brintos

brintos / llvm-project-archived public Read only

0
0
Text · 62.5 KiB · bfcdf9b Raw
906 lines · cpp
1// 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 _2// Test target codegen - host bc file has to be created first.3// RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -fopenmp-cuda-mode -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm-bc %s -o %t-ppc-host.bc4// RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -fopenmp-cuda-mode -x c++ -triple nvptx64-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm %s -fopenmp-is-target-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - | FileCheck %s --check-prefix=CHECK45-645// RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -fopenmp-cuda-mode -x c++ -triple i386-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm-bc %s -o %t-x86-host.bc6// RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -fopenmp-cuda-mode -x c++ -triple nvptx-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm %s -fopenmp-is-target-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck %s --check-prefix=CHECK45-327// RUN: %clang_cc1 -verify -fopenmp -fopenmp-version=45 -fopenmp-cuda-mode -fexceptions -fcxx-exceptions -x c++ -triple nvptx-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm %s -fopenmp-is-target-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck %s --check-prefix=CHECK45-32-EX8 9// RUN: %clang_cc1 -verify -fopenmp -fopenmp-cuda-mode -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm-bc %s -o %t-ppc-host.bc10// RUN: %clang_cc1 -verify -fopenmp -fopenmp-cuda-mode -x c++ -triple nvptx64-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm %s -fopenmp-is-target-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - | FileCheck %s --check-prefix=CHECK-6411// RUN: %clang_cc1 -verify -fopenmp -fopenmp-cuda-mode -x c++ -triple i386-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm-bc %s -o %t-x86-host.bc12// RUN: %clang_cc1 -verify -fopenmp -fopenmp-cuda-mode -x c++ -triple nvptx-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm %s -fopenmp-is-target-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck %s --check-prefix=CHECK-3213// RUN: %clang_cc1 -verify -fopenmp -fopenmp-cuda-mode -fexceptions -fcxx-exceptions -x c++ -triple nvptx-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm %s -fopenmp-is-target-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck %s --check-prefix=CHECK-32-EX14 15// expected-no-diagnostics16#ifndef HEADER17#define HEADER18 19// Check that the execution mode of all 3 target regions on the gpu is set to SPMD Mode.20 21template<typename tx>22tx ftemplate(int n) {23  tx a = 0;24  short aa = 0;25  tx b[10];26 27  #pragma omp target parallel proc_bind(master)28  {29  }30 31  #pragma omp target parallel proc_bind(spread)32  {33    aa += 1;34  }35 36  #pragma omp target parallel proc_bind(close)37  {38    a += 1;39    aa += 1;40    b[2] += 1;41  }42 43  return a;44}45 46int bar(int n){47  int a = 0;48 49  a += ftemplate<int>(n);50 51  return a;52}53 54#endif55// CHECK45-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l2756// CHECK45-64-SAME: (ptr noalias noundef [[DYN_PTR:%.*]]) #[[ATTR0:[0-9]+]] {57// CHECK45-64-NEXT:  entry:58// CHECK45-64-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 859// CHECK45-64-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [0 x ptr], align 860// CHECK45-64-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 861// CHECK45-64-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_kernel_environment, ptr [[DYN_PTR]])62// CHECK45-64-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -163// CHECK45-64-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]64// CHECK45-64:       user_code.entry:65// CHECK45-64-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1:[0-9]+]])66// CHECK45-64-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i64 0)67// CHECK45-64-NEXT:    call void @__kmpc_target_deinit()68// CHECK45-64-NEXT:    ret void69// CHECK45-64:       worker.exit:70// CHECK45-64-NEXT:    ret void71//72//73// CHECK45-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined74// CHECK45-64-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR1:[0-9]+]] {75// CHECK45-64-NEXT:  entry:76// CHECK45-64-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 877// CHECK45-64-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 878// CHECK45-64-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 879// CHECK45-64-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 880// CHECK45-64-NEXT:    ret void81//82//83// CHECK45-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l3184// CHECK45-64-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i64 noundef [[AA:%.*]]) #[[ATTR0]] {85// CHECK45-64-NEXT:  entry:86// CHECK45-64-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 887// CHECK45-64-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 888// CHECK45-64-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 889// CHECK45-64-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [1 x ptr], align 890// CHECK45-64-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 891// CHECK45-64-NEXT:    store i64 [[AA]], ptr [[AA_ADDR]], align 892// CHECK45-64-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_kernel_environment, ptr [[DYN_PTR]])93// CHECK45-64-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -194// CHECK45-64-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]95// CHECK45-64:       user_code.entry:96// CHECK45-64-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])97// CHECK45-64-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 298// CHECK45-64-NEXT:    store i16 [[TMP2]], ptr [[AA_CASTED]], align 299// CHECK45-64-NEXT:    [[TMP3:%.*]] = load i64, ptr [[AA_CASTED]], align 8100// CHECK45-64-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i64 0, i64 0101// CHECK45-64-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP3]] to ptr102// CHECK45-64-NEXT:    store ptr [[TMP5]], ptr [[TMP4]], align 8103// CHECK45-64-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i64 1)104// CHECK45-64-NEXT:    call void @__kmpc_target_deinit()105// CHECK45-64-NEXT:    ret void106// CHECK45-64:       worker.exit:107// CHECK45-64-NEXT:    ret void108//109//110// CHECK45-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined111// CHECK45-64-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i64 noundef [[AA:%.*]]) #[[ATTR1]] {112// CHECK45-64-NEXT:  entry:113// CHECK45-64-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 8114// CHECK45-64-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 8115// CHECK45-64-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8116// CHECK45-64-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 8117// CHECK45-64-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 8118// CHECK45-64-NEXT:    store i64 [[AA]], ptr [[AA_ADDR]], align 8119// CHECK45-64-NEXT:    [[TMP0:%.*]] = load i16, ptr [[AA_ADDR]], align 2120// CHECK45-64-NEXT:    [[CONV:%.*]] = sext i16 [[TMP0]] to i32121// CHECK45-64-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 1122// CHECK45-64-NEXT:    [[CONV1:%.