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/parallel_for_simd_aligned_c...

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

367 lines
24 KiB
C++
Raw Normal View History

[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --function-signature --include-generated-funcs --replace-value-regex "__omp_offloading_[0-9a-z]+_[0-9a-z]+" "reduction_size[.].+[.]" "pl_cond[.].+[.|,]" --prefix-filecheck-ir-name _
// RUN: %clang_cc1 -verify -fopenmp -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK1
// RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple x86_64-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK2
// RUN: %clang_cc1 -verify -fopenmp -x c++ -std=c++11 -DLAMBDA -triple x86_64-unknown-linux -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK3
// RUN: %clang_cc1 -verify -fopenmp -x c++ -fblocks -DBLOCKS -triple x86_64-unknown-linux -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK4
// RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
// RUN: %clang_cc1 -fopenmp-simd -x c++ -std=c++11 -triple x86_64-unknown-unknown -emit-pch -o %t %s
// RUN: %clang_cc1 -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
// RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -std=c++11 -DLAMBDA -triple x86_64-unknown-linux -emit-llvm %s -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
// RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -fblocks -DBLOCKS -triple x86_64-unknown-linux -emit-llvm %s -o - | FileCheck %s --implicit-check-not="{{__kmpc|__tgt}}"
// expected-no-diagnostics
#ifndef HEADER
#define HEADER
void foo(float *c) {
#pragma omp parallel for simd aligned(c)
for (int i = 0; i < 10; ++i);
}
#endif
// CHECK1-LABEL: define {{[^@]+}}@_Z3fooPf
// CHECK1-SAME: (float* [[C:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[C_ADDR:%.*]] = alloca float*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 8
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: store float* [[C]], float** [[C_ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK1-NEXT: store float** [[C_ADDR]], float*** [[TMP0]], align 8
// CHECK1-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: ret void
//
//
// CHECK1-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[0-9]+]] {
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: entry:
// CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK1-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 8
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[TMP:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
// CHECK1-NEXT: [[I:%.*]] = alloca i32, align 4
// CHECK1-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK1-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK1-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK1-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK1-NEXT: [[TMP2:%.*]] = load float**, float*** [[TMP1]], align 8
// CHECK1-NEXT: [[TMP3:%.*]] = load float*, float** [[TMP2]], align 8
// CHECK1-NEXT: call void @llvm.assume(i1 true) [ "align"(float* [[TMP3]], i64 16) ]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4
// CHECK1-NEXT: store i32 9, i32* [[DOTOMP_UB]], align 4
// CHECK1-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4
// CHECK1-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[TMP4:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK1-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP4]], align 4
// CHECK1-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP5]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
// CHECK1-NEXT: [[TMP6:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
// CHECK1-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 9
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
// CHECK1: cond.true:
// CHECK1-NEXT: br label [[COND_END:%.*]]
// CHECK1: cond.false:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[TMP7:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: br label [[COND_END]]
// CHECK1: cond.end:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[COND:%.*]] = phi i32 [ 9, [[COND_TRUE]] ], [ [[TMP7]], [[COND_FALSE]] ]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
// CHECK1-NEXT: store i32 [[TMP8]], i32* [[DOTOMP_IV]], align 4
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: br label [[OMP_INNER_FOR_COND:%.*]]
// CHECK1: omp.inner.for.cond:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK1-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !llvm.access.group !3
// CHECK1-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP9]], [[TMP10]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
// CHECK1: omp.inner.for.body:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK1-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP11]], 1
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]]
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: store i32 [[ADD]], i32* [[I]], align 4, !llvm.access.group !3
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK1: omp.body.continue:
// CHECK1-NEXT: br label [[OMP_INNER_FOR_INC:%.*]]
// CHECK1: omp.inner.for.inc:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK1-NEXT: [[ADD2:%.*]] = add nsw i32 [[TMP12]], 1
