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/target_map_codegen_03.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1163 lines
85 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 _
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
// expected-no-diagnostics
#ifndef HEADER
#define HEADER
///==========================================================================///
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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 -DCK4 -verify -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK1
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
// RUN: %clang_cc1 -DCK4 -fopenmp -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -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-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK2
// RUN: %clang_cc1 -DCK4 -verify -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK3
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
// RUN: %clang_cc1 -DCK4 -fopenmp -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -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-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK4
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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 -DCK4 -verify -fopenmp -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK5
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
// RUN: %clang_cc1 -DCK4 -fopenmp -fopenmp-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -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-version=45 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK6
// RUN: %clang_cc1 -DCK4 -verify -fopenmp -fopenmp-version=45 -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK7
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
// RUN: %clang_cc1 -DCK4 -fopenmp -fopenmp-version=45 -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -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-version=45 -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK8
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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 -DCK4 -verify -fopenmp -fopenmp-version=50 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK9
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
// RUN: %clang_cc1 -DCK4 -fopenmp -fopenmp-version=50 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -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-version=50 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK10
// RUN: %clang_cc1 -DCK4 -verify -fopenmp -fopenmp-version=50 -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK11
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
// RUN: %clang_cc1 -DCK4 -fopenmp -fopenmp-version=50 -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -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-version=50 -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck -allow-deprecated-dag-overlap %s --check-prefix=CHECK12
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +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 -DCK4 -verify -fopenmp-simd -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -emit-llvm %s -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
// RUN: %clang_cc1 -DCK4 -fopenmp-simd -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -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 -fopenmp-targets=powerpc64le-ibm-linux-gnu -x c++ -triple powerpc64le-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
// RUN: %clang_cc1 -DCK4 -verify -fopenmp-simd -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -emit-llvm %s -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
// RUN: %clang_cc1 -DCK4 -fopenmp-simd -fopenmp-targets=i386-pc-linux-gnu -x c++ -std=c++11 -triple i386-unknown-unknown -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 -fopenmp-targets=i386-pc-linux-gnu -x c++ -triple i386-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
[OpenMP] Split OpenMP/target_map_codegen test [NFC] The test file is the single longest test among clang's tests and ends up about doubling the wall time of clang tests on machines with high number of cores. The test appears to consist of multiple independent subtests and does not have to be in one file. Splitting it into smaller parts reduces test time on my machine from ~80s down to ~45. Differential Revision: https://reviews.llvm.org/D85551
2020-08-08 03:59:49 +08:00
#ifdef CK4
// Map types: OMP_MAP_PRIVATE_VAL | OMP_MAP_TARGET_PARAM | OMP_MAP_IMPLICIT = 800
void implicit_maps_nested_integer (int a){
int i = a;
// The captures in parallel are by reference. Only the capture in target is by
// copy.
#pragma omp parallel
{
#pragma omp target
{
#pragma omp parallel
{
++i;
}
}
}
}
#endif // CK4
#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 {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK1-SAME: (i32 signext [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], 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 [[A]], i32* [[A_ADDR]], align 4
// CHECK1-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK1-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK1-NEXT: store i32* [[I]], i32** [[TMP1]], align 8
// CHECK1-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, 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: [[I_CASTED:%.*]] = alloca i64, align 8
// CHECK1-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK1-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK1-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
// CHECK1-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK1-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK1-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK1-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK1-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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: [[CONV:%.*]] = bitcast i64* [[I_CASTED]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: store i32 [[TMP3]], i32* [[CONV]], align 4
// CHECK1-NEXT: [[TMP4:%.*]] = load i64, i64* [[I_CASTED]], align 8
// CHECK1-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: store i64 [[TMP4]], i64* [[TMP6]], align 8
// CHECK1-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK1-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
// CHECK1-NEXT: store i64 [[TMP4]], i64* [[TMP8]], align 8
// CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
// CHECK1-NEXT: store i8* null, i8** [[TMP9]], align 8
// CHECK1-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK1-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK1-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK1-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK1-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i64 [[TMP4]]) #[[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: br label [[OMP_OFFLOAD_CONT]]
