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/for_reduction_codegen.cpp

1244 lines
63 KiB
C++
Raw Normal View History

[OPENMP] -fopenmp enables OpenMP support (fix for http://llvm.org/PR23492) -fopenmp turns on OpenMP support and links libiomp5 as OpenMP library. Also there is -fopenmp={libiomp5|libgomp} option that allows to override effect of -fopenmp and link libgomp library (if -fopenmp=libgomp is specified). Differential Revision: http://reviews.llvm.org/D9736 llvm-svn: 237769
2015-05-20 12:24:19 +08:00
// RUN: %clang_cc1 -verify -fopenmp -x c++ -triple x86_64-apple-darwin10 -emit-llvm %s -o - | FileCheck %s
// RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple x86_64-apple-darwin10 -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-apple-darwin10 -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s
// RUN: %clang_cc1 -verify -fopenmp -x c++ -std=c++11 -DLAMBDA -triple x86_64-apple-darwin10 -emit-llvm %s -o - | FileCheck -check-prefix=LAMBDA %s
// RUN: %clang_cc1 -verify -fopenmp -x c++ -fblocks -DBLOCKS -triple x86_64-apple-darwin10 -emit-llvm %s -o - | FileCheck -check-prefix=BLOCKS %s
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// expected-no-diagnostics
Revert "Revert r241620 and follow-up commits" and move the initialization of the llvm targets from clang/CodeGen into ClangCheck.cpp and CIndex.cpp. llvm-svn: 241653
2015-07-08 09:00:30 +08:00
// REQUIRES: x86-registered-target
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
#ifndef HEADER
#define HEADER
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
volatile double g, g_orig;
volatile double &g1 = g_orig;
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
template <class T>
struct S {
T f;
S(T a) : f(a + g) {}
S() : f(g) {}
operator T() { return T(); }
S &operator&(const S &) { return *this; }
~S() {}
};
// CHECK-DAG: [[S_FLOAT_TY:%.+]] = type { float }
// CHECK-DAG: [[S_INT_TY:%.+]] = type { i{{[0-9]+}} }
// CHECK-DAG: [[ATOMIC_REDUCE_BARRIER_LOC:@.+]] = private unnamed_addr constant %{{.+}} { i32 0, i32 18, i32 0, i32 0, i8*
// CHECK-DAG: [[IMPLICIT_BARRIER_LOC:@.+]] = private unnamed_addr constant %{{.+}} { i32 0, i32 66, i32 0, i32 0, i8*
// CHECK-DAG: [[REDUCTION_LOC:@.+]] = private unnamed_addr constant %{{.+}} { i32 0, i32 18, i32 0, i32 0, i8*
// CHECK-DAG: [[REDUCTION_LOCK:@.+]] = common global [8 x i32] zeroinitializer
template <typename T>
T tmain() {
T t;
S<T> test;
T t_var = T(), t_var1;
T vec[] = {1, 2};
S<T> s_arr[] = {1, 2};
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
S<T> &var = test;
S<T> var1;
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
#pragma omp parallel
#pragma omp for reduction(+:t_var) reduction(&:var) reduction(&& : var1) reduction(min: t_var1) nowait
for (int i = 0; i < 2; ++i) {
vec[i] = t_var;
s_arr[i] = var;
}
[OPENMP] Fix crash on reductions codegen for short circuit reduction operations. llvm-svn: 236084
2015-04-29 13:21:03 +08:00
#pragma omp parallel
#pragma omp for reduction(&& : t_var)
for (int i = 0; i < 2; ++i) {
vec[i] = t_var;
s_arr[i] = var;
}
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
return T();
}
[OPENMP 4.0] Fixed support of array sections/array subscripts. Codegen for array sections/array subscripts worked only for expressions with arrays as base. Patch fixes codegen for bases with pointer/reference types. llvm-svn: 259776
2016-02-04 19:27:03 +08:00
extern S<float> **foo();
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
int main() {
#ifdef LAMBDA
// LAMBDA: [[G:@.+]] = global double
// LAMBDA-LABEL: @main
// LAMBDA: call void [[OUTER_LAMBDA:@.+]](
[&]() {
// LAMBDA: define{{.*}} internal{{.*}} void [[OUTER_LAMBDA]](
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// LAMBDA: call void {{.+}} @__kmpc_fork_call({{.+}}, i32 0, {{.+}}* [[OMP_REGION:@.+]] to {{.+}})
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
#pragma omp parallel
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
#pragma omp for reduction(+:g, g1)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
for (int i = 0; i < 2; ++i) {
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// LAMBDA: define{{.*}} internal{{.*}} void [[OMP_REGION]](i32* noalias %{{.+}}, i32* noalias %{{.+}})
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// LAMBDA: [[G_PRIVATE_ADDR:%.+]] = alloca double,
// Reduction list for runtime.
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
// LAMBDA: [[RED_LIST:%.+]] = alloca [2 x i8*],
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// LAMBDA: store double 0.0{{.+}}, double* [[G_PRIVATE_ADDR]]
// LAMBDA: call void @__kmpc_for_static_init_4(
g = 1;
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
g1 = 1;
[OPENMP] Drop type qualifiers from private variables. If the variable is marked as private in OpenMP construct, the reference to this variable should not keep type qualifiers for the original variable. Private copy is not volatile or constant, so we can use unqualified type for private copy. llvm-svn: 242133
2015-07-14 18:32:29 +08:00
// LAMBDA: store double 1.0{{.+}}, double* [[G_PRIVATE_ADDR]],
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// LAMBDA: [[G_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
// LAMBDA: store double* [[G_PRIVATE_ADDR]], double** [[G_PRIVATE_ADDR_REF]]
// LAMBDA: call void [[INNER_LAMBDA:@.+]](%{{.+}}* [[ARG]])
// LAMBDA: call void @__kmpc_for_static_fini(
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// LAMBDA: [[G_PRIV_REF:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[RED_LIST]], i64 0, i64 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// LAMBDA: [[BITCAST:%.+]] = bitcast double* [[G_PRIVATE_ADDR]] to i8*
// LAMBDA: store i8* [[BITCAST]], i8** [[G_PRIV_REF]],
// LAMBDA: call i32 @__kmpc_reduce(
// LAMBDA: switch i32 %{{.+}}, label %[[REDUCTION_DONE:.+]] [
// LAMBDA: i32 1, label %[[CASE1:.+]]
// LAMBDA: i32 2, label %[[CASE2:.+]]
// LAMBDA: [[CASE1]]
// LAMBDA: [[G_VAL:%.+]] = load double, double* [[G]]
// LAMBDA: [[G_PRIV_VAL:%.+]] = load double, double* [[G_PRIVATE_ADDR]]
// LAMBDA: [[ADD:%.+]] = fadd double [[G_VAL]], [[G_PRIV_VAL]]
// LAMBDA: store double [[ADD]], double* [[G]]
// LAMBDA: call void @__kmpc_end_reduce(
// LAMBDA: br label %[[REDUCTION_DONE]]
// LAMBDA: [[CASE2]]
// LAMBDA: [[G_PRIV_VAL:%.+]] = load double, double* [[G_PRIVATE_ADDR]]
// LAMBDA: fadd double
// LAMBDA: cmpxchg i64*
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// LAMBDA: call void @__kmpc_end_reduce(
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// LAMBDA: br label %[[REDUCTION_DONE]]
// LAMBDA: [[REDUCTION_DONE]]
// LAMBDA: ret void
[&]() {
// LAMBDA: define {{.+}} void [[INNER_LAMBDA]](%{{.+}}* [[ARG_PTR:%.+]])
// LAMBDA: store %{{.+}}* [[ARG_PTR]], %{{.+}}** [[ARG_PTR_REF:%.+]],
g = 2;
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
g1 = 2;
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// LAMBDA: [[ARG_PTR:%.+]] = load %{{.+}}*, %{{.+}}** [[ARG_PTR_REF]]
// LAMBDA: [[G_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
// LAMBDA: [[G_REF:%.+]] = load double*, double** [[G_PTR_REF]]
[OPENMP] Drop type qualifiers from private variables. If the variable is marked as private in OpenMP construct, the reference to this variable should not keep type qualifiers for the original variable. Private copy is not volatile or constant, so we can use unqualified type for private copy. llvm-svn: 242133
2015-07-14 18:32:29 +08:00
// LAMBDA: store double 2.0{{.+}}, double* [[G_REF]]
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
}();
}
}();
return 0;
#elif defined(BLOCKS)
// BLOCKS: [[G:@.+]] = global double
// BLOCKS-LABEL: @main
// BLOCKS: call void {{%.+}}(i8
^{
// BLOCKS: define{{.*}} internal{{.*}} void {{.+}}(i8*
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// BLOCKS: call void {{.+}} @__kmpc_fork_call({{.+}}, i32 0, {{.+}}* [[OMP_REGION:@.+]] to {{.+}})
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
#pragma omp parallel
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
#pragma omp for reduction(-:g, g1)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
for (int i = 0; i < 2; ++i) {
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// BLOCKS: define{{.*}} internal{{.*}} void [[OMP_REGION]](i32* noalias %{{.+}}, i32* noalias %{{.+}})
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// BLOCKS: [[G_PRIVATE_ADDR:%.+]] = alloca double,
// Reduction list for runtime.