*]] = trunc i32 [[ADD]] to i16123// CHECK45-64-NEXT:    store i16 [[CONV1]], ptr [[AA_ADDR]], align 2124// CHECK45-64-NEXT:    ret void125//126//127// CHECK45-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36128// CHECK45-64-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i64 noundef [[A:%.*]], i64 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {129// CHECK45-64-NEXT:  entry:130// CHECK45-64-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 8131// CHECK45-64-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8132// CHECK45-64-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8133// CHECK45-64-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 8134// CHECK45-64-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8135// CHECK45-64-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8136// CHECK45-64-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x ptr], align 8137// CHECK45-64-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 8138// CHECK45-64-NEXT:    store i64 [[A]], ptr [[A_ADDR]], align 8139// CHECK45-64-NEXT:    store i64 [[AA]], ptr [[AA_ADDR]], align 8140// CHECK45-64-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 8141// CHECK45-64-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 8142// CHECK45-64-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_kernel_environment, ptr [[DYN_PTR]])143// CHECK45-64-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP1]], -1144// CHECK45-64-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]145// CHECK45-64:       user_code.entry:146// CHECK45-64-NEXT:    [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])147// CHECK45-64-NEXT:    [[TMP3:%.*]] = load i32, ptr [[A_ADDR]], align 4148// CHECK45-64-NEXT:    store i32 [[TMP3]], ptr [[A_CASTED]], align 4149// CHECK45-64-NEXT:    [[TMP4:%.*]] = load i64, ptr [[A_CASTED]], align 8150// CHECK45-64-NEXT:    [[TMP5:%.*]] = load i16, ptr [[AA_ADDR]], align 2151// CHECK45-64-NEXT:    store i16 [[TMP5]], ptr [[AA_CASTED]], align 2152// CHECK45-64-NEXT:    [[TMP6:%.*]] = load i64, ptr [[AA_CASTED]], align 8153// CHECK45-64-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i64 0, i64 0154// CHECK45-64-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP4]] to ptr155// CHECK45-64-NEXT:    store ptr [[TMP8]], ptr [[TMP7]], align 8156// CHECK45-64-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i64 0, i64 1157// CHECK45-64-NEXT:    [[TMP10:%.*]] = inttoptr i64 [[TMP6]] to ptr158// CHECK45-64-NEXT:    store ptr [[TMP10]], ptr [[TMP9]], align 8159// CHECK45-64-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i64 0, i64 2160// CHECK45-64-NEXT:    store ptr [[TMP0]], ptr [[TMP11]], align 8161// CHECK45-64-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP2]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i64 3)162// CHECK45-64-NEXT:    call void @__kmpc_target_deinit()163// CHECK45-64-NEXT:    ret void164// CHECK45-64:       worker.exit:165// CHECK45-64-NEXT:    ret void166//167//168// CHECK45-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined169// CHECK45-64-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i64 noundef [[A:%.*]], i64 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR1]] {170// CHECK45-64-NEXT:  entry:171// CHECK45-64-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 8172// CHECK45-64-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 8173// CHECK45-64-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8174// CHECK45-64-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8175// CHECK45-64-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 8176// CHECK45-64-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 8177// CHECK45-64-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 8178// CHECK45-64-NEXT:    store i64 [[A]], ptr [[A_ADDR]], align 8179// CHECK45-64-NEXT:    store i64 [[AA]], ptr [[AA_ADDR]], align 8180// CHECK45-64-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 8181// CHECK45-64-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 8182// CHECK45-64-NEXT:    [[TMP1:%.*]] = load i32, ptr [[A_ADDR]], align 4183// CHECK45-64-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1184// CHECK45-64-NEXT:    store i32 [[ADD]], ptr [[A_ADDR]], align 4185// CHECK45-64-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2186// CHECK45-64-NEXT:    [[CONV:%.*]] = sext i16 [[TMP2]] to i32187// CHECK45-64-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1188// CHECK45-64-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16189// CHECK45-64-NEXT:    store i16 [[CONV2]], ptr [[AA_ADDR]], align 2190// CHECK45-64-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], ptr [[TMP0]], i64 0, i64 2191// CHECK45-64-NEXT:    [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX]], align 4192// CHECK45-64-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP3]], 1193// CHECK45-64-NEXT:    store i32 [[ADD3]], ptr [[ARRAYIDX]], align 4194// CHECK45-64-NEXT:    ret void195//196//197// CHECK45-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27198// CHECK45-32-SAME: (ptr noalias noundef [[DYN_PTR:%.*]]) #[[ATTR0:[0-9]+]] {199// CHECK45-32-NEXT:  entry:200// CHECK45-32-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4201// CHECK45-32-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [0 x ptr], align 4202// CHECK45-32-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4203// CHECK45-32-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_kernel_environment, ptr [[DYN_PTR]])204// CHECK45-32-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1205// CHECK45-32-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]206// CHECK45-32:       user_code.entry:207// CHECK45-32-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1:[0-9]+]])208// CHECK45-32-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 0)209// CHECK45-32-NEXT:    call void @__kmpc_target_deinit()210// CHECK45-32-NEXT:    ret void211// CHECK45-32:       worker.exit:212// CHECK45-32-NEXT:    ret void213//214//215// CHECK45-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined216// CHECK45-32-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR1:[0-9]+]] {217// CHECK45-32-NEXT:  entry:218// CHECK45-32-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4219// CHECK45-32-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4220// CHECK45-32-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4221// CHECK45-32-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4222// CHECK45-32-NEXT:    ret void223//224//225// CHECK45-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31226// CHECK45-32-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i32 noundef [[AA:%.