// CHECK1: omp_offload.cont:
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK1-SAME: (i64 [[I:%.*]]) #[[ATTR2:[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: [[I_ADDR:%.*]] = alloca i64, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], 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 [[I]], i64* [[I_ADDR]], align 8
// CHECK1-NEXT: [[CONV:%.*]] = bitcast i64* [[I_ADDR]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK1-NEXT: store i32* [[CONV]], i32** [[TMP0]], align 8
// 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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, 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
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK1-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK1-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK1-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK1-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK1-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK1-SAME: () #[[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: entry:
// CHECK1-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK1-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK2-SAME: (i32 signext [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], 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 [[A]], i32* [[A_ADDR]], align 4
// CHECK2-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK2-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK2-NEXT: store i32* [[I]], i32** [[TMP1]], align 8
// CHECK2-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, 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: [[I_CASTED:%.*]] = alloca i64, align 8
// CHECK2-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK2-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK2-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
// CHECK2-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK2-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK2-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK2-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK2-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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: [[CONV:%.*]] = bitcast i64* [[I_CASTED]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: store i32 [[TMP3]], i32* [[CONV]], align 4
// CHECK2-NEXT: [[TMP4:%.*]] = load i64, i64* [[I_CASTED]], align 8
// CHECK2-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: store i64 [[TMP4]], i64* [[TMP6]], align 8
// CHECK2-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK2-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
// CHECK2-NEXT: store i64 [[TMP4]], i64* [[TMP8]], align 8
// CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
// CHECK2-NEXT: store i8* null, i8** [[TMP9]], align 8
// CHECK2-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK2-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK2-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK2-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK2-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i64 [[TMP4]]) #[[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: br label [[OMP_OFFLOAD_CONT]]
// CHECK2: omp_offload.cont:
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK2-SAME: (i64 [[I:%.*]]) #[[ATTR2:[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: [[I_ADDR:%.*]] = alloca i64, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], 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 [[I]], i64* [[I_ADDR]], align 8
// CHECK2-NEXT: [[CONV:%.*]] = bitcast i64* [[I_ADDR]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK2-NEXT: store i32* [[CONV]], i32** [[TMP0]], align 8
// 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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, 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
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK2-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK2-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK2-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK2-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK2-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK2-SAME: () #[[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: entry:
// CHECK2-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK2-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK3-SAME: (i32 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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 [[A]], i32* [[A_ADDR]], align 4
// CHECK3-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK3-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK3-NEXT: store i32* [[I]], i32** [[TMP1]], align 4
// CHECK3-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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:%.*]] = alloca i32*, align 4
// CHECK3-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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: [[I_CASTED:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK3-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK3-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
// CHECK3-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK3-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK3-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK3-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK3-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK3-NEXT: store i32 [[TMP3]], i32* [[I_CASTED]], align 4
// CHECK3-NEXT: [[TMP4:%.*]] = load i32, i32* [[I_CASTED]], align 4
// CHECK3-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK3-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: store i32 [[TMP4]], i32* [[TMP6]], align 4
// CHECK3-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK3-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
// CHECK3-NEXT: store i32 [[TMP4]], i32* [[TMP8]], align 4
// CHECK3-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
// CHECK3-NEXT: store i8* null, i8** [[TMP9]], align 4
// CHECK3-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK3-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK3-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK3-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK3-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i32 [[TMP4]]) #[[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: br label [[OMP_OFFLOAD_CONT]]
// CHECK3: omp_offload.cont:
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK3-SAME: (i32 [[I:%.*]]) #[[ATTR2:[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: [[I_ADDR:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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 [[I]], i32* [[I_ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK3-NEXT: store i32* [[I_ADDR]], i32** [[TMP0]], align 4