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
// BLOCKS: [[RED_LIST:%.+]] = alloca [2 x i8*],
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// BLOCKS: store double 0.0{{.+}}, double* [[G_PRIVATE_ADDR]]
g = 1;
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
g1 = 1;
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// BLOCKS: call void @__kmpc_for_static_init_4(
[OPENMP] Drop type qualifiers from private variables. If the variable is marked as private in OpenMP construct, the reference to this variable should not keep type qualifiers for the original variable. Private copy is not volatile or constant, so we can use unqualified type for private copy. llvm-svn: 242133
2015-07-14 18:32:29 +08:00
// BLOCKS: store double 1.0{{.+}}, double* [[G_PRIVATE_ADDR]],
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// BLOCKS-NOT: [[G]]{{[[^:word:]]}}
// BLOCKS: double* [[G_PRIVATE_ADDR]]
// BLOCKS-NOT: [[G]]{{[[^:word:]]}}
// BLOCKS: call void {{%.+}}(i8
// BLOCKS: call void @__kmpc_for_static_fini(
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// BLOCKS: [[G_PRIV_REF:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[RED_LIST]], i64 0, i64 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// BLOCKS: [[BITCAST:%.+]] = bitcast double* [[G_PRIVATE_ADDR]] to i8*
// BLOCKS: store i8* [[BITCAST]], i8** [[G_PRIV_REF]],
// BLOCKS: call i32 @__kmpc_reduce(
// BLOCKS: switch i32 %{{.+}}, label %[[REDUCTION_DONE:.+]] [
// BLOCKS: i32 1, label %[[CASE1:.+]]
// BLOCKS: i32 2, label %[[CASE2:.+]]
// BLOCKS: [[CASE1]]
// BLOCKS: [[G_VAL:%.+]] = load double, double* [[G]]
// BLOCKS: [[G_PRIV_VAL:%.+]] = load double, double* [[G_PRIVATE_ADDR]]
// BLOCKS: [[ADD:%.+]] = fadd double [[G_VAL]], [[G_PRIV_VAL]]
// BLOCKS: store double [[ADD]], double* [[G]]
// BLOCKS: call void @__kmpc_end_reduce(
// BLOCKS: br label %[[REDUCTION_DONE]]
// BLOCKS: [[CASE2]]
// BLOCKS: [[G_PRIV_VAL:%.+]] = load double, double* [[G_PRIVATE_ADDR]]
// BLOCKS: fadd double
// BLOCKS: cmpxchg i64*
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// BLOCKS: call void @__kmpc_end_reduce(
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// BLOCKS: br label %[[REDUCTION_DONE]]
// BLOCKS: [[REDUCTION_DONE]]
// BLOCKS: ret void
^{
// BLOCKS: define {{.+}} void {{@.+}}(i8*
g = 2;
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
g1 = 2;
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// BLOCKS-NOT: [[G]]{{[[^:word:]]}}
[OPENMP] Drop type qualifiers from private variables. If the variable is marked as private in OpenMP construct, the reference to this variable should not keep type qualifiers for the original variable. Private copy is not volatile or constant, so we can use unqualified type for private copy. llvm-svn: 242133
2015-07-14 18:32:29 +08:00
// BLOCKS: store double 2.0{{.+}}, double*
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// BLOCKS-NOT: [[G]]{{[[^:word:]]}}
// BLOCKS: ret
}();
}
}();
return 0;
#else
S<float> test;
float t_var = 0, t_var1;
int vec[] = {1, 2};
S<float> s_arr[] = {1, 2};
[OPENMP] Fix for http://llvm.org/PR24674: assertion failed and and abort trap Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later. llvm-svn: 246846
2015-09-04 19:26:21 +08:00
S<float> &var = test;
[OPENMP 4.1] Codegen for array sections/subscripts in 'reduction' clause. OpenMP 4.1 adds support for array sections/subscripts in 'reduction' clause. Patch adds codegen for this feature. llvm-svn: 249672
2015-10-08 17:10:53 +08:00
S<float> var1, arrs[10][4];
[OPENMP 4.0] Fixed support of array sections/array subscripts. Codegen for array sections/array subscripts worked only for expressions with arrays as base. Patch fixes codegen for bases with pointer/reference types. llvm-svn: 259776
2016-02-04 19:27:03 +08:00
S<float> **var2 = foo();
S<float> vvar2[2];
S<float> (&var3)[2] = s_arr;
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
#pragma omp parallel
#pragma omp for reduction(+:t_var) reduction(&:var) reduction(&& : var1) reduction(min: t_var1)
for (int i = 0; i < 2; ++i) {
vec[i] = t_var;
s_arr[i] = var;
}
[OPENMP 4.1] Codegen for array sections/subscripts in 'reduction' clause. OpenMP 4.1 adds support for array sections/subscripts in 'reduction' clause. Patch adds codegen for this feature. llvm-svn: 249672
2015-10-08 17:10:53 +08:00
int arr[10][vec[1]];
#pragma omp parallel for reduction(+:arr[1][:vec[1]]) reduction(&:arrs[1:vec[1]][1:2])
for (int i = 0; i < 10; ++i)
++arr[1][i];
[OPENMP 4.5] Allow arrays in 'reduction' clause. OpenMP 4.5, alogn with array sections, allows to use variables of array type in reductions. llvm-svn: 258804
2016-01-26 20:20:39 +08:00
#pragma omp parallel
#pragma omp for reduction(+:arr) reduction(&:arrs)
for (int i = 0; i < 10; ++i)
++arr[1][i];
[OPENMP 4.0] Fixed support of array sections/array subscripts. Codegen for array sections/array subscripts worked only for expressions with arrays as base. Patch fixes codegen for bases with pointer/reference types. llvm-svn: 259776
2016-02-04 19:27:03 +08:00
#pragma omp parallel
#pragma omp for reduction(& : var2[0 : 5][1 : 6])
for (int i = 0; i < 10; ++i)
;
#pragma omp parallel
#pragma omp for reduction(& : vvar2[0 : 5])
for (int i = 0; i < 10; ++i)
;
#pragma omp parallel
#pragma omp for reduction(& : var3[1 : 2])
for (int i = 0; i < 10; ++i)
;
#pragma omp parallel
#pragma omp for reduction(& : var3)
for (int i = 0; i < 10; ++i)
;
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
return tmain<int>();
#endif
}
// CHECK: define {{.*}}i{{[0-9]+}} @main()
// CHECK: [[TEST:%.+]] = alloca [[S_FLOAT_TY]],
// CHECK: call {{.*}} [[S_FLOAT_TY_CONSTR:@.+]]([[S_FLOAT_TY]]* [[TEST]])
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 6, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, float*, [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]*, float*, [2 x i32]*, [2 x [[S_FLOAT_TY]]]*)* [[MAIN_MICROTASK:@.+]] to void
[OPENMP 4.1] Codegen for array sections/subscripts in 'reduction' clause. OpenMP 4.1 adds support for array sections/subscripts in 'reduction' clause. Patch adds codegen for this feature. llvm-svn: 249672
2015-10-08 17:10:53 +08:00
// CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 5, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, i64, i64, i32*, [2 x i32]*, [10 x [4 x [[S_FLOAT_TY]]]]*)* [[MAIN_MICROTASK1:@.+]] to void
[OPENMP 4.5] Allow arrays in 'reduction' clause. OpenMP 4.5, alogn with array sections, allows to use variables of array type in reductions. llvm-svn: 258804
2016-01-26 20:20:39 +08:00
// CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 4, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, i64, i64, i32*, [10 x [4 x [[S_FLOAT_TY]]]]*)* [[MAIN_MICROTASK2:@.+]] to void
[OPENMP 4.0] Fixed support of array sections/array subscripts. Codegen for array sections/array subscripts worked only for expressions with arrays as base. Patch fixes codegen for bases with pointer/reference types. llvm-svn: 259776
2016-02-04 19:27:03 +08:00
// CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 1, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, [[S_FLOAT_TY]]***)* [[MAIN_MICROTASK3:@.+]] to void
// CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 1, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, [2 x [[S_FLOAT_TY]]]*)* [[MAIN_MICROTASK4:@.+]] to void
// CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 1, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, [2 x [[S_FLOAT_TY]]]*)* [[MAIN_MICROTASK5:@.+]] to void
// CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 1, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, [2 x [[S_FLOAT_TY]]]*)* [[MAIN_MICROTASK6:@.+]] to void
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: = call {{.*}}i{{.+}} [[TMAIN_INT:@.+]]()
// CHECK: call {{.*}} [[S_FLOAT_TY_DESTR:@.+]]([[S_FLOAT_TY]]*
// CHECK: ret
//
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// CHECK: define internal void [[MAIN_MICROTASK]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, float* dereferenceable(4) %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(4) %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(4) %{{.+}}, float* dereferenceable(4) %{{.+}}, [2 x i32]* dereferenceable(8) %vec, [2 x [[S_FLOAT_TY]]]* dereferenceable(8) %{{.+}})
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR_PRIV:%.+]] = alloca float,
// CHECK: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]],
// CHECK: [[VAR1_PRIV:%.+]] = alloca [[S_FLOAT_TY]],
// CHECK: [[T_VAR1_PRIV:%.+]] = alloca float,
// Reduction list for runtime.
// CHECK: [[RED_LIST:%.+]] = alloca [4 x i8*],
// CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]],
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// CHECK: [[T_VAR_REF:%.+]] = load float*, float** %
// CHECK: [[VAR1_REF:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** %
// CHECK: [[T_VAR1_REF:%.+]] = load float*, float** %
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// For + reduction operation initial value of private variable is 0.
// CHECK: store float 0.0{{.+}}, float* [[T_VAR_PRIV]],
// For & reduction operation initial value of private variable is ones in all bits.
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// CHECK: [[VAR_REF:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** %
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: call {{.*}} [[S_FLOAT_TY_CONSTR:@.+]]([[S_FLOAT_TY]]* [[VAR_PRIV]])
// For && reduction operation initial value of private variable is 1.0.
// CHECK: call {{.*}} [[S_FLOAT_TY_CONSTR:@.+]]([[S_FLOAT_TY]]* [[VAR1_PRIV]])
// For min reduction operation initial value of private variable is largest repesentable value.
// CHECK: store float 0x47EFFFFFE0000000, float* [[T_VAR1_PRIV]],
// CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]]
// CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]]
// CHECK: call void @__kmpc_for_static_init_4(
// Skip checks for internal operations.