*]]) #[[ATTR0]] {227// CHECK45-32-NEXT:  entry:228// CHECK45-32-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4229// CHECK45-32-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4230// CHECK45-32-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4231// CHECK45-32-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [1 x ptr], align 4232// CHECK45-32-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4233// CHECK45-32-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4234// CHECK45-32-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_kernel_environment, ptr [[DYN_PTR]])235// CHECK45-32-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1236// CHECK45-32-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]237// CHECK45-32:       user_code.entry:238// CHECK45-32-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])239// CHECK45-32-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2240// CHECK45-32-NEXT:    store i16 [[TMP2]], ptr [[AA_CASTED]], align 2241// CHECK45-32-NEXT:    [[TMP3:%.*]] = load i32, ptr [[AA_CASTED]], align 4242// CHECK45-32-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 0243// CHECK45-32-NEXT:    [[TMP5:%.*]] = inttoptr i32 [[TMP3]] to ptr244// CHECK45-32-NEXT:    store ptr [[TMP5]], ptr [[TMP4]], align 4245// CHECK45-32-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 1)246// CHECK45-32-NEXT:    call void @__kmpc_target_deinit()247// CHECK45-32-NEXT:    ret void248// CHECK45-32:       worker.exit:249// CHECK45-32-NEXT:    ret void250//251//252// CHECK45-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined253// CHECK45-32-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i32 noundef [[AA:%.*]]) #[[ATTR1]] {254// CHECK45-32-NEXT:  entry:255// CHECK45-32-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4256// CHECK45-32-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4257// CHECK45-32-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4258// CHECK45-32-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4259// CHECK45-32-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4260// CHECK45-32-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4261// CHECK45-32-NEXT:    [[TMP0:%.*]] = load i16, ptr [[AA_ADDR]], align 2262// CHECK45-32-NEXT:    [[CONV:%.*]] = sext i16 [[TMP0]] to i32263// CHECK45-32-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 1264// CHECK45-32-NEXT:    [[CONV1:%.*]] = trunc i32 [[ADD]] to i16265// CHECK45-32-NEXT:    store i16 [[CONV1]], ptr [[AA_ADDR]], align 2266// CHECK45-32-NEXT:    ret void267//268//269// CHECK45-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36270// CHECK45-32-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i32 noundef [[A:%.*]], i32 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {271// CHECK45-32-NEXT:  entry:272// CHECK45-32-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4273// CHECK45-32-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4274// CHECK45-32-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4275// CHECK45-32-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 4276// CHECK45-32-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4277// CHECK45-32-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4278// CHECK45-32-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x ptr], align 4279// CHECK45-32-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4280// CHECK45-32-NEXT:    store i32 [[A]], ptr [[A_ADDR]], align 4281// CHECK45-32-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4282// CHECK45-32-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 4283// CHECK45-32-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 4284// CHECK45-32-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_kernel_environment, ptr [[DYN_PTR]])285// CHECK45-32-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP1]], -1286// CHECK45-32-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]287// CHECK45-32:       user_code.entry:288// CHECK45-32-NEXT:    [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])289// CHECK45-32-NEXT:    [[TMP3:%.*]] = load i32, ptr [[A_ADDR]], align 4290// CHECK45-32-NEXT:    store i32 [[TMP3]], ptr [[A_CASTED]], align 4291// CHECK45-32-NEXT:    [[TMP4:%.*]] = load i32, ptr [[A_CASTED]], align 4292// CHECK45-32-NEXT:    [[TMP5:%.*]] = load i16, ptr [[AA_ADDR]], align 2293// CHECK45-32-NEXT:    store i16 [[TMP5]], ptr [[AA_CASTED]], align 2294// CHECK45-32-NEXT:    [[TMP6:%.*]] = load i32, ptr [[AA_CASTED]], align 4295// CHECK45-32-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 0296// CHECK45-32-NEXT:    [[TMP8:%.*]] = inttoptr i32 [[TMP4]] to ptr297// CHECK45-32-NEXT:    store ptr [[TMP8]], ptr [[TMP7]], align 4298// CHECK45-32-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 1299// CHECK45-32-NEXT:    [[TMP10:%.*]] = inttoptr i32 [[TMP6]] to ptr300// CHECK45-32-NEXT:    store ptr [[TMP10]], ptr [[TMP9]], align 4301// CHECK45-32-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 2302// CHECK45-32-NEXT:    store ptr [[TMP0]], ptr [[TMP11]], align 4303// CHECK45-32-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP2]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 3)304// CHECK45-32-NEXT:    call void @__kmpc_target_deinit()305// CHECK45-32-NEXT:    ret void306// CHECK45-32:       worker.exit:307// CHECK45-32-NEXT:    ret void308//309//310// CHECK45-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined311// CHECK45-32-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i32 noundef [[A:%.*]], i32 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR1]] {312// CHECK45-32-NEXT:  entry:313// CHECK45-32-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4314// CHECK45-32-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4315// CHECK45-32-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4316// CHECK45-32-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4317// CHECK45-32-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 4318// CHECK45-32-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4319// CHECK45-32-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4320// CHECK45-32-NEXT:    store i32 [[A]], ptr [[A_ADDR]], align 4321// CHECK45-32-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4322// CHECK45-32-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 4323// CHECK45-32-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 4324// CHECK45-32-NEXT:    [[TMP1:%.*]] = load i32, ptr [[A_ADDR]], align 4325// CHECK45-32-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1326// CHECK45-32-NEXT:    store i32 [[ADD]], ptr [[A_ADDR]], align 4327// CHECK45-32-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2328// CHECK45-32-NEXT:    [[CONV:%.*]] = sext i16 [[TMP2]] to i32329// CHECK45-32-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1330// CHECK45-32-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16331// CHECK45-32-NEXT:    store i16 [[CONV2]], ptr [[AA_ADDR]], align 2332// CHECK45-32-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], ptr [[TMP0]], i32 0, i32 2333// CHECK45-32-NEXT:    [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX]], align 4334// CHECK45-32-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP3]], 1335// CHECK45-32-NEXT:    store i32 [[ADD3]], ptr [[ARRAYIDX]], align 4336// CHECK45-32-NEXT:    ret void337//338//339// CHECK45-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27340// CHECK45-32-EX-SAME: (ptr noalias noundef [[DYN_PTR:%.