// CHECK3-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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:%.*]] = alloca i32*, align 4
// CHECK3-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK3-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK3-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK3-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK3-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK3-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK3-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK3-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK3-SAME: () #[[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: entry:
// CHECK3-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK3-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK4-SAME: (i32 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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 [[A]], i32* [[A_ADDR]], align 4
// CHECK4-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK4-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK4-NEXT: store i32* [[I]], i32** [[TMP1]], align 4
// CHECK4-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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:%.*]] = alloca i32*, align 4
// CHECK4-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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: [[I_CASTED:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK4-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK4-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
// CHECK4-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK4-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK4-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK4-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK4-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK4-NEXT: store i32 [[TMP3]], i32* [[I_CASTED]], align 4
// CHECK4-NEXT: [[TMP4:%.*]] = load i32, i32* [[I_CASTED]], align 4
// CHECK4-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK4-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: store i32 [[TMP4]], i32* [[TMP6]], align 4
// CHECK4-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK4-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
// CHECK4-NEXT: store i32 [[TMP4]], i32* [[TMP8]], align 4
// CHECK4-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
// CHECK4-NEXT: store i8* null, i8** [[TMP9]], align 4
// CHECK4-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK4-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK4-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK4-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK4-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i32 [[TMP4]]) #[[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: br label [[OMP_OFFLOAD_CONT]]
// CHECK4: omp_offload.cont:
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK4-SAME: (i32 [[I:%.*]]) #[[ATTR2:[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: [[I_ADDR:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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 [[I]], i32* [[I_ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK4-NEXT: store i32* [[I_ADDR]], i32** [[TMP0]], align 4
// CHECK4-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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:%.*]] = alloca i32*, align 4
// CHECK4-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK4-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK4-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK4-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK4-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK4-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK4-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK4-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code 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_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK4-SAME: () #[[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: entry:
// CHECK4-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK4-NEXT: ret void
//
//
// CHECK5-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK5-SAME: (i32 signext [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK5-NEXT: entry:
// CHECK5-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK5-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], 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
// CHECK5-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4
// CHECK5-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK5-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK5-NEXT: store i32* [[I]], i32** [[TMP1]], align 8
// CHECK5-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK5-NEXT: ret void
//
//
// CHECK5-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
// CHECK5-NEXT: entry:
// CHECK5-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK5-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, 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
// CHECK5-NEXT: [[I_CASTED:%.*]] = alloca i64, align 8
// CHECK5-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK5-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK5-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
// CHECK5-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK5-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK5-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK5-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK5-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK5-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK5-NEXT: [[CONV:%.*]] = bitcast i64* [[I_CASTED]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: store i32 [[TMP3]], i32* [[CONV]], align 4
// CHECK5-NEXT: [[TMP4:%.*]] = load i64, i64* [[I_CASTED]], align 8
// CHECK5-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK5-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: store i64 [[TMP4]], i64* [[TMP6]], align 8
// CHECK5-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK5-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
// CHECK5-NEXT: store i64 [[TMP4]], i64* [[TMP8]], align 8
// CHECK5-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
// CHECK5-NEXT: store i8* null, i8** [[TMP9]], align 8
// CHECK5-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK5-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK5-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK5-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK5-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK5: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i64 [[TMP4]]) #[[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