// CHECK: call void @__kmpc_for_static_fini(
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BITCAST:%.+]] = bitcast float* [[T_VAR_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[T_VAR_PRIV_REF]],
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 1
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[VAR_PRIV_REF]],
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR1_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 2
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR1_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[VAR1_PRIV_REF]],
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR1_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 3
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BITCAST:%.+]] = bitcast float* [[T_VAR1_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[T_VAR1_PRIV_REF]],
// res = __kmpc_reduce(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>);
// CHECK: [[BITCAST:%.+]] = bitcast [4 x i8*]* [[RED_LIST]] to i8*
// CHECK: [[RES:%.+]] = call i32 @__kmpc_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], i32 4, i64 32, i8* [[BITCAST]], void (i8*, i8*)* [[REDUCTION_FUNC:@.+]], [8 x i32]* [[REDUCTION_LOCK]])
// switch(res)
// CHECK: switch i32 [[RES]], label %[[RED_DONE:.+]] [
// CHECK: i32 1, label %[[CASE1:.+]]
// CHECK: i32 2, label %[[CASE2:.+]]
// CHECK: ]
// case 1:
// t_var += t_var_reduction;
// CHECK: [[T_VAR_VAL:%.+]] = load float, float* [[T_VAR_REF]],
// CHECK: [[T_VAR_PRIV_VAL:%.+]] = load float, float* [[T_VAR_PRIV]],
// CHECK: [[UP:%.+]] = fadd float [[T_VAR_VAL]], [[T_VAR_PRIV_VAL]]
// CHECK: store float [[UP]], float* [[T_VAR_REF]],
// var = var.operator &(var_reduction);
// CHECK: [[UP:%.+]] = call dereferenceable(4) [[S_FLOAT_TY]]* @{{.+}}([[S_FLOAT_TY]]* [[VAR_REF]], [[S_FLOAT_TY]]* dereferenceable(4) [[VAR_PRIV]])
// CHECK: [[BC1:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_REF]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_FLOAT_TY]]* [[UP]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// var1 = var1.operator &&(var1_reduction);
// CHECK: [[TO_FLOAT:%.+]] = call float @{{.+}}([[S_FLOAT_TY]]* [[VAR1_REF]])
// CHECK: [[VAR1_BOOL:%.+]] = fcmp une float [[TO_FLOAT]], 0.0
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[VAR1_BOOL]], label %[[TRUE:.+]], label %[[END2:.+]]
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[TRUE]]
// CHECK: [[TO_FLOAT:%.+]] = call float @{{.+}}([[S_FLOAT_TY]]* [[VAR1_PRIV]])
// CHECK: [[VAR1_REDUCTION_BOOL:%.+]] = fcmp une float [[TO_FLOAT]], 0.0
// CHECK: br label %[[END2]]
// CHECK: [[END2]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: [[COND_LVALUE:%.+]] = phi i1 [ false, %{{.+}} ], [ [[VAR1_REDUCTION_BOOL]], %[[TRUE]] ]
// CHECK: [[CONV:%.+]] = uitofp i1 [[COND_LVALUE]] to float
// CHECK: call void @{{.+}}([[S_FLOAT_TY]]* [[COND_LVALUE:%.+]], float [[CONV]])
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BC1:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR1_REF]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_FLOAT_TY]]* [[COND_LVALUE]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// t_var1 = min(t_var1, t_var1_reduction);
// CHECK: [[T_VAR1_VAL:%.+]] = load float, float* [[T_VAR1_REF]],
// CHECK: [[T_VAR1_PRIV_VAL:%.+]] = load float, float* [[T_VAR1_PRIV]],
// CHECK: [[CMP:%.+]] = fcmp olt float [[T_VAR1_VAL]], [[T_VAR1_PRIV_VAL]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[CMP]]
// CHECK: [[UP:%.+]] = phi float
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: store float [[UP]], float* [[T_VAR1_REF]],
// __kmpc_end_reduce(<loc>, <gtid>, &<lock>);
// CHECK: call void @__kmpc_end_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]])
// break;
// CHECK: br label %[[RED_DONE]]
// case 2:
// t_var += t_var_reduction;
// CHECK: load float, float* [[T_VAR_PRIV]]
// CHECK: [[T_VAR_REF_INT:%.+]] = bitcast float* [[T_VAR_REF]] to i32*
[OPENMP] Fixed bug in atomic update/capture/write constructs. Fixed a bug with codegen for destination atomic l-value with padding and junk in this padding bytes. llvm-svn: 237422
2015-05-15 16:36:34 +08:00
// CHECK: [[OLD1:%.+]] = load atomic i32, i32* [[T_VAR_REF_INT]] monotonic,
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: br label %[[CONT:.+]]
// CHECK: [[CONT]]
[OPENMP] Fixed bug in atomic update/capture/write constructs. Fixed a bug with codegen for destination atomic l-value with padding and junk in this padding bytes. llvm-svn: 237422
2015-05-15 16:36:34 +08:00
// CHECK: [[ORIG_OLD_INT:%.+]] = phi i32 [ [[OLD1]], %{{.+}} ], [ [[OLD2:%.+]], %[[CONT]] ]
// CHECK: fadd float
// CHECK: [[UP_INT:%.+]] = load i32, i32*
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR_REF_INT:%.+]] = bitcast float* [[T_VAR_REF]] to i32*
// CHECK: [[RES:%.+]] = cmpxchg i32* [[T_VAR_REF_INT]], i32 [[ORIG_OLD_INT]], i32 [[UP_INT]] monotonic monotonic
[OPENMP] Fixed bug in atomic update/capture/write constructs. Fixed a bug with codegen for destination atomic l-value with padding and junk in this padding bytes. llvm-svn: 237422
2015-05-15 16:36:34 +08:00
// CHECK: [[OLD2:%.+]] = extractvalue { i32, i1 } [[RES]], 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[SUCCESS_FAIL:%.+]] = extractvalue { i32, i1 } [[RES]], 1
// CHECK: br i1 [[SUCCESS_FAIL]], label %[[ATOMIC_DONE:.+]], label %[[CONT]]
// CHECK: [[ATOMIC_DONE]]
// var = var.operator &(var_reduction);
// CHECK: call void @__kmpc_critical(
// CHECK: [[UP:%.+]] = call dereferenceable(4) [[S_FLOAT_TY]]* @{{.+}}([[S_FLOAT_TY]]* [[VAR_REF]], [[S_FLOAT_TY]]* dereferenceable(4) [[VAR_PRIV]])
// CHECK: [[BC1:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_REF]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_FLOAT_TY]]* [[UP]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: call void @__kmpc_end_critical(
// var1 = var1.operator &&(var1_reduction);
// CHECK: call void @__kmpc_critical(
// CHECK: [[TO_FLOAT:%.+]] = call float @{{.+}}([[S_FLOAT_TY]]* [[VAR1_REF]])
// CHECK: [[VAR1_BOOL:%.+]] = fcmp une float [[TO_FLOAT]], 0.0
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[VAR1_BOOL]], label %[[TRUE:.+]], label %[[END2:.+]]
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[TRUE]]
// CHECK: [[TO_FLOAT:%.+]] = call float @{{.+}}([[S_FLOAT_TY]]* [[VAR1_PRIV]])
// CHECK: [[VAR1_REDUCTION_BOOL:%.+]] = fcmp une float [[TO_FLOAT]], 0.0
// CHECK: br label %[[END2]]
// CHECK: [[END2]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: [[COND_LVALUE:%.+]] = phi i1 [ false, %{{.+}} ], [ [[VAR1_REDUCTION_BOOL]], %[[TRUE]] ]
// CHECK: [[CONV:%.+]] = uitofp i1 [[COND_LVALUE]] to float
// CHECK: call void @{{.+}}([[S_FLOAT_TY]]* [[COND_LVALUE:%.+]], float [[CONV]])
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BC1:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR1_REF]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_FLOAT_TY]]* [[COND_LVALUE]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: call void @__kmpc_end_critical(
// t_var1 = min(t_var1, t_var1_reduction);
// CHECK: load float, float* [[T_VAR1_PRIV]]
// CHECK: [[T_VAR1_REF_INT:%.+]] = bitcast float* [[T_VAR1_REF]] to i32*
[OPENMP] Fixed bug in atomic update/capture/write constructs. Fixed a bug with codegen for destination atomic l-value with padding and junk in this padding bytes. llvm-svn: 237422
2015-05-15 16:36:34 +08:00
// CHECK: [[OLD1:%.+]] = load atomic i32, i32* [[T_VAR1_REF_INT]] monotonic,
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: br label %[[CONT:.+]]
// CHECK: [[CONT]]
[OPENMP] Fixed bug in atomic update/capture/write constructs. Fixed a bug with codegen for destination atomic l-value with padding and junk in this padding bytes. llvm-svn: 237422
2015-05-15 16:36:34 +08:00
// CHECK: [[ORIG_OLD_INT:%.+]] = phi i32 [ [[OLD1]], %{{.+}} ], [ [[OLD2:%.+]], %{{.+}} ]
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[CMP:%.+]] = fcmp olt float
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[CMP]]
[OPENMP] Fixed bug in atomic update/capture/write constructs. Fixed a bug with codegen for destination atomic l-value with padding and junk in this padding bytes. llvm-svn: 237422
2015-05-15 16:36:34 +08:00
// CHECK: phi float
// CHECK: [[UP_INT:%.+]] = load i32
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR1_REF_INT:%.+]] = bitcast float* [[T_VAR1_REF]] to i32*
// CHECK: [[RES:%.+]] = cmpxchg i32* [[T_VAR1_REF_INT]], i32 [[ORIG_OLD_INT]], i32 [[UP_INT]] monotonic monotonic
[OPENMP] Fixed bug in atomic update/capture/write constructs. Fixed a bug with codegen for destination atomic l-value with padding and junk in this padding bytes. llvm-svn: 237422
2015-05-15 16:36:34 +08:00
// CHECK: [[OLD2:%.+]] = extractvalue { i32, i1 } [[RES]], 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[SUCCESS_FAIL:%.+]] = extractvalue { i32, i1 } [[RES]], 1
// CHECK: br i1 [[SUCCESS_FAIL]], label %[[ATOMIC_DONE:.+]], label %[[CONT]]
// CHECK: [[ATOMIC_DONE]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// __kmpc_end_reduce(<loc>, <gtid>, &<lock>);
// CHECK: call void @__kmpc_end_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]])
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// break;
// CHECK: br label %[[RED_DONE]]
// CHECK: [[RED_DONE]]
// CHECK-DAG: call {{.*}} [[S_FLOAT_TY_DESTR]]([[S_FLOAT_TY]]* [[VAR_PRIV]])
// CHECK-DAG: call {{.*}} [[S_FLOAT_TY_DESTR]]([[S_FLOAT_TY]]*
[OPENMP] Emit __kmpc_cancel_barrier() and code for 'cancellation point' only if 'cancel' is found. Patch improves codegen for OpenMP constructs. If the OpenMP region does not have internal 'cancel' construct, a call to 'void __kmpc_barrier()' runtime function is generated for all implicit/explicit barriers. If the region has inner 'cancel' directive, then ``` if (__kmpc_cancel_barrier()) exit from outer construct; ``` code is generated. Also, the code for 'canellation point' directive is not generated if parent directive does not have 'cancel' directive. llvm-svn: 247681
2015-09-15 20:52:43 +08:00
// CHECK: call void @__kmpc_barrier(%{{.+}}* [[IMPLICIT_BARRIER_LOC]], i{{[0-9]+}} [[GTID]])
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: ret void
// void reduce_func(void *lhs[<n>], void *rhs[<n>]) {
// *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]);
// ...
