpenMP] Add Module metadata for OpenMP compilation This patch adds a module level metadata flag indicating that the module was compiled with the `-fopenmp` flag. This will make it easier for passes like OpenMPOpt to determine if it should be run. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D102361
2021-06-25 02:39:12 +08:00
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META113:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META116:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META118:![0-9]+]])
// CHECK4-NEXT: call void @llvm.experimental.noalias.scope.decl(metadata [[META120:![0-9]+]])
// CHECK4-NEXT: store i32 [[TMP2]], i32* [[DOTGLOBAL_TID__ADDR_I]], align 4, !noalias !122
// CHECK4-NEXT: store i32* [[TMP5]], i32** [[DOTPART_ID__ADDR_I]], align 8, !noalias !122
// CHECK4-NEXT: store i8* null, i8** [[DOTPRIVATES__ADDR_I]], align 8, !noalias !122
// CHECK4-NEXT: store void (i8*, ...)* null, void (i8*, ...)** [[DOTCOPY_FN__ADDR_I]], align 8, !noalias !122
// CHECK4-NEXT: store i8* [[TMP9]], i8** [[DOTTASK_T__ADDR_I]], align 8, !noalias !122
// CHECK4-NEXT: store %struct.anon.21* [[TMP8]], %struct.anon.21** [[__CONTEXT_ADDR_I]], align 8, !noalias !122
// CHECK4-NEXT: [[TMP10:%.*]] = load %struct.anon.21*, %struct.anon.21** [[__CONTEXT_ADDR_I]], align 8, !noalias !122
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: [[TMP11:%.*]] = getelementptr inbounds [[STRUCT_ANON_21:%.*]], %struct.anon.21* [[TMP10]], i32 0, i32 0
// CHECK4-NEXT: [[TMP12:%.*]] = load %struct.S1*, %struct.S1** [[TMP11]], align 8
// CHECK4-NEXT: [[A_I:%.*]] = getelementptr inbounds [[STRUCT_S1:%.*]], %struct.S1* [[TMP12]], i32 0, i32 0
// CHECK4-NEXT: store i32 0, i32* [[A_I]], align 4
// CHECK4-NEXT: ret i32 0
//
//
// CHECK4-LABEL: define {{[^@]+}}@_GLOBAL__sub_I_task_codegen.cpp
[OpenMP] Overhaul `declare target` handling This patch fixes various issues with our prior `declare target` handling and extends it to support `omp begin declare target` as well. This started with PR49649 in mind, trying to provide a way for users to avoid the "ref" global use introduced for globals with internal linkage. From there it went down the rabbit hole, e.g., all variables, even `nohost` ones, were emitted into the device code so it was impossible to determine if "ref" was needed late in the game (based on the name only). To make it really useful, `begin declare target` was needed as it can carry the `device_type`. Not emitting variables eagerly had a ripple effect. Finally, the precedence of the (explicit) declare target list items needed to be taken into account, that meant we cannot just look for any declare target attribute to make a decision. This caused the handling of functions to require fixup as well. I tried to clean up things while I was at it, e.g., we should not "parse declarations and defintions" as part of OpenMP parsing, this will always break at some point. Instead, we keep track what region we are in and act on definitions and declarations instead, this is what we do for declare variant and other begin/end directives already. Highlights: - new diagnosis for restrictions specificed in the standard, - delayed emission of globals not mentioned in an explicit list of a declare target, - omission of `nohost` globals on the host and `host` globals on the device, - no explicit parsing of declarations in-between `omp [begin] declare variant` and the corresponding end anymore, regular parsing instead, - precedence for explicit mentions in `declare target` lists over implicit mentions in the declaration-definition-seq, and - `omp allocate` declarations will now replace an earlier emitted global, if necessary. --- Notes: The patch is larger than I hoped but it turns out that most changes do on their own lead to "inconsistent states", which seem less desirable overall. After working through this I feel the standard should remove the explicit declare target forms as the delayed emission is horrible. That said, while we delay things anyway, it seems to me we check too often for the current status even though that is often not sufficient to act upon. There seems to be a lot of duplication that can probably be trimmed down. Eagerly emitting some things seems pretty weak as an argument to keep so much logic around. --- Reviewed By: ABataev Differential Revision: https://reviews.llvm.org/D101030
2021-04-22 13:57:28 +08:00
// CHECK4-SAME: () #[[ATTR7]] section "__TEXT,__StaticInit,regular,pure_instructions" {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: call void @__cxx_global_var_init()
// CHECK4-NEXT: ret void
//