// CHECK5-NEXT: br label [[OMP_OFFLOAD_CONT]]
// CHECK5: omp_offload.cont:
// CHECK5-NEXT: ret void
//
//
// CHECK5-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK5-SAME: (i64 [[I:%.*]]) #[[ATTR2:[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
// CHECK5-NEXT: entry:
// CHECK5-NEXT: [[I_ADDR:%.*]] = alloca i64, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], 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
// CHECK5-NEXT: store i64 [[I]], i64* [[I_ADDR]], align 8
// CHECK5-NEXT: [[CONV:%.*]] = bitcast i64* [[I_ADDR]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK5-NEXT: store i32* [[CONV]], i32** [[TMP0]], align 8
// CHECK5-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK5-NEXT: ret void
//
//
// CHECK5-LABEL: define {{[^@]+}}@.omp_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK5-NEXT: entry:
// CHECK5-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK5-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, 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
// CHECK5-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK5-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK5-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK5-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK5-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK5-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK5-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK5-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK5-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK5-NEXT: ret void
//
//
// CHECK5-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK5-SAME: () #[[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
// CHECK5-NEXT: entry:
// CHECK5-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK5-NEXT: ret void
//
//
// CHECK6-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK6-SAME: (i32 signext [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK6-NEXT: entry:
// CHECK6-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK6-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], 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
// CHECK6-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4
// CHECK6-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK6-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK6-NEXT: store i32* [[I]], i32** [[TMP1]], align 8
// CHECK6-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK6-NEXT: ret void
//
//
// CHECK6-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
// CHECK6-NEXT: entry:
// CHECK6-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK6-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, 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
// CHECK6-NEXT: [[I_CASTED:%.*]] = alloca i64, align 8
// CHECK6-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK6-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK6-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
// CHECK6-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK6-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK6-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK6-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK6-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK6-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK6-NEXT: [[CONV:%.*]] = bitcast i64* [[I_CASTED]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: store i32 [[TMP3]], i32* [[CONV]], align 4
// CHECK6-NEXT: [[TMP4:%.*]] = load i64, i64* [[I_CASTED]], align 8
// CHECK6-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK6-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: store i64 [[TMP4]], i64* [[TMP6]], align 8
// CHECK6-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK6-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
// CHECK6-NEXT: store i64 [[TMP4]], i64* [[TMP8]], align 8
// CHECK6-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
// CHECK6-NEXT: store i8* null, i8** [[TMP9]], align 8
// CHECK6-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK6-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK6-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK6-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK6-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK6: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i64 [[TMP4]]) #[[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
// CHECK6-NEXT: br label [[OMP_OFFLOAD_CONT]]
// CHECK6: omp_offload.cont:
// CHECK6-NEXT: ret void
//
//
// CHECK6-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK6-SAME: (i64 [[I:%.*]]) #[[ATTR2:[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
// CHECK6-NEXT: entry:
// CHECK6-NEXT: [[I_ADDR:%.*]] = alloca i64, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], 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
// CHECK6-NEXT: store i64 [[I]], i64* [[I_ADDR]], align 8
// CHECK6-NEXT: [[CONV:%.*]] = bitcast i64* [[I_ADDR]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK6-NEXT: store i32* [[CONV]], i32** [[TMP0]], align 8
// CHECK6-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK6-NEXT: ret void
//
//
// CHECK6-LABEL: define {{[^@]+}}@.omp_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK6-NEXT: entry:
// CHECK6-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK6-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, 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
// CHECK6-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK6-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK6-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK6-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK6-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK6-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK6-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK6-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK6-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK6-NEXT: ret void
//
//
// CHECK6-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK6-SAME: () #[[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
// CHECK6-NEXT: entry:
// CHECK6-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK6-NEXT: ret void
//
//
// CHECK7-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK7-SAME: (i32 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK7-NEXT: entry:
// CHECK7-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK7-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK7-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4
// CHECK7-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK7-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK7-NEXT: store i32* [[I]], i32** [[TMP1]], align 4
// CHECK7-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK7-NEXT: ret void
//
//
// CHECK7-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