// CHECK1-NEXT: store i32 [[ADD2]], i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK1-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1: omp.inner.for.end:
// CHECK1-NEXT: br label [[OMP_LOOP_EXIT:%.*]]
// CHECK1: omp.loop.exit:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK1-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP5]])
// CHECK1-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IS_LAST]], align 4
// CHECK1-NEXT: [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
// CHECK1-NEXT: br i1 [[TMP14]], label [[DOTOMP_FINAL_THEN:%.*]], label [[DOTOMP_FINAL_DONE:%.*]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK1: .omp.final.then:
// CHECK1-NEXT: store i32 10, i32* [[I]], align 4
// CHECK1-NEXT: br label [[DOTOMP_FINAL_DONE]]
// CHECK1: .omp.final.done:
// CHECK1-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@_Z3fooPf
// CHECK2-SAME: (float* [[C:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[C_ADDR:%.*]] = alloca float*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 8
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: store float* [[C]], float** [[C_ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK2-NEXT: store float** [[C_ADDR]], float*** [[TMP0]], align 8
// CHECK2-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: ret void
//
//
// CHECK2-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[0-9]+]] {
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: entry:
// CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK2-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 8
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[TMP:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
// CHECK2-NEXT: [[I:%.*]] = alloca i32, align 4
// CHECK2-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK2-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK2-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK2-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK2-NEXT: [[TMP2:%.*]] = load float**, float*** [[TMP1]], align 8
// CHECK2-NEXT: [[TMP3:%.*]] = load float*, float** [[TMP2]], align 8
// CHECK2-NEXT: call void @llvm.assume(i1 true) [ "align"(float* [[TMP3]], i64 16) ]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4
// CHECK2-NEXT: store i32 9, i32* [[DOTOMP_UB]], align 4
// CHECK2-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4
// CHECK2-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[TMP4:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK2-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP4]], align 4
// CHECK2-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP5]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
// CHECK2-NEXT: [[TMP6:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
// CHECK2-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 9
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
// CHECK2: cond.true:
// CHECK2-NEXT: br label [[COND_END:%.*]]
// CHECK2: cond.false:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[TMP7:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: br label [[COND_END]]
// CHECK2: cond.end:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[COND:%.*]] = phi i32 [ 9, [[COND_TRUE]] ], [ [[TMP7]], [[COND_FALSE]] ]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
// CHECK2-NEXT: store i32 [[TMP8]], i32* [[DOTOMP_IV]], align 4
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: br label [[OMP_INNER_FOR_COND:%.*]]
// CHECK2: omp.inner.for.cond:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK2-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !llvm.access.group !3
// CHECK2-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP9]], [[TMP10]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
// CHECK2: omp.inner.for.body:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK2-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP11]], 1
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]]
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: store i32 [[ADD]], i32* [[I]], align 4, !llvm.access.group !3
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK2: omp.body.continue:
// CHECK2-NEXT: br label [[OMP_INNER_FOR_INC:%.*]]
// CHECK2: omp.inner.for.inc:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK2-NEXT: [[ADD2:%.*]] = add nsw i32 [[TMP12]], 1
// CHECK2-NEXT: store i32 [[ADD2]], i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK2-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2: omp.inner.for.end:
// CHECK2-NEXT: br label [[OMP_LOOP_EXIT:%.*]]
// CHECK2: omp.loop.exit:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK2-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP5]])
// CHECK2-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IS_LAST]], align 4
// CHECK2-NEXT: [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
// CHECK2-NEXT: br i1 [[TMP14]], label [[DOTOMP_FINAL_THEN:%.*]], label [[DOTOMP_FINAL_DONE:%.*]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK2: .omp.final.then:
// CHECK2-NEXT: store i32 10, i32* [[I]], align 4
// CHECK2-NEXT: br label [[DOTOMP_FINAL_DONE]]
// CHECK2: .omp.final.done:
// CHECK2-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@_Z3fooPf