// *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1],
// *(Type<n>-1*)rhs[<n>-1]);
// }
// CHECK: define internal void [[REDUCTION_FUNC]](i8*, i8*)
// t_var_lhs = (float*)lhs[0];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS:%.+]], i64 0, i64 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR_RHS_VOID:%.+]] = load i8*, i8** [[T_VAR_RHS_REF]],
// CHECK: [[T_VAR_RHS:%.+]] = bitcast i8* [[T_VAR_RHS_VOID]] to float*
// t_var_rhs = (float*)rhs[0];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS:%.+]], i64 0, i64 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR_LHS_VOID:%.+]] = load i8*, i8** [[T_VAR_LHS_REF]],
// CHECK: [[T_VAR_LHS:%.+]] = bitcast i8* [[T_VAR_LHS_VOID]] to float*
// var_lhs = (S<float>*)lhs[1];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 1
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[VAR_RHS_VOID:%.+]] = load i8*, i8** [[VAR_RHS_REF]],
// CHECK: [[VAR_RHS:%.+]] = bitcast i8* [[VAR_RHS_VOID]] to [[S_FLOAT_TY]]*
// var_rhs = (S<float>*)rhs[1];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 1
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[VAR_LHS_VOID:%.+]] = load i8*, i8** [[VAR_LHS_REF]],
// CHECK: [[VAR_LHS:%.+]] = bitcast i8* [[VAR_LHS_VOID]] to [[S_FLOAT_TY]]*
// var1_lhs = (S<float>*)lhs[2];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR1_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 2
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[VAR1_RHS_VOID:%.+]] = load i8*, i8** [[VAR1_RHS_REF]],
// CHECK: [[VAR1_RHS:%.+]] = bitcast i8* [[VAR1_RHS_VOID]] to [[S_FLOAT_TY]]*
// var1_rhs = (S<float>*)rhs[2];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR1_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 2
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[VAR1_LHS_VOID:%.+]] = load i8*, i8** [[VAR1_LHS_REF]],
// CHECK: [[VAR1_LHS:%.+]] = bitcast i8* [[VAR1_LHS_VOID]] to [[S_FLOAT_TY]]*
// t_var1_lhs = (float*)lhs[3];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR1_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 3
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR1_RHS_VOID:%.+]] = load i8*, i8** [[T_VAR1_RHS_REF]],
// CHECK: [[T_VAR1_RHS:%.+]] = bitcast i8* [[T_VAR1_RHS_VOID]] to float*
// t_var1_rhs = (float*)rhs[3];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR1_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 3
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR1_LHS_VOID:%.+]] = load i8*, i8** [[T_VAR1_LHS_REF]],
// CHECK: [[T_VAR1_LHS:%.+]] = bitcast i8* [[T_VAR1_LHS_VOID]] to float*
// t_var_lhs += t_var_rhs;
// CHECK: [[T_VAR_LHS_VAL:%.+]] = load float, float* [[T_VAR_LHS]],
// CHECK: [[T_VAR_RHS_VAL:%.+]] = load float, float* [[T_VAR_RHS]],
// CHECK: [[UP:%.+]] = fadd float [[T_VAR_LHS_VAL]], [[T_VAR_RHS_VAL]]
// CHECK: store float [[UP]], float* [[T_VAR_LHS]],
// var_lhs = var_lhs.operator &(var_rhs);
// CHECK: [[UP:%.+]] = call dereferenceable(4) [[S_FLOAT_TY]]* @{{.+}}([[S_FLOAT_TY]]* [[VAR_LHS]], [[S_FLOAT_TY]]* dereferenceable(4) [[VAR_RHS]])
// CHECK: [[BC1:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_LHS]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_FLOAT_TY]]* [[UP]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// var1_lhs = var1_lhs.operator &&(var1_rhs);
// CHECK: [[TO_FLOAT:%.+]] = call float @{{.+}}([[S_FLOAT_TY]]* [[VAR1_LHS]])
// CHECK: [[VAR1_BOOL:%.+]] = fcmp une float [[TO_FLOAT]], 0.0
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[VAR1_BOOL]], label %[[TRUE:.+]], label %[[END2:.+]]
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[TRUE]]
// CHECK: [[TO_FLOAT:%.+]] = call float @{{.+}}([[S_FLOAT_TY]]* [[VAR1_RHS]])
// CHECK: [[VAR1_REDUCTION_BOOL:%.+]] = fcmp une float [[TO_FLOAT]], 0.0
// CHECK: br label %[[END2]]
// CHECK: [[END2]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: [[COND_LVALUE:%.+]] = phi i1 [ false, %{{.+}} ], [ [[VAR1_REDUCTION_BOOL]], %[[TRUE]] ]
// CHECK: [[CONV:%.+]] = uitofp i1 [[COND_LVALUE]] to float
// CHECK: call void @{{.+}}([[S_FLOAT_TY]]* [[COND_LVALUE:%.+]], float [[CONV]])
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BC1:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR1_LHS]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_FLOAT_TY]]* [[COND_LVALUE]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// t_var1_lhs = min(t_var1_lhs, t_var1_rhs);
// CHECK: [[T_VAR1_LHS_VAL:%.+]] = load float, float* [[T_VAR1_LHS]],
// CHECK: [[T_VAR1_RHS_VAL:%.+]] = load float, float* [[T_VAR1_RHS]],
// CHECK: [[CMP:%.+]] = fcmp olt float [[T_VAR1_LHS_VAL]], [[T_VAR1_RHS_VAL]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[CMP]]
// CHECK: [[UP:%.+]] = phi float
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: store float [[UP]], float* [[T_VAR1_LHS]],
// CHECK: ret void
[OPENMP 4.1] Codegen for array sections/subscripts in 'reduction' clause. OpenMP 4.1 adds support for array sections/subscripts in 'reduction' clause. Patch adds codegen for this feature. llvm-svn: 249672
2015-10-08 17:10:53 +08:00
// CHECK: define internal void [[MAIN_MICROTASK1]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, i64 %{{.+}}, i64 %{{.+}}, i32* nonnull %{{.+}}, [2 x i32]* dereferenceable(8) %{{.+}}, [10 x [4 x [[S_FLOAT_TY]]]]* dereferenceable(160) %{{.+}})
// Reduction list for runtime.
// CHECK: [[RED_LIST:%.+]] = alloca [4 x i8*],
// CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]],
// CHECK: [[IDX1:%.+]] = mul nsw i64 1, %{{.+}}
// CHECK: [[LB1:%.+]] = getelementptr inbounds i32, i32* %{{.+}}, i64 [[IDX1]]
// CHECK: [[LB1_0:%.+]] = getelementptr inbounds i32, i32* [[LB1]], i64 0
// CHECK: [[IDX1:%.+]] = mul nsw i64 1, %{{.+}}
// CHECK: [[UB1:%.+]] = getelementptr inbounds i32, i32* %{{.+}}, i64 [[IDX1]]
// CHECK: [[UB1_UP:%.+]] = getelementptr inbounds i32, i32* [[UB1]], i64 %
// CHECK: [[UB_CAST:%.+]] = ptrtoint i32* [[UB1_UP]] to i64
// CHECK: [[LB_CAST:%.+]] = ptrtoint i32* [[LB1_0]] to i64
// CHECK: [[DIFF:%.+]] = sub i64 [[UB_CAST]], [[LB_CAST]]
// CHECK: [[SIZE_1:%.+]] = sdiv exact i64 [[DIFF]], ptrtoint (i32* getelementptr (i32, i32* null, i32 1) to i64)
// CHECK: [[ARR_SIZE:%.+]] = add nuw i64 [[SIZE_1]], 1
// CHECK: call i8* @llvm.stacksave()
// CHECK: [[ARR_PRIV:%.+]] = alloca i32, i64 [[ARR_SIZE]],
// Check initialization of private copy.
// CHECK: [[END:%.+]] = getelementptr i32, i32* [[ARR_PRIV]], i64 [[ARR_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[ARR_PRIV]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi i32*
// CHECK: store i32 0, i32* %
// CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// CHECK: [[ARRS_PRIV:%.+]] = alloca [[S_FLOAT_TY]], i64 [[ARRS_SIZE:%.+]],
// Check initialization of private copy.
// CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_PRIV]], i64 [[ARRS_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_PRIV]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: call void @_ZN1SIfEC1Ev([[S_FLOAT_TY]]* %
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]]
// CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]]
// CHECK: call void @__kmpc_for_static_init_4(
// Skip checks for internal operations.
// CHECK: call void @__kmpc_for_static_fini(
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
// CHECK: [[ARR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 0
// CHECK: [[BITCAST:%.+]] = bitcast i32* [[ARR_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[ARR_PRIV_REF]],
// CHECK: [[ARR_SIZE_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 1
// CHECK: [[BITCAST:%.+]] = inttoptr i64 [[ARR_SIZE]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[ARR_SIZE_REF]],
// CHECK: [[ARRS_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 2
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[ARRS_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[ARRS_PRIV_REF]],
// CHECK: [[ARRS_SIZE_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 3
// CHECK: [[BITCAST:%.+]] = inttoptr i64 [[ARRS_SIZE]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[ARRS_SIZE_REF]],
// res = __kmpc_reduce(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>);
// CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]]
// CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]]
// CHECK: [[BITCAST:%.+]] = bitcast [4 x i8*]* [[RED_LIST]] to i8*
// CHECK: [[RES:%.+]] = call i32 @__kmpc_reduce_nowait(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], i32 2, i64 32, i8* [[BITCAST]], void (i8*, i8*)* [[REDUCTION_FUNC:@.+]], [8 x i32]* [[REDUCTION_LOCK]])
// switch(res)
// CHECK: switch i32 [[RES]], label %[[RED_DONE:.+]] [
// CHECK: i32 1, label %[[CASE1:.+]]
// CHECK: i32 2, label %[[CASE2:.+]]
// CHECK: ]
// case 1:
// CHECK: [[CASE1]]
// arr[:] += arr_reduction[:];
// CHECK: [[END:%.+]] = getelementptr i32, i32* [[LB1_0]], i64 [[ARR_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[LB1_0]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi i32*
// CHECK: [[ADD:%.+]] = add nsw i32 %
// CHECK: store i32 [[ADD]], i32* %
// CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// arrs[:] = var.operator &(arrs_reduction[:]);
// CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_LB:%.+]], i64 [[ARRS_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_LB]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: [[AND:%.+]] = call dereferenceable(4) [[S_FLOAT_TY]]* @_ZN1SIfEanERKS0_([[S_FLOAT_TY]]* %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(4) %{{.+}})
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[AND]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* %{{.+}}, i8* [[BITCAST]], i64 4, i32 4, i1 false)
[OPENMP 4.1] Codegen for array sections/subscripts in 'reduction' clause. OpenMP 4.1 adds support for array sections/subscripts in 'reduction' clause. Patch adds codegen for this feature. llvm-svn: 249672
2015-10-08 17:10:53 +08:00
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// __kmpc_end_reduce(<loc>, <gtid>, &<lock>);
// CHECK: call void @__kmpc_end_reduce_nowait(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]])
// break;
// CHECK: br label %[[RED_DONE]]
// case 2:
// CHECK: [[CASE2]]
// arr[:] += arr_reduction[:];
// CHECK: [[END:%.+]] = getelementptr i32, i32* [[LB1_0]], i64 [[ARR_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[LB1_0]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi i32*
// CHECK: atomicrmw add i32* %{{.+}}, i32 %{{.+}} monotonic
// CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// arrs[:] = var.operator &(arrs_reduction[:]);
// CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_LB:%.+]], i64 [[ARRS_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_LB]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: call void @__kmpc_critical(
// CHECK: [[AND:%.+]] = call dereferenceable(4) [[S_FLOAT_TY]]* @_ZN1SIfEanERKS0_([[S_FLOAT_TY]]* %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(4) %{{.+}})
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[AND]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* %{{.+}}, i8* [[BITCAST]], i64 4, i32 4, i1 false)
[OPENMP 4.1] Codegen for array sections/subscripts in 'reduction' clause. OpenMP 4.1 adds support for array sections/subscripts in 'reduction' clause. Patch adds codegen for this feature. llvm-svn: 249672
2015-10-08 17:10:53 +08:00
// CHECK: call void @__kmpc_end_critical(
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// break;
// CHECK: br label %[[RED_DONE]]
// CHECK: [[RED_DONE]]
// Check destruction of private copy.