// CHECK7-NEXT: entry:
// CHECK7-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
// CHECK7-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK7-NEXT: [[I_CASTED:%.*]] = alloca i32, align 4
// CHECK7-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK7-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK7-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
// CHECK7-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK7-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK7-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK7-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK7-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK7-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK7-NEXT: store i32 [[TMP3]], i32* [[I_CASTED]], align 4
// CHECK7-NEXT: [[TMP4:%.*]] = load i32, i32* [[I_CASTED]], align 4
// CHECK7-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK7-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: store i32 [[TMP4]], i32* [[TMP6]], align 4
// CHECK7-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK7-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
// CHECK7-NEXT: store i32 [[TMP4]], i32* [[TMP8]], align 4
// CHECK7-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
// CHECK7-NEXT: store i8* null, i8** [[TMP9]], align 4
// CHECK7-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK7-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK7-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK7-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK7-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK7: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i32 [[TMP4]]) #[[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
// CHECK7-NEXT: br label [[OMP_OFFLOAD_CONT]]
// CHECK7: omp_offload.cont:
// CHECK7-NEXT: ret void
//
//
// CHECK7-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK7-SAME: (i32 [[I:%.*]]) #[[ATTR2:[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
// CHECK7-NEXT: entry:
// CHECK7-NEXT: [[I_ADDR:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK7-NEXT: store i32 [[I]], i32* [[I_ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK7-NEXT: store i32* [[I_ADDR]], i32** [[TMP0]], align 4
// CHECK7-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK7-NEXT: ret void
//
//
// CHECK7-LABEL: define {{[^@]+}}@.omp_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK7-NEXT: entry:
// CHECK7-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
// CHECK7-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK7-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK7-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK7-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK7-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK7-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK7-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK7-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK7-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK7-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK7-NEXT: ret void
//
//
// CHECK7-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK7-SAME: () #[[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
// CHECK7-NEXT: entry:
// CHECK7-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK7-NEXT: ret void
//
//
// CHECK8-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK8-SAME: (i32 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK8-NEXT: entry:
// CHECK8-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK8-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK8-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4
// CHECK8-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK8-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK8-NEXT: store i32* [[I]], i32** [[TMP1]], align 4
// CHECK8-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK8-NEXT: ret void
//
//
// CHECK8-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
// CHECK8-NEXT: entry:
// CHECK8-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
// CHECK8-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK8-NEXT: [[I_CASTED:%.*]] = alloca i32, align 4
// CHECK8-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK8-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK8-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
// CHECK8-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK8-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK8-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK8-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK8-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK8-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK8-NEXT: store i32 [[TMP3]], i32* [[I_CASTED]], align 4
// CHECK8-NEXT: [[TMP4:%.*]] = load i32, i32* [[I_CASTED]], align 4
// CHECK8-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK8-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: store i32 [[TMP4]], i32* [[TMP6]], align 4
// CHECK8-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK8-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
// CHECK8-NEXT: store i32 [[TMP4]], i32* [[TMP8]], align 4
// CHECK8-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
// CHECK8-NEXT: store i8* null, i8** [[TMP9]], align 4
// CHECK8-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK8-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK8-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK8-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK8-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK8: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i32 [[TMP4]]) #[[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
// CHECK8-NEXT: br label [[OMP_OFFLOAD_CONT]]
// CHECK8: omp_offload.cont:
// CHECK8-NEXT: ret void
//
//
// CHECK8-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK8-SAME: (i32 [[I:%.*]]) #[[ATTR2:[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
// CHECK8-NEXT: entry:
// CHECK8-NEXT: [[I_ADDR:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK8-NEXT: store i32 [[I]], i32* [[I_ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK8-NEXT: store i32* [[I_ADDR]], i32** [[TMP0]], align 4
// CHECK8-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK8-NEXT: ret void
//
//
// CHECK8-LABEL: define {{[^@]+}}@.omp_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK8-NEXT: entry:
// CHECK8-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
// CHECK8-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK8-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK8-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK8-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK8-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK8-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK8-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK8-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK8-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK8-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK8-NEXT: ret void