*]]) #[[ATTR0:[0-9]+]] {341// CHECK45-32-EX-NEXT:  entry:342// CHECK45-32-EX-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4343// CHECK45-32-EX-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [0 x ptr], align 4344// CHECK45-32-EX-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4345// CHECK45-32-EX-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_kernel_environment, ptr [[DYN_PTR]])346// CHECK45-32-EX-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1347// CHECK45-32-EX-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]348// CHECK45-32-EX:       user_code.entry:349// CHECK45-32-EX-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1:[0-9]+]])350// CHECK45-32-EX-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 0)351// CHECK45-32-EX-NEXT:    call void @__kmpc_target_deinit()352// CHECK45-32-EX-NEXT:    ret void353// CHECK45-32-EX:       worker.exit:354// CHECK45-32-EX-NEXT:    ret void355//356//357// CHECK45-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined358// CHECK45-32-EX-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR1:[0-9]+]] {359// CHECK45-32-EX-NEXT:  entry:360// CHECK45-32-EX-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4361// CHECK45-32-EX-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4362// CHECK45-32-EX-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4363// CHECK45-32-EX-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4364// CHECK45-32-EX-NEXT:    ret void365//366//367// CHECK45-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31368// CHECK45-32-EX-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i32 noundef [[AA:%.*]]) #[[ATTR0]] {369// CHECK45-32-EX-NEXT:  entry:370// CHECK45-32-EX-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4371// CHECK45-32-EX-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4372// CHECK45-32-EX-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4373// CHECK45-32-EX-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [1 x ptr], align 4374// CHECK45-32-EX-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4375// CHECK45-32-EX-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4376// CHECK45-32-EX-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_kernel_environment, ptr [[DYN_PTR]])377// CHECK45-32-EX-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1378// CHECK45-32-EX-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]379// CHECK45-32-EX:       user_code.entry:380// CHECK45-32-EX-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])381// CHECK45-32-EX-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2382// CHECK45-32-EX-NEXT:    store i16 [[TMP2]], ptr [[AA_CASTED]], align 2383// CHECK45-32-EX-NEXT:    [[TMP3:%.*]] = load i32, ptr [[AA_CASTED]], align 4384// CHECK45-32-EX-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 0385// CHECK45-32-EX-NEXT:    [[TMP5:%.*]] = inttoptr i32 [[TMP3]] to ptr386// CHECK45-32-EX-NEXT:    store ptr [[TMP5]], ptr [[TMP4]], align 4387// CHECK45-32-EX-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 1)388// CHECK45-32-EX-NEXT:    call void @__kmpc_target_deinit()389// CHECK45-32-EX-NEXT:    ret void390// CHECK45-32-EX:       worker.exit:391// CHECK45-32-EX-NEXT:    ret void392//393//394// CHECK45-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined395// CHECK45-32-EX-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i32 noundef [[AA:%.*]]) #[[ATTR1]] {396// CHECK45-32-EX-NEXT:  entry:397// CHECK45-32-EX-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4398// CHECK45-32-EX-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4399// CHECK45-32-EX-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4400// CHECK45-32-EX-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4401// CHECK45-32-EX-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4402// CHECK45-32-EX-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4403// CHECK45-32-EX-NEXT:    [[TMP0:%.*]] = load i16, ptr [[AA_ADDR]], align 2404// CHECK45-32-EX-NEXT:    [[CONV:%.*]] = sext i16 [[TMP0]] to i32405// CHECK45-32-EX-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 1406// CHECK45-32-EX-NEXT:    [[CONV1:%.*]] = trunc i32 [[ADD]] to i16407// CHECK45-32-EX-NEXT:    store i16 [[CONV1]], ptr [[AA_ADDR]], align 2408// CHECK45-32-EX-NEXT:    ret void409//410//411// CHECK45-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36412// CHECK45-32-EX-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i32 noundef [[A:%.*]], i32 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {413// CHECK45-32-EX-NEXT:  entry:414// CHECK45-32-EX-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4415// CHECK45-32-EX-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4416// CHECK45-32-EX-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4417// CHECK45-32-EX-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 4418// CHECK45-32-EX-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4419// CHECK45-32-EX-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4420// CHECK45-32-EX-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x ptr], align 4421// CHECK45-32-EX-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4422// CHECK45-32-EX-NEXT:    store i32 [[A]], ptr [[A_ADDR]], align 4423// CHECK45-32-EX-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4424// CHECK45-32-EX-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 4425// CHECK45-32-EX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 4426// CHECK45-32-EX-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_kernel_environment, ptr [[DYN_PTR]])427// CHECK45-32-EX-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP1]], -1428// CHECK45-32-EX-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]429// CHECK45-32-EX:       user_code.entry:430// CHECK45-32-EX-NEXT:    [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])431// CHECK45-32-EX-NEXT:    [[TMP3:%.*]] = load i32, ptr [[A_ADDR]], align 4432// CHECK45-32-EX-NEXT:    store i32 [[TMP3]], ptr [[A_CASTED]], align 4433// CHECK45-32-EX-NEXT:    [[TMP4:%.*]] = load i32, ptr [[A_CASTED]], align 4434// CHECK45-32-EX-NEXT:    [[TMP5:%.*]] = load i16, ptr [[AA_ADDR]], align 2435// CHECK45-32-EX-NEXT:    store i16 [[TMP5]], ptr [[AA_CASTED]], align 2436// CHECK45-32-EX-NEXT:    [[TMP6:%.*]] = load i32, ptr [[AA_CASTED]], align 4437// CHECK45-32-EX-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 0438// CHECK45-32-EX-NEXT:    [[TMP8:%.*]] = inttoptr i32 [[TMP4]] to ptr439// CHECK45-32-EX-NEXT:    store ptr [[TMP8]], ptr [[TMP7]], align 4440// CHECK45-32-EX-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 1441// CHECK45-32-EX-NEXT:    [[TMP10:%.*]] = inttoptr i32 [[TMP6]] to ptr442// CHECK45-32-EX-NEXT:    store ptr [[TMP10]], ptr [[TMP9]], align 4443// CHECK45-32-EX-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 2444// CHECK45-32-EX-NEXT:    store ptr [[TMP0]], ptr [[TMP11]], align 4445// CHECK45-32-EX-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP2]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 3)446// CHECK45-32-EX-NEXT:    call void @__kmpc_target_deinit()447// CHECK45-32-EX-NEXT:    ret void448// CHECK45-32-EX:       worker.exit:449// CHECK45-32-EX-NEXT:    ret void450//451//452// CHECK45-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined453// CHECK45-32-EX-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i32 noundef [[A:%.