// CHECK: [[END:%.+]] = getelementptr inbounds [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_PRIV]], i64 [[ARRS_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_PRIV]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: call void @_ZN1SIfED1Ev([[S_FLOAT_TY]]* %
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[ARRS_PRIV]]
// CHECK: br i1 [[DONE]],
// CHECK: call void @llvm.stackrestore(i8*
// CHECK: ret void
// void reduce_func(void *lhs[<n>], void *rhs[<n>]) {
// *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]);
// ...
// *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1],
// *(Type<n>-1*)rhs[<n>-1]);
// }
// CHECK: define internal void [[REDUCTION_FUNC]](i8*, i8*)
// arr_rhs = (int*)rhs[0];
// CHECK: [[ARR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS:%.+]], i64 0, i64 0
// CHECK: [[ARR_RHS_VOID:%.+]] = load i8*, i8** [[ARR_RHS_REF]],
// CHECK: [[ARR_RHS:%.+]] = bitcast i8* [[ARR_RHS_VOID]] to i32*
// arr_lhs = (int*)lhs[0];
// CHECK: [[ARR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS:%.+]], i64 0, i64 0
// CHECK: [[ARR_LHS_VOID:%.+]] = load i8*, i8** [[ARR_LHS_REF]],
// CHECK: [[ARR_LHS:%.+]] = bitcast i8* [[ARR_LHS_VOID]] to i32*
// arr_size = (size_t)lhs[1];
// CHECK: [[ARR_SIZE_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 1
// CHECK: [[ARR_SIZE_VOID:%.+]] = load i8*, i8** [[ARR_SIZE_REF]],
// CHECK: [[ARR_SIZE:%.+]] = ptrtoint i8* [[ARR_SIZE_VOID]] to i64
// arrs_rhs = (S<float>*)rhs[2];
// CHECK: [[ARRS_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 2
// CHECK: [[ARRS_RHS_VOID:%.+]] = load i8*, i8** [[ARRS_RHS_REF]],
// CHECK: [[ARRS_RHS:%.+]] = bitcast i8* [[ARRS_RHS_VOID]] to [[S_FLOAT_TY]]*
// arrs_lhs = (S<float>*)lhs[2];
// CHECK: [[ARRS_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 2
// CHECK: [[ARRS_LHS_VOID:%.+]] = load i8*, i8** [[ARRS_LHS_REF]],
// CHECK: [[ARRS_LHS:%.+]] = bitcast i8* [[ARRS_LHS_VOID]] to [[S_FLOAT_TY]]*
// arrs_size = (size_t)lhs[3];
// CHECK: [[ARRS_SIZE_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 3
// CHECK: [[ARRS_SIZE_VOID:%.+]] = load i8*, i8** [[ARRS_SIZE_REF]],
// CHECK: [[ARRS_SIZE:%.+]] = ptrtoint i8* [[ARRS_SIZE_VOID]] to i64
// arr_lhs[:] += arr_rhs[:];
// CHECK: [[END:%.+]] = getelementptr i32, i32* [[ARR_LHS]], i64 [[ARR_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[ARR_LHS]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi i32*
// CHECK: [[ADD:%.+]] = add nsw i32 %
// CHECK: store i32 [[ADD]], i32* %
// CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// arrs_lhs = arrs_lhs.operator &(arrs_rhs);
// CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_LB:%.+]], i64 [[ARRS_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_LB]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: [[AND:%.+]] = call dereferenceable(4) [[S_FLOAT_TY]]* @_ZN1SIfEanERKS0_([[S_FLOAT_TY]]* %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(4) %{{.+}})
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[AND]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* %{{.+}}, i8* [[BITCAST]], i64 4, i32 4, i1 false)
[OPENMP 4.1] Codegen for array sections/subscripts in 'reduction' clause. OpenMP 4.1 adds support for array sections/subscripts in 'reduction' clause. Patch adds codegen for this feature. llvm-svn: 249672
2015-10-08 17:10:53 +08:00
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// CHECK: ret void
[OPENMP 4.5] Allow arrays in 'reduction' clause. OpenMP 4.5, alogn with array sections, allows to use variables of array type in reductions. llvm-svn: 258804
2016-01-26 20:20:39 +08:00
// CHECK: define internal void [[MAIN_MICROTASK2]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, i64 %{{.+}}, i64 %{{.+}}, i32* nonnull %{{.+}}, [10 x [4 x [[S_FLOAT_TY]]]]* dereferenceable(160) %{{.+}})
// CHECK: [[ARRS_PRIV:%.+]] = alloca [10 x [4 x [[S_FLOAT_TY]]]],
// Reduction list for runtime.
// CHECK: [[RED_LIST:%.+]] = alloca [3 x i8*],
// CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]],
// CHECK: [[ARR_SIZE:%.+]] = mul nuw i64 %{{.+}}, 4
// CHECK: call i8* @llvm.stacksave()
// CHECK: [[ARR_PRIV:%.+]] = alloca i32, i64 [[ARR_SIZE]],
// Check initialization of private copy.
// CHECK: [[END:%.+]] = getelementptr i32, i32* [[ARR_PRIV]], i64 [[ARR_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[ARR_PRIV]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi i32*
// CHECK: store i32 0, i32* %
// CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// Check initialization of private copy.
// CHECK: [[LHS_BEGIN:%.+]] = bitcast [10 x [4 x [[S_FLOAT_TY]]]]* %{{.+}} to [[S_FLOAT_TY]]*
// CHECK: [[BEGIN:%.+]] = getelementptr inbounds [10 x [4 x [[S_FLOAT_TY]]]], [10 x [4 x [[S_FLOAT_TY]]]]* [[ARRS_PRIV]], i32 0, i32 0, i32 0
// CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[BEGIN]], i64 40
// CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[BEGIN]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: call void @_ZN1SIfEC1Ev([[S_FLOAT_TY]]* %
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// CHECK: [[ARRS_PRIV_BEGIN:%.+]] = bitcast [10 x [4 x [[S_FLOAT_TY]]]]* [[ARRS_PRIV]] to [[S_FLOAT_TY]]*
// CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]]
// CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]]
// CHECK: call void @__kmpc_for_static_init_4(
// Skip checks for internal operations.
// CHECK: call void @__kmpc_for_static_fini(
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
// CHECK: [[ARR_PRIV_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST]], i64 0, i64 0
// CHECK: [[BITCAST:%.+]] = bitcast i32* [[ARR_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[ARR_PRIV_REF]],
// CHECK: [[ARR_SIZE_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST]], i64 0, i64 1
// CHECK: [[BITCAST:%.+]] = inttoptr i64 [[ARR_SIZE]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[ARR_SIZE_REF]],
// CHECK: [[ARRS_PRIV_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST]], i64 0, i64 2
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[ARRS_PRIV_BEGIN]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[ARRS_PRIV_REF]],
// res = __kmpc_reduce(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>);
// CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]]
// CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]]
// CHECK: [[BITCAST:%.+]] = bitcast [3 x i8*]* [[RED_LIST]] to i8*
// CHECK: [[RES:%.+]] = call i32 @__kmpc_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], i32 2, i64 24, i8* [[BITCAST]], void (i8*, i8*)* [[REDUCTION_FUNC:@.+]], [8 x i32]* [[REDUCTION_LOCK]])
// switch(res)
// CHECK: switch i32 [[RES]], label %[[RED_DONE:.+]] [
// CHECK: i32 1, label %[[CASE1:.+]]
// CHECK: i32 2, label %[[CASE2:.+]]
// CHECK: ]
// case 1:
// CHECK: [[CASE1]]
// arr[:] += arr_reduction[:];
// CHECK: [[END:%.+]] = getelementptr i32, i32* [[LB1_0:%.+]], i64 [[ARR_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[LB1_0]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi i32*
// CHECK: [[ADD:%.+]] = add nsw i32 %
// CHECK: store i32 [[ADD]], i32* %
// CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// arrs[:] = var.operator &(arrs_reduction[:]);
// CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[LHS_BEGIN]], i64 40
// CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[LHS_BEGIN]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: [[AND:%.+]] = call dereferenceable(4) [[S_FLOAT_TY]]* @_ZN1SIfEanERKS0_([[S_FLOAT_TY]]* %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(4) %{{.+}})
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[AND]] to i8*
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* %{{.+}}, i8* [[BITCAST]], i64 4, i32 4, i1 false)
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// __kmpc_end_reduce(<loc>, <gtid>, &<lock>);
// CHECK: call void @__kmpc_end_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]])
// break;
// CHECK: br label %[[RED_DONE]]
// case 2:
// CHECK: [[CASE2]]
// arr[:] += arr_reduction[:];
// CHECK: [[END:%.+]] = getelementptr i32, i32* [[LB1_0]], i64 [[ARR_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[LB1_0]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi i32*
// CHECK: atomicrmw add i32* %{{.+}}, i32 %{{.+}} monotonic
// CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// arrs[:] = var.operator &(arrs_reduction[:]);
// CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[LHS_BEGIN]], i64 40
// CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[LHS_BEGIN]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: call void @__kmpc_critical(
// CHECK: [[AND:%.+]] = call dereferenceable(4) [[S_FLOAT_TY]]* @_ZN1SIfEanERKS0_([[S_FLOAT_TY]]* %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(4) %{{.+}})
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[AND]] to i8*
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* %{{.+}}, i8* [[BITCAST]], i64 4, i32 4, i1 false)
// CHECK: call void @__kmpc_end_critical(
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// break;
// CHECK: br label %[[RED_DONE]]
// CHECK: [[RED_DONE]]
// Check destruction of private copy.