//
//
// CHECK8-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK8-SAME: () #[[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
// CHECK8-NEXT: entry:
// CHECK8-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK8-NEXT: ret void
//
//
// CHECK9-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK9-SAME: (i32 signext [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK9-NEXT: entry:
// CHECK9-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK9-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], 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
// CHECK9-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4
// CHECK9-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK9-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK9-NEXT: store i32* [[I]], i32** [[TMP1]], align 8
// CHECK9-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK9-NEXT: ret void
//
//
// CHECK9-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
// CHECK9-NEXT: entry:
// CHECK9-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK9-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, 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
// CHECK9-NEXT: [[I_CASTED:%.*]] = alloca i64, align 8
// CHECK9-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK9-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK9-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
// CHECK9-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK9-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK9-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK9-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK9-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK9-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK9-NEXT: [[CONV:%.*]] = bitcast i64* [[I_CASTED]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: store i32 [[TMP3]], i32* [[CONV]], align 4
// CHECK9-NEXT: [[TMP4:%.*]] = load i64, i64* [[I_CASTED]], align 8
// CHECK9-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK9-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: store i64 [[TMP4]], i64* [[TMP6]], align 8
// CHECK9-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK9-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
// CHECK9-NEXT: store i64 [[TMP4]], i64* [[TMP8]], align 8
// CHECK9-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
// CHECK9-NEXT: store i8* null, i8** [[TMP9]], align 8
// CHECK9-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK9-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK9-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK9-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK9-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK9: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i64 [[TMP4]]) #[[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
// CHECK9-NEXT: br label [[OMP_OFFLOAD_CONT]]
// CHECK9: omp_offload.cont:
// CHECK9-NEXT: ret void
//
//
// CHECK9-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK9-SAME: (i64 [[I:%.*]]) #[[ATTR2:[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
// CHECK9-NEXT: entry:
// CHECK9-NEXT: [[I_ADDR:%.*]] = alloca i64, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], 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
// CHECK9-NEXT: store i64 [[I]], i64* [[I_ADDR]], align 8
// CHECK9-NEXT: [[CONV:%.*]] = bitcast i64* [[I_ADDR]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK9-NEXT: store i32* [[CONV]], i32** [[TMP0]], align 8
// CHECK9-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK9-NEXT: ret void
//
//
// CHECK9-LABEL: define {{[^@]+}}@.omp_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK9-NEXT: entry:
// CHECK9-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK9-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, 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
// CHECK9-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK9-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK9-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK9-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK9-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK9-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK9-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK9-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK9-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK9-NEXT: ret void
//
//
// CHECK9-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK9-SAME: () #[[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
// CHECK9-NEXT: entry:
// CHECK9-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK9-NEXT: ret void
//
//
// CHECK10-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK10-SAME: (i32 signext [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK10-NEXT: entry:
// CHECK10-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK10-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], 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
// CHECK10-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4
// CHECK10-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK10-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK10-NEXT: store i32* [[I]], i32** [[TMP1]], align 8
// CHECK10-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK10-NEXT: ret void
//
//
// CHECK10-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
// CHECK10-NEXT: entry:
// CHECK10-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK10-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, 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
// CHECK10-NEXT: [[I_CASTED:%.*]] = alloca i64, align 8
// CHECK10-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK10-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 8
// CHECK10-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 8
// CHECK10-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK10-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK10-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK10-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK10-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK10-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK10-NEXT: [[CONV:%.*]] = bitcast i64* [[I_CASTED]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: store i32 [[TMP3]], i32* [[CONV]], align 4
// CHECK10-NEXT: [[TMP4:%.*]] = load i64, i64* [[I_CASTED]], align 8
// CHECK10-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK10-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i64*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: store i64 [[TMP4]], i64* [[TMP6]], align 8