*]], i32 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR1]] {454// CHECK45-32-EX-NEXT:  entry:455// CHECK45-32-EX-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4456// CHECK45-32-EX-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4457// CHECK45-32-EX-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4458// CHECK45-32-EX-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4459// CHECK45-32-EX-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 4460// CHECK45-32-EX-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4461// CHECK45-32-EX-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4462// CHECK45-32-EX-NEXT:    store i32 [[A]], ptr [[A_ADDR]], align 4463// CHECK45-32-EX-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4464// CHECK45-32-EX-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 4465// CHECK45-32-EX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 4466// CHECK45-32-EX-NEXT:    [[TMP1:%.*]] = load i32, ptr [[A_ADDR]], align 4467// CHECK45-32-EX-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1468// CHECK45-32-EX-NEXT:    store i32 [[ADD]], ptr [[A_ADDR]], align 4469// CHECK45-32-EX-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2470// CHECK45-32-EX-NEXT:    [[CONV:%.*]] = sext i16 [[TMP2]] to i32471// CHECK45-32-EX-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1472// CHECK45-32-EX-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16473// CHECK45-32-EX-NEXT:    store i16 [[CONV2]], ptr [[AA_ADDR]], align 2474// CHECK45-32-EX-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], ptr [[TMP0]], i32 0, i32 2475// CHECK45-32-EX-NEXT:    [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX]], align 4476// CHECK45-32-EX-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP3]], 1477// CHECK45-32-EX-NEXT:    store i32 [[ADD3]], ptr [[ARRAYIDX]], align 4478// CHECK45-32-EX-NEXT:    ret void479//480//481// CHECK-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27482// CHECK-64-SAME: (ptr noalias noundef [[DYN_PTR:%.*]]) #[[ATTR0:[0-9]+]] {483// CHECK-64-NEXT:  entry:484// CHECK-64-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 8485// CHECK-64-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [0 x ptr], align 8486// CHECK-64-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 8487// CHECK-64-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_kernel_environment, ptr [[DYN_PTR]])488// CHECK-64-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1489// CHECK-64-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]490// CHECK-64:       user_code.entry:491// CHECK-64-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1:[0-9]+]])492// CHECK-64-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i64 0)493// CHECK-64-NEXT:    call void @__kmpc_target_deinit()494// CHECK-64-NEXT:    ret void495// CHECK-64:       worker.exit:496// CHECK-64-NEXT:    ret void497//498//499// CHECK-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined500// CHECK-64-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR1:[0-9]+]] {501// CHECK-64-NEXT:  entry:502// CHECK-64-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 8503// CHECK-64-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 8504// CHECK-64-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 8505// CHECK-64-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 8506// CHECK-64-NEXT:    ret void507//508//509// CHECK-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31510// CHECK-64-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i64 noundef [[AA:%.*]]) #[[ATTR0]] {511// CHECK-64-NEXT:  entry:512// CHECK-64-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 8513// CHECK-64-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8514// CHECK-64-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8515// CHECK-64-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [1 x ptr], align 8516// CHECK-64-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 8517// CHECK-64-NEXT:    store i64 [[AA]], ptr [[AA_ADDR]], align 8518// CHECK-64-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_kernel_environment, ptr [[DYN_PTR]])519// CHECK-64-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1520// CHECK-64-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]521// CHECK-64:       user_code.entry:522// CHECK-64-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])523// CHECK-64-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2524// CHECK-64-NEXT:    store i16 [[TMP2]], ptr [[AA_CASTED]], align 2525// CHECK-64-NEXT:    [[TMP3:%.*]] = load i64, ptr [[AA_CASTED]], align 8526// CHECK-64-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i64 0, i64 0527// CHECK-64-NEXT:    [[TMP5:%.*]] = inttoptr i64 [[TMP3]] to ptr528// CHECK-64-NEXT:    store ptr [[TMP5]], ptr [[TMP4]], align 8529// CHECK-64-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i64 1)530// CHECK-64-NEXT:    call void @__kmpc_target_deinit()531// CHECK-64-NEXT:    ret void532// CHECK-64:       worker.exit:533// CHECK-64-NEXT:    ret void534//535//536// CHECK-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined537// CHECK-64-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i64 noundef [[AA:%.*]]) #[[ATTR1]] {538// CHECK-64-NEXT:  entry:539// CHECK-64-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 8540// CHECK-64-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 8541// CHECK-64-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8542// CHECK-64-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 8543// CHECK-64-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 8544// CHECK-64-NEXT:    store i64 [[AA]], ptr [[AA_ADDR]], align 8545// CHECK-64-NEXT:    [[TMP0:%.*]] = load i16, ptr [[AA_ADDR]], align 2546// CHECK-64-NEXT:    [[CONV:%.*]] = sext i16 [[TMP0]] to i32547// CHECK-64-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 1548// CHECK-64-NEXT:    [[CONV1:%.*]] = trunc i32 [[ADD]] to i16549// CHECK-64-NEXT:    store i16 [[CONV1]], ptr [[AA_ADDR]], align 2550// CHECK-64-NEXT:    ret void551//552//553// CHECK-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36554// CHECK-64-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i64 noundef [[A:%.*]], i64 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {555// CHECK-64-NEXT:  entry:556// CHECK-64-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 8557// CHECK-64-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8558// CHECK-64-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8559// CHECK-64-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 8560// CHECK-64-NEXT:    [[A_CASTED:%.*]] = alloca i64, align 8561// CHECK-64-NEXT:    [[AA_CASTED:%.