// CHECK: [[BEGIN:%.+]] = getelementptr inbounds [10 x [4 x [[S_FLOAT_TY]]]], [10 x [4 x [[S_FLOAT_TY]]]]* [[ARRS_PRIV]], i32 0, i32 0, i32 0
// CHECK: [[END:%.+]] = getelementptr inbounds [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[BEGIN]], i64 40
// CHECK: br
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: call void @_ZN1SIfED1Ev([[S_FLOAT_TY]]* %
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[BEGIN]]
// CHECK: br i1 [[DONE]],
// CHECK: call void @llvm.stackrestore(i8*
// CHECK: call void @__kmpc_barrier(
// CHECK: ret void
// void reduce_func(void *lhs[<n>], void *rhs[<n>]) {
// *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]);
// ...
// *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1],
// *(Type<n>-1*)rhs[<n>-1]);
// }
// CHECK: define internal void [[REDUCTION_FUNC]](i8*, i8*)
// arr_rhs = (int*)rhs[0];
// CHECK: [[ARR_RHS_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_RHS:%.+]], i64 0, i64 0
// CHECK: [[ARR_RHS_VOID:%.+]] = load i8*, i8** [[ARR_RHS_REF]],
// CHECK: [[ARR_RHS:%.+]] = bitcast i8* [[ARR_RHS_VOID]] to i32*
// arr_lhs = (int*)lhs[0];
// CHECK: [[ARR_LHS_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_LHS:%.+]], i64 0, i64 0
// CHECK: [[ARR_LHS_VOID:%.+]] = load i8*, i8** [[ARR_LHS_REF]],
// CHECK: [[ARR_LHS:%.+]] = bitcast i8* [[ARR_LHS_VOID]] to i32*
// arr_size = (size_t)lhs[1];
// CHECK: [[ARR_SIZE_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_LHS]], i64 0, i64 1
// CHECK: [[ARR_SIZE_VOID:%.+]] = load i8*, i8** [[ARR_SIZE_REF]],
// CHECK: [[ARR_SIZE:%.+]] = ptrtoint i8* [[ARR_SIZE_VOID]] to i64
// arrs_rhs = (S<float>*)rhs[2];
// CHECK: [[ARRS_RHS_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_RHS]], i64 0, i64 2
// CHECK: [[ARRS_RHS_VOID:%.+]] = load i8*, i8** [[ARRS_RHS_REF]],
// CHECK: [[ARRS_RHS:%.+]] = bitcast i8* [[ARRS_RHS_VOID]] to [[S_FLOAT_TY]]*
// arrs_lhs = (S<float>*)lhs[2];
// CHECK: [[ARRS_LHS_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_LHS]], i64 0, i64 2
// CHECK: [[ARRS_LHS_VOID:%.+]] = load i8*, i8** [[ARRS_LHS_REF]],
// CHECK: [[ARRS_LHS:%.+]] = bitcast i8* [[ARRS_LHS_VOID]] to [[S_FLOAT_TY]]*
// arr_lhs[:] += arr_rhs[:];
// CHECK: [[END:%.+]] = getelementptr i32, i32* [[ARR_LHS]], i64 [[ARR_SIZE]]
// CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[ARR_LHS]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi i32*
// CHECK: [[ADD:%.+]] = add nsw i32 %
// CHECK: store i32 [[ADD]], i32* %
// CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// arrs_lhs = arrs_lhs.operator &(arrs_rhs);
// CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_LB:%.+]], i64 40
// CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_LB]], [[END]]
// CHECK: br i1 [[ISEMPTY]],
// CHECK: phi [[S_FLOAT_TY]]*
// CHECK: [[AND:%.+]] = call dereferenceable(4) [[S_FLOAT_TY]]* @_ZN1SIfEanERKS0_([[S_FLOAT_TY]]* %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(4) %{{.+}})
// CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[AND]] to i8*
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* %{{.+}}, i8* [[BITCAST]], i64 4, i32 4, i1 false)
// CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]]
// CHECK: br i1 [[DONE]],
// CHECK: ret void
[OPENMP 4.0] Fixed support of array sections/array subscripts. Codegen for array sections/array subscripts worked only for expressions with arrays as base. Patch fixes codegen for bases with pointer/reference types. llvm-svn: 259776
2016-02-04 19:27:03 +08:00
// CHECK: define internal void [[MAIN_MICROTASK3]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [[S_FLOAT_TY]]*** dereferenceable(8) %{{.+}})
// CHECK: [[VAR2_ORIG_ADDR:%.+]] = alloca [[S_FLOAT_TY]]***,
// Reduction list for runtime.
// CHECK: [[RED_LIST:%.+]] = alloca [2 x i8*],
// CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]],
// CHECK: [[VAR2_ORIG:%.+]] = load [[S_FLOAT_TY]]***, [[S_FLOAT_TY]]**** [[VAR2_ORIG_ADDR]],
// CHECK: load [[S_FLOAT_TY]]**, [[S_FLOAT_TY]]*** [[VAR2_ORIG]],
// CHECK: getelementptr inbounds [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** %{{.+}}, i64 0
// CHECK: load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** %
// CHECK: [[LOW:%.+]] = getelementptr inbounds [[S_FLOAT_TY]], [[S_FLOAT_TY]]* %{{.+}}, i64 1
// CHECK: load [[S_FLOAT_TY]]**, [[S_FLOAT_TY]]*** [[VAR2_ORIG]],
// CHECK: getelementptr inbounds [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** %{{.+}}, i64 4
// CHECK: load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** %
// CHECK: getelementptr inbounds [[S_FLOAT_TY]], [[S_FLOAT_TY]]* %{{.+}}, i64 6
// CHECK: [[LD:%.+]] = load [[S_FLOAT_TY]]**, [[S_FLOAT_TY]]*** [[VAR2_ORIG]],
// CHECK: [[ORIG_START:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[LD]],
// CHECK: [[LAST:%.+]] = ptrtoint [[S_FLOAT_TY]]* %{{.+}} to i64
// CHECK: [[FIRST:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW]] to i64
// CHECK: [[BYTE_DIF:%.+]] = sub i64 [[LAST]], [[FIRST]]
// CHECK: [[DIF:%.+]] = sdiv exact i64 [[BYTE_DIF]], ptrtoint (float* getelementptr (float, float* null, i32 1) to i64)
// CHECK: [[SIZE:%.+]] = add nuw i64 [[DIF]], 1
// CHECK: call i8* @llvm.stacksave()
// CHECK: [[VAR2_PRIV:%.+]] = alloca [[S_FLOAT_TY]], i64 [[SIZE]],
// CHECK: [[START:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[ORIG_START]] to i64
// CHECK: [[LOW_BOUND:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW]] to i64
// CHECK: [[OFFSET_BYTES:%.+]] = sub i64 [[START]], [[LOW_BOUND]]
// CHECK: [[OFFSET:%.+]] = sdiv exact i64 [[OFFSET_BYTES]], ptrtoint (float* getelementptr (float, float* null, i32 1) to i64)
// CHECK: [[PSEUDO_VAR2_PRIV:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[VAR2_PRIV]], i64 [[OFFSET]]
// CHECK: store [[S_FLOAT_TY]]** [[REF:.+]], [[S_FLOAT_TY]]*** %
// CHECK: store [[S_FLOAT_TY]]* [[PSEUDO_VAR2_PRIV]], [[S_FLOAT_TY]]** [[REF]]
// CHECK: ret void
// CHECK: define internal void [[MAIN_MICROTASK4]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [2 x [[S_FLOAT_TY]]]* dereferenceable(8) %{{.+}})
// CHECK: [[VVAR2_ORIG_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*,
// Reduction list for runtime.
// CHECK: [[RED_LIST:%.+]] = alloca [2 x i8*],
// CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]],
// CHECK: [[VVAR2_ORIG:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[VVAR2_ORIG_ADDR]],
// CHECK: [[LOW:%.+]] = getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[VVAR2_ORIG]], i64 0, i64 0
// CHECK: getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[VVAR2_ORIG]], i64 0, i64 4
// CHECK: [[ORIG_START:%.+]] = bitcast [2 x [[S_FLOAT_TY]]]* [[VVAR2_ORIG]] to [[S_FLOAT_TY]]*
// CHECK: [[LAST:%.+]] = ptrtoint [[S_FLOAT_TY]]* %{{.+}} to i64
// CHECK: [[FIRST:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW]] to i64
// CHECK: [[BYTE_DIF:%.+]] = sub i64 [[LAST]], [[FIRST]]
// CHECK: [[DIF:%.+]] = sdiv exact i64 [[BYTE_DIF]], ptrtoint (float* getelementptr (float, float* null, i32 1) to i64)
// CHECK: [[SIZE:%.+]] = add nuw i64 [[DIF]], 1
// CHECK: call i8* @llvm.stacksave()
// CHECK: [[VVAR2_PRIV:%.+]] = alloca [[S_FLOAT_TY]], i64 [[SIZE]],
// CHECK: [[START:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[ORIG_START]] to i64
// CHECK: [[LOW_BOUND:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW]] to i64
// CHECK: [[OFFSET_BYTES:%.+]] = sub i64 [[START]], [[LOW_BOUND]]
// CHECK: [[OFFSET:%.+]] = sdiv exact i64 [[OFFSET_BYTES]], ptrtoint (float* getelementptr (float, float* null, i32 1) to i64)
// CHECK: [[PSEUDO_VVAR2_PRIV:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[VVAR2_PRIV]], i64 [[OFFSET]]
// CHECK: [[VVAR2_PRIV:%.+]] = bitcast [[S_FLOAT_TY]]* [[PSEUDO_VVAR2_PRIV]] to [2 x [[S_FLOAT_TY]]]*
// CHECK: ret void
// CHECK: define internal void [[MAIN_MICROTASK5]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [2 x [[S_FLOAT_TY]]]* dereferenceable(8) %{{.+}})
// CHECK: [[VAR3_ORIG_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*,
// Reduction list for runtime.