// CHECK3-SAME: (float* [[C:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[C_ADDR:%.*]] = alloca float*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 8
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: store float* [[C]], float** [[C_ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK3-NEXT: store float** [[C_ADDR]], float*** [[TMP0]], align 8
// CHECK3-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: ret void
//
//
// CHECK3-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[0-9]+]] {
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: entry:
// CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK3-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 8
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[TMP:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
// CHECK3-NEXT: [[I:%.*]] = alloca i32, align 4
// CHECK3-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK3-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK3-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK3-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK3-NEXT: [[TMP2:%.*]] = load float**, float*** [[TMP1]], align 8
// CHECK3-NEXT: [[TMP3:%.*]] = load float*, float** [[TMP2]], align 8
// CHECK3-NEXT: call void @llvm.assume(i1 true) [ "align"(float* [[TMP3]], i64 16) ]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4
// CHECK3-NEXT: store i32 9, i32* [[DOTOMP_UB]], align 4
// CHECK3-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4
// CHECK3-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[TMP4:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK3-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP4]], align 4
// CHECK3-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP5]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
// CHECK3-NEXT: [[TMP6:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
// CHECK3-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 9
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
// CHECK3: cond.true:
// CHECK3-NEXT: br label [[COND_END:%.*]]
// CHECK3: cond.false:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[TMP7:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: br label [[COND_END]]
// CHECK3: cond.end:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[COND:%.*]] = phi i32 [ 9, [[COND_TRUE]] ], [ [[TMP7]], [[COND_FALSE]] ]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
// CHECK3-NEXT: store i32 [[TMP8]], i32* [[DOTOMP_IV]], align 4
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: br label [[OMP_INNER_FOR_COND:%.*]]
// CHECK3: omp.inner.for.cond:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK3-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !llvm.access.group !3
// CHECK3-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP9]], [[TMP10]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
// CHECK3: omp.inner.for.body:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK3-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP11]], 1
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]]
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: store i32 [[ADD]], i32* [[I]], align 4, !llvm.access.group !3
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK3: omp.body.continue:
// CHECK3-NEXT: br label [[OMP_INNER_FOR_INC:%.*]]
// CHECK3: omp.inner.for.inc:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK3-NEXT: [[ADD2:%.*]] = add nsw i32 [[TMP12]], 1
// CHECK3-NEXT: store i32 [[ADD2]], i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK3-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3: omp.inner.for.end:
// CHECK3-NEXT: br label [[OMP_LOOP_EXIT:%.*]]
// CHECK3: omp.loop.exit:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK3-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP5]])
// CHECK3-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IS_LAST]], align 4
// CHECK3-NEXT: [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
// CHECK3-NEXT: br i1 [[TMP14]], label [[DOTOMP_FINAL_THEN:%.*]], label [[DOTOMP_FINAL_DONE:%.*]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK3: .omp.final.then:
// CHECK3-NEXT: store i32 10, i32* [[I]], align 4
// CHECK3-NEXT: br label [[DOTOMP_FINAL_DONE]]
// CHECK3: .omp.final.done:
// CHECK3-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@_Z3fooPf
// CHECK4-SAME: (float* [[C:%.*]]) #[[ATTR0:[0-9]+]] {
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[C_ADDR:%.*]] = alloca float*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[OMP_OUTLINED_ARG_AGG_:%.*]] = alloca [[STRUCT_ANON:%.*]], align 8
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: store float* [[C]], float** [[C_ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[TMP0:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[OMP_OUTLINED_ARG_AGG_]], i32 0, i32 0
// CHECK4-NEXT: store float** [[C_ADDR]], float*** [[TMP0]], align 8
// CHECK4-NEXT: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 1, void (i32*, i32*, ...)* bitcast (void (i32*, i32*, %struct.anon*)* @.omp_outlined. to void (i32*, i32*, ...)*), %struct.anon* [[OMP_OUTLINED_ARG_AGG_]])
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: ret void
//
//
// CHECK4-LABEL: define {{[^@]+}}@.omp_outlined.