*]] = alloca i64, align 8562// CHECK-64-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x ptr], align 8563// CHECK-64-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 8564// CHECK-64-NEXT:    store i64 [[A]], ptr [[A_ADDR]], align 8565// CHECK-64-NEXT:    store i64 [[AA]], ptr [[AA_ADDR]], align 8566// CHECK-64-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 8567// CHECK-64-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 8568// CHECK-64-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_kernel_environment, ptr [[DYN_PTR]])569// CHECK-64-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP1]], -1570// CHECK-64-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]571// CHECK-64:       user_code.entry:572// CHECK-64-NEXT:    [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])573// CHECK-64-NEXT:    [[TMP3:%.*]] = load i32, ptr [[A_ADDR]], align 4574// CHECK-64-NEXT:    store i32 [[TMP3]], ptr [[A_CASTED]], align 4575// CHECK-64-NEXT:    [[TMP4:%.*]] = load i64, ptr [[A_CASTED]], align 8576// CHECK-64-NEXT:    [[TMP5:%.*]] = load i16, ptr [[AA_ADDR]], align 2577// CHECK-64-NEXT:    store i16 [[TMP5]], ptr [[AA_CASTED]], align 2578// CHECK-64-NEXT:    [[TMP6:%.*]] = load i64, ptr [[AA_CASTED]], align 8579// CHECK-64-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i64 0, i64 0580// CHECK-64-NEXT:    [[TMP8:%.*]] = inttoptr i64 [[TMP4]] to ptr581// CHECK-64-NEXT:    store ptr [[TMP8]], ptr [[TMP7]], align 8582// CHECK-64-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i64 0, i64 1583// CHECK-64-NEXT:    [[TMP10:%.*]] = inttoptr i64 [[TMP6]] to ptr584// CHECK-64-NEXT:    store ptr [[TMP10]], ptr [[TMP9]], align 8585// CHECK-64-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i64 0, i64 2586// CHECK-64-NEXT:    store ptr [[TMP0]], ptr [[TMP11]], align 8587// CHECK-64-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP2]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i64 3)588// CHECK-64-NEXT:    call void @__kmpc_target_deinit()589// CHECK-64-NEXT:    ret void590// CHECK-64:       worker.exit:591// CHECK-64-NEXT:    ret void592//593//594// CHECK-64-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined595// CHECK-64-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i64 noundef [[A:%.*]], i64 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR1]] {596// CHECK-64-NEXT:  entry:597// CHECK-64-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 8598// CHECK-64-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 8599// CHECK-64-NEXT:    [[A_ADDR:%.*]] = alloca i64, align 8600// CHECK-64-NEXT:    [[AA_ADDR:%.*]] = alloca i64, align 8601// CHECK-64-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 8602// CHECK-64-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 8603// CHECK-64-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 8604// CHECK-64-NEXT:    store i64 [[A]], ptr [[A_ADDR]], align 8605// CHECK-64-NEXT:    store i64 [[AA]], ptr [[AA_ADDR]], align 8606// CHECK-64-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 8607// CHECK-64-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 8608// CHECK-64-NEXT:    [[TMP1:%.*]] = load i32, ptr [[A_ADDR]], align 4609// CHECK-64-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1610// CHECK-64-NEXT:    store i32 [[ADD]], ptr [[A_ADDR]], align 4611// CHECK-64-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2612// CHECK-64-NEXT:    [[CONV:%.*]] = sext i16 [[TMP2]] to i32613// CHECK-64-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1614// CHECK-64-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16615// CHECK-64-NEXT:    store i16 [[CONV2]], ptr [[AA_ADDR]], align 2616// CHECK-64-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], ptr [[TMP0]], i64 0, i64 2617// CHECK-64-NEXT:    [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX]], align 4618// CHECK-64-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP3]], 1619// CHECK-64-NEXT:    store i32 [[ADD3]], ptr [[ARRAYIDX]], align 4620// CHECK-64-NEXT:    ret void621//622//623// CHECK-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27624// CHECK-32-SAME: (ptr noalias noundef [[DYN_PTR:%.*]]) #[[ATTR0:[0-9]+]] {625// CHECK-32-NEXT:  entry:626// CHECK-32-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4627// CHECK-32-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [0 x ptr], align 4628// CHECK-32-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4629// CHECK-32-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_kernel_environment, ptr [[DYN_PTR]])630// CHECK-32-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1631// CHECK-32-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]632// CHECK-32:       user_code.entry:633// CHECK-32-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1:[0-9]+]])634// CHECK-32-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 0)635// CHECK-32-NEXT:    call void @__kmpc_target_deinit()636// CHECK-32-NEXT:    ret void637// CHECK-32:       worker.exit:638// CHECK-32-NEXT:    ret void639//640//641// CHECK-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined642// CHECK-32-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR1:[0-9]+]] {643// CHECK-32-NEXT:  entry:644// CHECK-32-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4645// CHECK-32-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4646// CHECK-32-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4647// CHECK-32-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4648// CHECK-32-NEXT:    ret void649//650//651// CHECK-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31652// CHECK-32-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i32 noundef [[AA:%.*]]) #[[ATTR0]] {653// CHECK-32-NEXT:  entry:654// CHECK-32-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4655// CHECK-32-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4656// CHECK-32-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4657// CHECK-32-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [1 x ptr], align 4658// CHECK-32-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4659// CHECK-32-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4660// CHECK-32-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_kernel_environment, ptr [[DYN_PTR]])661// CHECK-32-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1662// CHECK-32-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]663// CHECK-32:       user_code.entry:664// CHECK-32-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])665// CHECK-32-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2666// CHECK-32-NEXT:    store i16 [[TMP2]], ptr [[AA_CASTED]], align 2667// CHECK-32-NEXT:    [[TMP3:%.*]] = load i32, ptr [[AA_CASTED]], align 4668// CHECK-32-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 0669// CHECK-32-NEXT:    [[TMP5:%.*]] = inttoptr i32 [[TMP3]] to ptr670// CHECK-32-NEXT:    store ptr [[TMP5]], ptr [[TMP4]], align 4671// CHECK-32-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 1)672// CHECK-32-NEXT:    call void @__kmpc_target_deinit()673// CHECK-32-NEXT:    ret void674// CHECK-32:       worker.exit:675// CHECK-32-NEXT:    ret void676//677//678// CHECK-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined679// CHECK-32-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i32 noundef [[AA:%.