// CHECK: [[RED_LIST:%.+]] = alloca [2 x i8*],
// CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]],
// CHECK: [[VAR3_ORIG:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]],
// CHECK: [[LOW:%.+]] = getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[VAR3_ORIG]], i64 0, i64 1
// CHECK: [[VAR3_ORIG:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]],
// CHECK: getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[VAR3_ORIG]], i64 0, i64 2
// CHECK: [[VAR3_ORIG:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]],
// CHECK: [[ORIG_START:%.+]] = bitcast [2 x [[S_FLOAT_TY]]]* [[VAR3_ORIG]] to [[S_FLOAT_TY]]*
// CHECK: [[LAST:%.+]] = ptrtoint [[S_FLOAT_TY]]* %{{.+}} to i64
// CHECK: [[FIRST:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW]] to i64
// CHECK: [[BYTE_DIF:%.+]] = sub i64 [[LAST]], [[FIRST]]
// CHECK: [[DIF:%.+]] = sdiv exact i64 [[BYTE_DIF]], ptrtoint (float* getelementptr (float, float* null, i32 1) to i64)
// CHECK: [[SIZE:%.+]] = add nuw i64 [[DIF]], 1
// CHECK: call i8* @llvm.stacksave()
// CHECK: [[VAR3_PRIV:%.+]] = alloca [[S_FLOAT_TY]], i64 [[SIZE]],
// CHECK: [[START:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[ORIG_START]] to i64
// CHECK: [[LOW_BOUND:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW]] to i64
// CHECK: [[OFFSET_BYTES:%.+]] = sub i64 [[START]], [[LOW_BOUND]]
// CHECK: [[OFFSET:%.+]] = sdiv exact i64 [[OFFSET_BYTES]], ptrtoint (float* getelementptr (float, float* null, i32 1) to i64)
// CHECK: [[PSEUDO_VAR3_PRIV:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[VAR3_PRIV]], i64 [[OFFSET]]
// CHECK: [[VAR3_PRIV:%.+]] = bitcast [[S_FLOAT_TY]]* [[PSEUDO_VAR3_PRIV]] to [2 x [[S_FLOAT_TY]]]*
// CHECK: store [2 x [[S_FLOAT_TY]]]* [[VAR3_PRIV]], [2 x [[S_FLOAT_TY]]]** %
// CHECK: ret void
// CHECK: define internal void [[MAIN_MICROTASK6]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [2 x [[S_FLOAT_TY]]]* dereferenceable(8) %{{.+}})
// CHECK: [[VAR3_ORIG_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*,
// CHECK: [[VAR3_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]],
// Reduction list for runtime.
// CHECK: [[RED_LIST:%.+]] = alloca [1 x i8*],
// CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]],
// CHECK: [[VAR3_ORIG:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]],
// CHECK: bitcast [2 x [[S_FLOAT_TY]]]* [[VAR3_ORIG]] to [[S_FLOAT_TY]]*
// CHECK: getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[VAR3_PRIV]], i32 0, i32 0
// CHECK: getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* %{{.+}}, i64 2
// CHECK: store [2 x [[S_FLOAT_TY]]]* [[VAR3_PRIV]], [2 x [[S_FLOAT_TY]]]** %
// CHECK: ret void
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: define {{.*}} i{{[0-9]+}} [[TMAIN_INT]]()
// CHECK: [[TEST:%.+]] = alloca [[S_INT_TY]],
// CHECK: call {{.*}} [[S_INT_TY_CONSTR:@.+]]([[S_INT_TY]]* [[TEST]])
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 6, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, i32*, [[S_INT_TY]]*, [[S_INT_TY]]*, i32*, [2 x i32]*, [2 x [[S_INT_TY]]]*)* [[TMAIN_MICROTASK:@.+]] to void
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: call {{.*}} [[S_INT_TY_DESTR:@.+]]([[S_INT_TY]]*
// CHECK: ret
//
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// CHECK: define internal void [[TMAIN_MICROTASK]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, i32* dereferenceable(4) %{{.+}}, [[S_INT_TY]]* dereferenceable(4) %{{.+}}, [[S_INT_TY]]* dereferenceable(4) %{{.+}}, i32* dereferenceable(4) %{{.+}}, [2 x i32]* dereferenceable(8) %{{.+}}, [2 x [[S_INT_TY]]]* dereferenceable(8) %{{.+}})
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: alloca i{{[0-9]+}},
// CHECK: alloca i{{[0-9]+}},
// CHECK: alloca i{{[0-9]+}},
// CHECK: alloca i{{[0-9]+}},
// CHECK: alloca i{{[0-9]+}},
// CHECK: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}},
// CHECK: [[VAR_PRIV:%.+]] = alloca [[S_INT_TY]],
// CHECK: [[VAR1_PRIV:%.+]] = alloca [[S_INT_TY]],
// CHECK: [[T_VAR1_PRIV:%.+]] = alloca i{{[0-9]+}},
// Reduction list for runtime.
// CHECK: [[RED_LIST:%.+]] = alloca [4 x i8*],
// CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]],
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// CHECK: [[T_VAR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** %
// CHECK: [[VAR1_REF:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** %
// CHECK: [[T_VAR1_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** %
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// For + reduction operation initial value of private variable is 0.
// CHECK: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[T_VAR_PRIV]],
// For & reduction operation initial value of private variable is ones in all bits.
[OPENMP] Outlined function for parallel and other regions with list of captured variables. Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least. Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record. llvm-svn: 247251
2015-09-10 16:12:02 +08:00
// CHECK: [[VAR_REF:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** %
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: call {{.*}} [[S_INT_TY_CONSTR:@.+]]([[S_INT_TY]]* [[VAR_PRIV]])
// For && reduction operation initial value of private variable is 1.0.
// CHECK: call {{.*}} [[S_INT_TY_CONSTR:@.+]]([[S_INT_TY]]* [[VAR1_PRIV]])
// For min reduction operation initial value of private variable is largest repesentable value.
// CHECK: store i{{[0-9]+}} 2147483647, i{{[0-9]+}}* [[T_VAR1_PRIV]],
// CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]]
// CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]]
// CHECK: call void @__kmpc_for_static_init_4(
// Skip checks for internal operations.
// CHECK: call void @__kmpc_for_static_fini(
// void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]};
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BITCAST:%.+]] = bitcast i{{[0-9]+}}* [[T_VAR_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[T_VAR_PRIV_REF]],
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 1
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BITCAST:%.+]] = bitcast [[S_INT_TY]]* [[VAR_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[VAR_PRIV_REF]],
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR1_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 2
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BITCAST:%.+]] = bitcast [[S_INT_TY]]* [[VAR1_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[VAR1_PRIV_REF]],
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR1_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 3
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BITCAST:%.+]] = bitcast i{{[0-9]+}}* [[T_VAR1_PRIV]] to i8*
// CHECK: store i8* [[BITCAST]], i8** [[T_VAR1_PRIV_REF]],
// res = __kmpc_reduce_nowait(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>);
// CHECK: [[BITCAST:%.+]] = bitcast [4 x i8*]* [[RED_LIST]] to i8*
// CHECK: [[RES:%.+]] = call i32 @__kmpc_reduce_nowait(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], i32 4, i64 32, i8* [[BITCAST]], void (i8*, i8*)* [[REDUCTION_FUNC:@.+]], [8 x i32]* [[REDUCTION_LOCK]])
// switch(res)
// CHECK: switch i32 [[RES]], label %[[RED_DONE:.+]] [
// CHECK: i32 1, label %[[CASE1:.+]]
// CHECK: i32 2, label %[[CASE2:.+]]
// CHECK: ]
// case 1:
// t_var += t_var_reduction;
// CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_REF]],
// CHECK: [[T_VAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]],
// CHECK: [[UP:%.+]] = add nsw i{{[0-9]+}} [[T_VAR_VAL]], [[T_VAR_PRIV_VAL]]
// CHECK: store i{{[0-9]+}} [[UP]], i{{[0-9]+}}* [[T_VAR_REF]],
// var = var.operator &(var_reduction);
// CHECK: [[UP:%.+]] = call dereferenceable(4) [[S_INT_TY]]* @{{.+}}([[S_INT_TY]]* [[VAR_REF]], [[S_INT_TY]]* dereferenceable(4) [[VAR_PRIV]])
// CHECK: [[BC1:%.+]] = bitcast [[S_INT_TY]]* [[VAR_REF]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_INT_TY]]* [[UP]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// var1 = var1.operator &&(var1_reduction);
// CHECK: [[TO_INT:%.+]] = call i{{[0-9]+}} @{{.+}}([[S_INT_TY]]* [[VAR1_REF]])
// CHECK: [[VAR1_BOOL:%.+]] = icmp ne i{{[0-9]+}} [[TO_INT]], 0
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[VAR1_BOOL]], label %[[TRUE:.+]], label %[[END2:.+]]
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[TRUE]]
// CHECK: [[TO_INT:%.+]] = call i{{[0-9]+}} @{{.+}}([[S_INT_TY]]* [[VAR1_PRIV]])
// CHECK: [[VAR1_REDUCTION_BOOL:%.+]] = icmp ne i{{[0-9]+}} [[TO_INT]], 0
// CHECK: br label %[[END2]]
// CHECK: [[END2]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: [[COND_LVALUE:%.+]] = phi i1 [ false, %{{.+}} ], [ [[VAR1_REDUCTION_BOOL]], %[[TRUE]] ]
// CHECK: [[CONV:%.+]] = zext i1 [[COND_LVALUE]] to i32
// CHECK: call void @{{.+}}([[S_INT_TY]]* [[COND_LVALUE:%.+]], i32 [[CONV]])
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BC1:%.+]] = bitcast [[S_INT_TY]]* [[VAR1_REF]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_INT_TY]]* [[COND_LVALUE]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// t_var1 = min(t_var1, t_var1_reduction);
// CHECK: [[T_VAR1_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR1_REF]],
// CHECK: [[T_VAR1_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR1_PRIV]],
// CHECK: [[CMP:%.+]] = icmp slt i{{[0-9]+}} [[T_VAR1_VAL]], [[T_VAR1_PRIV_VAL]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[CMP]]
// CHECK: [[UP:%.+]] = phi i32
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: store i{{[0-9]+}} [[UP]], i{{[0-9]+}}* [[T_VAR1_REF]],
// __kmpc_end_reduce_nowait(<loc>, <gtid>, &<lock>);
// CHECK: call void @__kmpc_end_reduce_nowait(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]])
// break;
// CHECK: br label %[[RED_DONE]]
// case 2:
// t_var += t_var_reduction;
// CHECK: [[T_VAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]]
// CHECK: atomicrmw add i32* [[T_VAR_REF]], i32 [[T_VAR_PRIV_VAL]] monotonic
// var = var.operator &(var_reduction);
// CHECK: call void @__kmpc_critical(
// CHECK: [[UP:%.+]] = call dereferenceable(4) [[S_INT_TY]]* @{{.+}}([[S_INT_TY]]* [[VAR_REF]], [[S_INT_TY]]* dereferenceable(4) [[VAR_PRIV]])
// CHECK: [[BC1:%.+]] = bitcast [[S_INT_TY]]* [[VAR_REF]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_INT_TY]]* [[UP]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: call void @__kmpc_end_critical(
// var1 = var1.operator &&(var1_reduction);
// CHECK: call void @__kmpc_critical(
// CHECK: [[TO_INT:%.+]] = call i{{[0-9]+}} @{{.+}}([[S_INT_TY]]* [[VAR1_REF]])
// CHECK: [[VAR1_BOOL:%.+]] = icmp ne i{{[0-9]+}} [[TO_INT]], 0
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[VAR1_BOOL]], label %[[TRUE:.+]], label %[[END2:.+]]
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[TRUE]]
// CHECK: [[TO_INT:%.+]] = call i{{[0-9]+}} @{{.+}}([[S_INT_TY]]* [[VAR1_PRIV]])
// CHECK: [[VAR1_REDUCTION_BOOL:%.+]] = icmp ne i{{[0-9]+}} [[TO_INT]], 0
// CHECK: br label %[[END2]]
// CHECK: [[END2]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: [[COND_LVALUE:%.+]] = phi i1 [ false, %{{.+}} ], [ [[VAR1_REDUCTION_BOOL]], %[[TRUE]] ]
// CHECK: [[CONV:%.+]] = zext i1 [[COND_LVALUE]] to i32
// CHECK: call void @{{.+}}([[S_INT_TY]]* [[COND_LVALUE:%.+]], i32 [[CONV]])
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BC1:%.+]] = bitcast [[S_INT_TY]]* [[VAR1_REF]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_INT_TY]]* [[COND_LVALUE]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: call void @__kmpc_end_critical(
// t_var1 = min(t_var1, t_var1_reduction);
// CHECK: [[T_VAR1_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR1_PRIV]]
// CHECK: atomicrmw min i32* [[T_VAR1_REF]], i32 [[T_VAR1_PRIV_VAL]] monotonic
// break;
// CHECK: br label %[[RED_DONE]]
// CHECK: [[RED_DONE]]
// CHECK-DAG: call {{.*}} [[S_INT_TY_DESTR]]([[S_INT_TY]]* [[VAR_PRIV]])
// CHECK-DAG: call {{.*}} [[S_INT_TY_DESTR]]([[S_INT_TY]]*
// CHECK: ret void
// void reduce_func(void *lhs[<n>], void *rhs[<n>]) {
// *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]);
// ...