// CHECK10-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK10-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i64*
// CHECK10-NEXT: store i64 [[TMP4]], i64* [[TMP8]], align 8
// CHECK10-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i64 0, i64 0
// CHECK10-NEXT: store i8* null, i8** [[TMP9]], align 8
// CHECK10-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK10-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK10-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK10-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK10-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK10: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i64 [[TMP4]]) #[[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
// CHECK10-NEXT: br label [[OMP_OFFLOAD_CONT]]
// CHECK10: omp_offload.cont:
// CHECK10-NEXT: ret void
//
//
// CHECK10-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK10-SAME: (i64 [[I:%.*]]) #[[ATTR2:[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
// CHECK10-NEXT: entry:
// CHECK10-NEXT: [[I_ADDR:%.*]] = alloca i64, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], 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
// CHECK10-NEXT: store i64 [[I]], i64* [[I_ADDR]], align 8
// CHECK10-NEXT: [[CONV:%.*]] = bitcast i64* [[I_ADDR]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK10-NEXT: store i32* [[CONV]], i32** [[TMP0]], align 8
// CHECK10-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK10-NEXT: ret void
//
//
// CHECK10-LABEL: define {{[^@]+}}@.omp_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK10-NEXT: entry:
// CHECK10-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK10-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, 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
// CHECK10-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK10-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK10-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK10-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 8
// CHECK10-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK10-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 8
// CHECK10-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK10-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK10-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK10-NEXT: ret void
//
//
// CHECK10-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK10-SAME: () #[[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
// CHECK10-NEXT: entry:
// CHECK10-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK10-NEXT: ret void
//
//
// CHECK11-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK11-SAME: (i32 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK11-NEXT: entry:
// CHECK11-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK11-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK11-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4
// CHECK11-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK11-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK11-NEXT: store i32* [[I]], i32** [[TMP1]], align 4
// CHECK11-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK11-NEXT: ret void
//
//
// CHECK11-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
// CHECK11-NEXT: entry:
// CHECK11-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
// CHECK11-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK11-NEXT: [[I_CASTED:%.*]] = alloca i32, align 4
// CHECK11-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK11-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK11-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
// CHECK11-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK11-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK11-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK11-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK11-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK11-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK11-NEXT: store i32 [[TMP3]], i32* [[I_CASTED]], align 4
// CHECK11-NEXT: [[TMP4:%.*]] = load i32, i32* [[I_CASTED]], align 4
// CHECK11-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK11-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: store i32 [[TMP4]], i32* [[TMP6]], align 4
// CHECK11-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK11-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
// CHECK11-NEXT: store i32 [[TMP4]], i32* [[TMP8]], align 4
// CHECK11-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
// CHECK11-NEXT: store i8* null, i8** [[TMP9]], align 4
// CHECK11-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK11-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK11-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK11-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK11-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK11: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i32 [[TMP4]]) #[[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
// CHECK11-NEXT: br label [[OMP_OFFLOAD_CONT]]
// CHECK11: omp_offload.cont:
// CHECK11-NEXT: ret void
//
//
// CHECK11-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK11-SAME: (i32 [[I:%.*]]) #[[ATTR2:[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
// CHECK11-NEXT: entry:
// CHECK11-NEXT: [[I_ADDR:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK11-NEXT: store i32 [[I]], i32* [[I_ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK11-NEXT: store i32* [[I_ADDR]], i32** [[TMP0]], align 4
// CHECK11-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK11-NEXT: ret void
//
//
// CHECK11-LABEL: define {{[^@]+}}@.omp_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK11-NEXT: entry:
// CHECK11-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
// CHECK11-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK11-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK11-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK11-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK11-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK11-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK11-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK11-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK11-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK11-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK11-NEXT: ret void
//
//
// CHECK11-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK11-SAME: () #[[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
// CHECK11-NEXT: entry:
// CHECK11-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK11-NEXT: ret void
//
//
// CHECK12-LABEL: define {{[^@]+}}@_Z28implicit_maps_nested_integeri
// CHECK12-SAME: (i32 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK12-NEXT: entry:
// CHECK12-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4