*]]) #[[ATTR1]] {680// CHECK-32-NEXT:  entry:681// CHECK-32-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4682// CHECK-32-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4683// CHECK-32-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4684// CHECK-32-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4685// CHECK-32-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4686// CHECK-32-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4687// CHECK-32-NEXT:    [[TMP0:%.*]] = load i16, ptr [[AA_ADDR]], align 2688// CHECK-32-NEXT:    [[CONV:%.*]] = sext i16 [[TMP0]] to i32689// CHECK-32-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 1690// CHECK-32-NEXT:    [[CONV1:%.*]] = trunc i32 [[ADD]] to i16691// CHECK-32-NEXT:    store i16 [[CONV1]], ptr [[AA_ADDR]], align 2692// CHECK-32-NEXT:    ret void693//694//695// CHECK-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36696// CHECK-32-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i32 noundef [[A:%.*]], i32 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {697// CHECK-32-NEXT:  entry:698// CHECK-32-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4699// CHECK-32-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4700// CHECK-32-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4701// CHECK-32-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 4702// CHECK-32-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4703// CHECK-32-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4704// CHECK-32-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x ptr], align 4705// CHECK-32-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4706// CHECK-32-NEXT:    store i32 [[A]], ptr [[A_ADDR]], align 4707// CHECK-32-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4708// CHECK-32-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 4709// CHECK-32-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 4710// CHECK-32-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_kernel_environment, ptr [[DYN_PTR]])711// CHECK-32-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP1]], -1712// CHECK-32-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]713// CHECK-32:       user_code.entry:714// CHECK-32-NEXT:    [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])715// CHECK-32-NEXT:    [[TMP3:%.*]] = load i32, ptr [[A_ADDR]], align 4716// CHECK-32-NEXT:    store i32 [[TMP3]], ptr [[A_CASTED]], align 4717// CHECK-32-NEXT:    [[TMP4:%.*]] = load i32, ptr [[A_CASTED]], align 4718// CHECK-32-NEXT:    [[TMP5:%.*]] = load i16, ptr [[AA_ADDR]], align 2719// CHECK-32-NEXT:    store i16 [[TMP5]], ptr [[AA_CASTED]], align 2720// CHECK-32-NEXT:    [[TMP6:%.*]] = load i32, ptr [[AA_CASTED]], align 4721// CHECK-32-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 0722// CHECK-32-NEXT:    [[TMP8:%.*]] = inttoptr i32 [[TMP4]] to ptr723// CHECK-32-NEXT:    store ptr [[TMP8]], ptr [[TMP7]], align 4724// CHECK-32-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 1725// CHECK-32-NEXT:    [[TMP10:%.*]] = inttoptr i32 [[TMP6]] to ptr726// CHECK-32-NEXT:    store ptr [[TMP10]], ptr [[TMP9]], align 4727// CHECK-32-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 2728// CHECK-32-NEXT:    store ptr [[TMP0]], ptr [[TMP11]], align 4729// CHECK-32-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP2]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 3)730// CHECK-32-NEXT:    call void @__kmpc_target_deinit()731// CHECK-32-NEXT:    ret void732// CHECK-32:       worker.exit:733// CHECK-32-NEXT:    ret void734//735//736// CHECK-32-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined737// CHECK-32-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i32 noundef [[A:%.*]], i32 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR1]] {738// CHECK-32-NEXT:  entry:739// CHECK-32-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4740// CHECK-32-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4741// CHECK-32-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4742// CHECK-32-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4743// CHECK-32-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 4744// CHECK-32-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4745// CHECK-32-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4746// CHECK-32-NEXT:    store i32 [[A]], ptr [[A_ADDR]], align 4747// CHECK-32-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4748// CHECK-32-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 4749// CHECK-32-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 4750// CHECK-32-NEXT:    [[TMP1:%.*]] = load i32, ptr [[A_ADDR]], align 4751// CHECK-32-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1752// CHECK-32-NEXT:    store i32 [[ADD]], ptr [[A_ADDR]], align 4753// CHECK-32-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2754// CHECK-32-NEXT:    [[CONV:%.*]] = sext i16 [[TMP2]] to i32755// CHECK-32-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1756// CHECK-32-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16757// CHECK-32-NEXT:    store i16 [[CONV2]], ptr [[AA_ADDR]], align 2758// CHECK-32-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], ptr [[TMP0]], i32 0, i32 2759// CHECK-32-NEXT:    [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX]], align 4760// CHECK-32-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP3]], 1761// CHECK-32-NEXT:    store i32 [[ADD3]], ptr [[ARRAYIDX]], align 4762// CHECK-32-NEXT:    ret void763//764//765// CHECK-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27766// CHECK-32-EX-SAME: (ptr noalias noundef [[DYN_PTR:%.*]]) #[[ATTR0:[0-9]+]] {767// CHECK-32-EX-NEXT:  entry:768// CHECK-32-EX-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4769// CHECK-32-EX-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [0 x ptr], align 4770// CHECK-32-EX-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4771// CHECK-32-EX-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_kernel_environment, ptr [[DYN_PTR]])772// CHECK-32-EX-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1773// CHECK-32-EX-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]774// CHECK-32-EX:       user_code.entry:775// CHECK-32-EX-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1:[0-9]+]])776// CHECK-32-EX-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 0)777// CHECK-32-EX-NEXT:    call void @__kmpc_target_deinit()778// CHECK-32-EX-NEXT:    ret void779// CHECK-32-EX:       worker.exit:780// CHECK-32-EX-NEXT:    ret void781//782//783// CHECK-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l27_omp_outlined784// CHECK-32-EX-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]]) #[[ATTR1:[0-9]+]] {785// CHECK-32-EX-NEXT:  entry:786// CHECK-32-EX-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4787// CHECK-32-EX-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4788// CHECK-32-EX-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4789// CHECK-32-EX-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4790// CHECK-32-EX-NEXT:    ret void791//792//793// CHECK-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31794// CHECK-32-EX-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i32 noundef [[AA:%.