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], %struct.anon* noalias [[__CONTEXT:%.*]]) #[[ATTR1:[0-9]+]] {
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: entry:
// CHECK4-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8
// CHECK4-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[__CONTEXT_ADDR:%.*]] = alloca %struct.anon*, align 8
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[TMP:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4
// CHECK4-NEXT: [[I:%.*]] = alloca i32, align 4
// CHECK4-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK4-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: store %struct.anon* [[__CONTEXT]], %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK4-NEXT: [[TMP0:%.*]] = load %struct.anon*, %struct.anon** [[__CONTEXT_ADDR]], align 8
// CHECK4-NEXT: [[TMP1:%.*]] = getelementptr inbounds [[STRUCT_ANON:%.*]], %struct.anon* [[TMP0]], i32 0, i32 0
// CHECK4-NEXT: [[TMP2:%.*]] = load float**, float*** [[TMP1]], align 8
// CHECK4-NEXT: [[TMP3:%.*]] = load float*, float** [[TMP2]], align 8
// CHECK4-NEXT: call void @llvm.assume(i1 true) [ "align"(float* [[TMP3]], i64 16) ]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4
// CHECK4-NEXT: store i32 9, i32* [[DOTOMP_UB]], align 4
// CHECK4-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4
// CHECK4-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[TMP4:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8
// CHECK4-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP4]], align 4
// CHECK4-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB1:[0-9]+]], i32 [[TMP5]], i32 34, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1)
// CHECK4-NEXT: [[TMP6:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
// CHECK4-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP6]], 9
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]]
// CHECK4: cond.true:
// CHECK4-NEXT: br label [[COND_END:%.*]]
// CHECK4: cond.false:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[TMP7:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: br label [[COND_END]]
// CHECK4: cond.end:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[COND:%.*]] = phi i32 [ 9, [[COND_TRUE]] ], [ [[TMP7]], [[COND_FALSE]] ]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4
// CHECK4-NEXT: store i32 [[TMP8]], i32* [[DOTOMP_IV]], align 4
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: br label [[OMP_INNER_FOR_COND:%.*]]
// CHECK4: omp.inner.for.cond:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK4-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !llvm.access.group !3
// CHECK4-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP9]], [[TMP10]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_END:%.*]]
// CHECK4: omp.inner.for.body:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK4-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP11]], 1
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]]
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: store i32 [[ADD]], i32* [[I]], align 4, !llvm.access.group !3
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4-NEXT: br label [[OMP_BODY_CONTINUE:%.*]]
// CHECK4: omp.body.continue:
// CHECK4-NEXT: br label [[OMP_INNER_FOR_INC:%.*]]
// CHECK4: omp.inner.for.inc:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK4-NEXT: [[ADD2:%.*]] = add nsw i32 [[TMP12]], 1
// CHECK4-NEXT: store i32 [[ADD2]], i32* [[DOTOMP_IV]], align 4, !llvm.access.group !3
// CHECK4-NEXT: br label [[OMP_INNER_FOR_COND]], !llvm.loop [[LOOP4:![0-9]+]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4: omp.inner.for.end:
// CHECK4-NEXT: br label [[OMP_LOOP_EXIT:%.*]]
// CHECK4: omp.loop.exit:
[OpenMP] Codegen aggregate for outlined function captures Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call. Reviewed By: jdoerfert, jhuber6 Differential Revision: https://reviews.llvm.org/D102107
2021-09-21 08:12:14 +08:00
// CHECK4-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB1]], i32 [[TMP5]])
// CHECK4-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IS_LAST]], align 4
// CHECK4-NEXT: [[TMP14:%.*]] = icmp ne i32 [[TMP13]], 0
// CHECK4-NEXT: br i1 [[TMP14]], label [[DOTOMP_FINAL_THEN:%.*]], label [[DOTOMP_FINAL_DONE:%.*]]
[OPENMP]Fix PR50699: capture locals in combine directrives for aligned clause. Need to capture locals in aligned clauses for the combined directives to be fix the crash in the codegen. Differential Revision: https://reviews.llvm.org/D104258
2021-06-15 03:28:39 +08:00
// CHECK4: .omp.final.then:
// CHECK4-NEXT: store i32 10, i32* [[I]], align 4
// CHECK4-NEXT: br label [[DOTOMP_FINAL_DONE]]
// CHECK4: .omp.final.done:
// CHECK4-NEXT: ret void
//