// CHECK12-NEXT: [[I:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK12-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4
// CHECK12-NEXT: [[TMP0:%.*]] = load i32, i32* [[A_ADDR]], align 4
// CHECK12-NEXT: store i32 [[TMP0]], i32* [[I]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK12-NEXT: store i32* [[I]], i32** [[TMP1]], align 4
// CHECK12-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK12-NEXT: ret void
//
//
// CHECK12-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[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
// CHECK12-NEXT: entry:
// CHECK12-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
// CHECK12-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK12-NEXT: [[I_CASTED:%.*]] = alloca i32, align 4
// CHECK12-NEXT: [[DOTOFFLOAD_BASEPTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK12-NEXT: [[DOTOFFLOAD_PTRS:%.*]] = alloca [1 x i8*], align 4
// CHECK12-NEXT: [[DOTOFFLOAD_MAPPERS:%.*]] = alloca [1 x i8*], align 4
// CHECK12-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK12-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK12-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 4
// CHECK12-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK12-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK12-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK12-NEXT: store i32 [[TMP3]], i32* [[I_CASTED]], align 4
// CHECK12-NEXT: [[TMP4:%.*]] = load i32, i32* [[I_CASTED]], align 4
// CHECK12-NEXT: [[TMP5:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], 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
// CHECK12-NEXT: [[TMP6:%.*]] = bitcast i8** [[TMP5]] to i32*
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: store i32 [[TMP4]], i32* [[TMP6]], align 4
// CHECK12-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK12-NEXT: [[TMP8:%.*]] = bitcast i8** [[TMP7]] to i32*
// CHECK12-NEXT: store i32 [[TMP4]], i32* [[TMP8]], align 4
// CHECK12-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_MAPPERS]], i32 0, i32 0
// CHECK12-NEXT: store i8* null, i8** [[TMP9]], align 4
// CHECK12-NEXT: [[TMP10:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_BASEPTRS]], i32 0, i32 0
// CHECK12-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOFFLOAD_PTRS]], i32 0, i32 0
// CHECK12-NEXT: [[TMP12:%.*]] = call i32 @__tgt_target_teams_mapper(%struct.ident_t* @[[GLOB1]], i64 -1, i8* @.{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48.region_id, i32 1, i8** [[TMP10]], i8** [[TMP11]], i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_sizes, i32 0, i32 0), i64* getelementptr inbounds ([1 x i64], [1 x i64]* @.offload_maptypes, i32 0, i32 0), i8** null, i8** null, i32 1, i32 0)
// CHECK12-NEXT: [[TMP13:%.*]] = icmp ne i32 [[TMP12]], 0
// CHECK12-NEXT: br i1 [[TMP13]], label [[OMP_OFFLOAD_FAILED:%.*]], label [[OMP_OFFLOAD_CONT:%.*]]
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK12: omp_offload.failed:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48(i32 [[TMP4]]) #[[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
// CHECK12-NEXT: br label [[OMP_OFFLOAD_CONT]]
// CHECK12: omp_offload.cont:
// CHECK12-NEXT: ret void
//
//
// CHECK12-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z28implicit_maps_nested_integeri_l48
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK12-SAME: (i32 [[I:%.*]]) #[[ATTR2:[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
// CHECK12-NEXT: entry:
// CHECK12-NEXT: [[I_ADDR:%.*]] = alloca i32, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON_0:%.*]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK12-NEXT: store i32 [[I]], i32* [[I_ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON_0]], %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK12-NEXT: store i32* [[I_ADDR]], i32** [[TMP0]], align 4
// CHECK12-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*, %struct.anon.0*)* @.omp_outlined..1 to void (i32*, i32*, ...)*), %struct.anon.0* [[OMP_OUTLINED_ARG_AGG_]])
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK12-NEXT: ret void
//
//
// CHECK12-LABEL: define {{[^@]+}}@.omp_outlined..1
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon.0* noalias [[__CONTEXT:%.*]]) #[[ATTR1]] {
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK12-NEXT: entry:
// CHECK12-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 4
// CHECK12-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon.0*, align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK12-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 4
// CHECK12-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK12-NEXT: store %struct.anon.0* [[__CONTEXT]], %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK12-NEXT: [[TMP0:%.*]] = load %struct.anon.0*, %struct.anon.0** [[__CONTEXT_ADDR]], align 4
// CHECK12-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON_0:%.*]], %struct.anon.0* [[TMP0]], i32 0, i32 0
// CHECK12-NEXT: [[TMP2:%.*]] = load i32*, i32** [[TMP1]], align 4
// CHECK12-NEXT: [[TMP3:%.*]] = load i32, i32* [[TMP2]], align 4
// CHECK12-NEXT: [[INC:%.*]] = add nsw i32 [[TMP3]], 1
// CHECK12-NEXT: store i32 [[INC]], i32* [[TMP2]], align 4
[OpenMP][NFC] Refactor Clang OpenMP tests using update_cc_test_checks This patch refactors a subset of Clang OpenMP tests, generating checklines using the update_cc_test_checks script. This refactoring facilitates updating the Clang OpenMP code generation codebase by automating test generation. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D101849
2021-05-06 06:13:14 +08:00
// CHECK12-NEXT: ret void
//
//
// CHECK12-LABEL: define {{[^@]+}}@.omp_offloading.requires_reg
[OpenMP] Creating the `omp_target_num_teams` and `omp_target_thread_limit` attributes to outlined functions The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant, these calls can be folded to the constant value. In commit D106033 we have the optimization phase. This commit adds the attributes to the outlined function for the grid size. the two attributes are `omp_target_num_teams` and `omp_target_thread_limit`. These values are added as long as they are constant. Two functions are created `getNumThreadsExprForTargetDirective` and `getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective` and `emitNumThreadsForTargetDirective` identify the expresion and emit the code. However, for the Device version of the outlined function, we cannot emit anything. Therefore, this is a first attempt to separate emision of code from deduction of the values. Reviewed By: jdoerfert Differential Revision: https://reviews.llvm.org/D106298
2021-07-28 05:20:47 +08:00
// CHECK12-SAME: () #[[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
// CHECK12-NEXT: entry:
// CHECK12-NEXT: call void @__tgt_register_requires(i64 1)
// CHECK12-NEXT: ret void
//