// *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1],
// *(Type<n>-1*)rhs[<n>-1]);
// }
// CHECK: define internal void [[REDUCTION_FUNC]](i8*, i8*)
// t_var_lhs = (i{{[0-9]+}}*)lhs[0];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS:%.+]], i64 0, i64 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR_RHS_VOID:%.+]] = load i8*, i8** [[T_VAR_RHS_REF]],
// CHECK: [[T_VAR_RHS:%.+]] = bitcast i8* [[T_VAR_RHS_VOID]] to i{{[0-9]+}}*
// t_var_rhs = (i{{[0-9]+}}*)rhs[0];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS:%.+]], i64 0, i64 0
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR_LHS_VOID:%.+]] = load i8*, i8** [[T_VAR_LHS_REF]],
// CHECK: [[T_VAR_LHS:%.+]] = bitcast i8* [[T_VAR_LHS_VOID]] to i{{[0-9]+}}*
// var_lhs = (S<i{{[0-9]+}}>*)lhs[1];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 1
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[VAR_RHS_VOID:%.+]] = load i8*, i8** [[VAR_RHS_REF]],
// CHECK: [[VAR_RHS:%.+]] = bitcast i8* [[VAR_RHS_VOID]] to [[S_INT_TY]]*
// var_rhs = (S<i{{[0-9]+}}>*)rhs[1];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 1
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[VAR_LHS_VOID:%.+]] = load i8*, i8** [[VAR_LHS_REF]],
// CHECK: [[VAR_LHS:%.+]] = bitcast i8* [[VAR_LHS_VOID]] to [[S_INT_TY]]*
// var1_lhs = (S<i{{[0-9]+}}>*)lhs[2];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR1_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 2
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[VAR1_RHS_VOID:%.+]] = load i8*, i8** [[VAR1_RHS_REF]],
// CHECK: [[VAR1_RHS:%.+]] = bitcast i8* [[VAR1_RHS_VOID]] to [[S_INT_TY]]*
// var1_rhs = (S<i{{[0-9]+}}>*)rhs[2];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR1_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 2
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[VAR1_LHS_VOID:%.+]] = load i8*, i8** [[VAR1_LHS_REF]],
// CHECK: [[VAR1_LHS:%.+]] = bitcast i8* [[VAR1_LHS_VOID]] to [[S_INT_TY]]*
// t_var1_lhs = (i{{[0-9]+}}*)lhs[3];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR1_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 3
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR1_RHS_VOID:%.+]] = load i8*, i8** [[T_VAR1_RHS_REF]],
// CHECK: [[T_VAR1_RHS:%.+]] = bitcast i8* [[T_VAR1_RHS_VOID]] to i{{[0-9]+}}*
// t_var1_rhs = (i{{[0-9]+}}*)rhs[3];
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[T_VAR1_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 3
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[T_VAR1_LHS_VOID:%.+]] = load i8*, i8** [[T_VAR1_LHS_REF]],
// CHECK: [[T_VAR1_LHS:%.+]] = bitcast i8* [[T_VAR1_LHS_VOID]] to i{{[0-9]+}}*
// t_var_lhs += t_var_rhs;
// CHECK: [[T_VAR_LHS_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_LHS]],
// CHECK: [[T_VAR_RHS_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_RHS]],
// CHECK: [[UP:%.+]] = add nsw i{{[0-9]+}} [[T_VAR_LHS_VAL]], [[T_VAR_RHS_VAL]]
// CHECK: store i{{[0-9]+}} [[UP]], i{{[0-9]+}}* [[T_VAR_LHS]],
// var_lhs = var_lhs.operator &(var_rhs);
// CHECK: [[UP:%.+]] = call dereferenceable(4) [[S_INT_TY]]* @{{.+}}([[S_INT_TY]]* [[VAR_LHS]], [[S_INT_TY]]* dereferenceable(4) [[VAR_RHS]])
// CHECK: [[BC1:%.+]] = bitcast [[S_INT_TY]]* [[VAR_LHS]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_INT_TY]]* [[UP]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// var1_lhs = var1_lhs.operator &&(var1_rhs);
// CHECK: [[TO_INT:%.+]] = call i{{[0-9]+}} @{{.+}}([[S_INT_TY]]* [[VAR1_LHS]])
// CHECK: [[VAR1_BOOL:%.+]] = icmp ne i{{[0-9]+}} [[TO_INT]], 0
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[VAR1_BOOL]], label %[[TRUE:.+]], label %[[END2:.+]]
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[TRUE]]
// CHECK: [[TO_INT:%.+]] = call i{{[0-9]+}} @{{.+}}([[S_INT_TY]]* [[VAR1_RHS]])
// CHECK: [[VAR1_REDUCTION_BOOL:%.+]] = icmp ne i{{[0-9]+}} [[TO_INT]], 0
// CHECK: br label %[[END2]]
// CHECK: [[END2]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: [[COND_LVALUE:%.+]] = phi i1 [ false, %{{.+}} ], [ [[VAR1_REDUCTION_BOOL]], %[[TRUE]] ]
// CHECK: [[CONV:%.+]] = zext i1 [[COND_LVALUE]] to i32
// CHECK: call void @{{.+}}([[S_INT_TY]]* [[COND_LVALUE:%.+]], i32 [[CONV]])
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: [[BC1:%.+]] = bitcast [[S_INT_TY]]* [[VAR1_LHS]] to i8*
// CHECK: [[BC2:%.+]] = bitcast [[S_INT_TY]]* [[COND_LVALUE]] to i8*
Revert "Change memcpy/memset/memmove to have dest and source alignments." This reverts commit r253512. This likely broke the bots in: http://lab.llvm.org:8011/builders/clang-ppc64-elf-linux2/builds/20202 http://bb.pgr.jp/builders/clang-3stage-i686-linux/builds/3787 llvm-svn: 253542
2015-11-19 13:55:59 +08:00
// CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[BC1]], i8* [[BC2]], i64 4, i32 4, i1 false)
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// t_var1_lhs = min(t_var1_lhs, t_var1_rhs);
// CHECK: [[T_VAR1_LHS_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR1_LHS]],
// CHECK: [[T_VAR1_RHS_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR1_RHS]],
// CHECK: [[CMP:%.+]] = icmp slt i{{[0-9]+}} [[T_VAR1_LHS_VAL]], [[T_VAR1_RHS_VAL]]
[OPENMP] Fixed codegen for 'reduction' clause. Fixed codegen for reduction operations min, max, && and ||. Codegen for them is quite similar and I was confused by this similarity. Also added a call to kmpc_end_reduce() in atomic part of reduction codegen (call to kmpc_end_reduce_nowait() is not required). Differential Revision: http://reviews.llvm.org/D9513 llvm-svn: 236689
2015-05-07 11:54:03 +08:00
// CHECK: br i1 [[CMP]]
// CHECK: [[UP:%.+]] = phi i32
[OPENMP] Codegen for 'reduction' clause in 'for' directive. Emit a code for reduction clause. Next code should be emitted for reductions: static kmp_critical_name lock = { 0 }; void reduce_func(void *lhs[<n>], void *rhs[<n>]) { *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); ... *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], *(Type<n>-1*)rhs[<n>-1]); } ... void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) { case 1: <LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]); ... <LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]); __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); break; case 2: Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0])); ... Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1])); break; default:; } Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation. Differential Revision: http://reviews.llvm.org/D9139 llvm-svn: 235506
2015-04-22 21:43:03 +08:00
// CHECK: store i{{[0-9]+}} [[UP]], i{{[0-9]+}}* [[T_VAR1_LHS]],
// CHECK: ret void
#endif