*]]) #[[ATTR0]] {795// CHECK-32-EX-NEXT:  entry:796// CHECK-32-EX-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4797// CHECK-32-EX-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4798// CHECK-32-EX-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4799// CHECK-32-EX-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [1 x ptr], align 4800// CHECK-32-EX-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4801// CHECK-32-EX-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4802// CHECK-32-EX-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_kernel_environment, ptr [[DYN_PTR]])803// CHECK-32-EX-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1804// CHECK-32-EX-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]805// CHECK-32-EX:       user_code.entry:806// CHECK-32-EX-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])807// CHECK-32-EX-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2808// CHECK-32-EX-NEXT:    store i16 [[TMP2]], ptr [[AA_CASTED]], align 2809// CHECK-32-EX-NEXT:    [[TMP3:%.*]] = load i32, ptr [[AA_CASTED]], align 4810// CHECK-32-EX-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [1 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 0811// CHECK-32-EX-NEXT:    [[TMP5:%.*]] = inttoptr i32 [[TMP3]] to ptr812// CHECK-32-EX-NEXT:    store ptr [[TMP5]], ptr [[TMP4]], align 4813// CHECK-32-EX-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP1]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 1)814// CHECK-32-EX-NEXT:    call void @__kmpc_target_deinit()815// CHECK-32-EX-NEXT:    ret void816// CHECK-32-EX:       worker.exit:817// CHECK-32-EX-NEXT:    ret void818//819//820// CHECK-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l31_omp_outlined821// CHECK-32-EX-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i32 noundef [[AA:%.*]]) #[[ATTR1]] {822// CHECK-32-EX-NEXT:  entry:823// CHECK-32-EX-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4824// CHECK-32-EX-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4825// CHECK-32-EX-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4826// CHECK-32-EX-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4827// CHECK-32-EX-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4828// CHECK-32-EX-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4829// CHECK-32-EX-NEXT:    [[TMP0:%.*]] = load i16, ptr [[AA_ADDR]], align 2830// CHECK-32-EX-NEXT:    [[CONV:%.*]] = sext i16 [[TMP0]] to i32831// CHECK-32-EX-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 1832// CHECK-32-EX-NEXT:    [[CONV1:%.*]] = trunc i32 [[ADD]] to i16833// CHECK-32-EX-NEXT:    store i16 [[CONV1]], ptr [[AA_ADDR]], align 2834// CHECK-32-EX-NEXT:    ret void835//836//837// CHECK-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36838// CHECK-32-EX-SAME: (ptr noalias noundef [[DYN_PTR:%.*]], i32 noundef [[A:%.*]], i32 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR0]] {839// CHECK-32-EX-NEXT:  entry:840// CHECK-32-EX-NEXT:    [[DYN_PTR_ADDR:%.*]] = alloca ptr, align 4841// CHECK-32-EX-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4842// CHECK-32-EX-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4843// CHECK-32-EX-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 4844// CHECK-32-EX-NEXT:    [[A_CASTED:%.*]] = alloca i32, align 4845// CHECK-32-EX-NEXT:    [[AA_CASTED:%.*]] = alloca i32, align 4846// CHECK-32-EX-NEXT:    [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x ptr], align 4847// CHECK-32-EX-NEXT:    store ptr [[DYN_PTR]], ptr [[DYN_PTR_ADDR]], align 4848// CHECK-32-EX-NEXT:    store i32 [[A]], ptr [[A_ADDR]], align 4849// CHECK-32-EX-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4850// CHECK-32-EX-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 4851// CHECK-32-EX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 4852// CHECK-32-EX-NEXT:    [[TMP1:%.*]] = call i32 @__kmpc_target_init(ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_kernel_environment, ptr [[DYN_PTR]])853// CHECK-32-EX-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP1]], -1854// CHECK-32-EX-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]855// CHECK-32-EX:       user_code.entry:856// CHECK-32-EX-NEXT:    [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(ptr @[[GLOB1]])857// CHECK-32-EX-NEXT:    [[TMP3:%.*]] = load i32, ptr [[A_ADDR]], align 4858// CHECK-32-EX-NEXT:    store i32 [[TMP3]], ptr [[A_CASTED]], align 4859// CHECK-32-EX-NEXT:    [[TMP4:%.*]] = load i32, ptr [[A_CASTED]], align 4860// CHECK-32-EX-NEXT:    [[TMP5:%.*]] = load i16, ptr [[AA_ADDR]], align 2861// CHECK-32-EX-NEXT:    store i16 [[TMP5]], ptr [[AA_CASTED]], align 2862// CHECK-32-EX-NEXT:    [[TMP6:%.*]] = load i32, ptr [[AA_CASTED]], align 4863// CHECK-32-EX-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 0864// CHECK-32-EX-NEXT:    [[TMP8:%.*]] = inttoptr i32 [[TMP4]] to ptr865// CHECK-32-EX-NEXT:    store ptr [[TMP8]], ptr [[TMP7]], align 4866// CHECK-32-EX-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 1867// CHECK-32-EX-NEXT:    [[TMP10:%.*]] = inttoptr i32 [[TMP6]] to ptr868// CHECK-32-EX-NEXT:    store ptr [[TMP10]], ptr [[TMP9]], align 4869// CHECK-32-EX-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [3 x ptr], ptr [[CAPTURED_VARS_ADDRS]], i32 0, i32 2870// CHECK-32-EX-NEXT:    store ptr [[TMP0]], ptr [[TMP11]], align 4871// CHECK-32-EX-NEXT:    call void @__kmpc_parallel_51(ptr @[[GLOB1]], i32 [[TMP2]], i32 1, i32 -1, i32 -1, ptr @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined, ptr null, ptr [[CAPTURED_VARS_ADDRS]], i32 3)872// CHECK-32-EX-NEXT:    call void @__kmpc_target_deinit()873// CHECK-32-EX-NEXT:    ret void874// CHECK-32-EX:       worker.exit:875// CHECK-32-EX-NEXT:    ret void876//877//878// CHECK-32-EX-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z9ftemplateIiET_i_l36_omp_outlined879// CHECK-32-EX-SAME: (ptr noalias noundef [[DOTGLOBAL_TID_:%.*]], ptr noalias noundef [[DOTBOUND_TID_:%.*]], i32 noundef [[A:%.*]], i32 noundef [[AA:%.*]], ptr noundef nonnull align 4 dereferenceable(40) [[B:%.*]]) #[[ATTR1]] {880// CHECK-32-EX-NEXT:  entry:881// CHECK-32-EX-NEXT:    [[DOTGLOBAL_TID__ADDR:%.*]] = alloca ptr, align 4882// CHECK-32-EX-NEXT:    [[DOTBOUND_TID__ADDR:%.*]] = alloca ptr, align 4883// CHECK-32-EX-NEXT:    [[A_ADDR:%.*]] = alloca i32, align 4884// CHECK-32-EX-NEXT:    [[AA_ADDR:%.*]] = alloca i32, align 4885// CHECK-32-EX-NEXT:    [[B_ADDR:%.*]] = alloca ptr, align 4886// CHECK-32-EX-NEXT:    store ptr [[DOTGLOBAL_TID_]], ptr [[DOTGLOBAL_TID__ADDR]], align 4887// CHECK-32-EX-NEXT:    store ptr [[DOTBOUND_TID_]], ptr [[DOTBOUND_TID__ADDR]], align 4888// CHECK-32-EX-NEXT:    store i32 [[A]], ptr [[A_ADDR]], align 4889// CHECK-32-EX-NEXT:    store i32 [[AA]], ptr [[AA_ADDR]], align 4890// CHECK-32-EX-NEXT:    store ptr [[B]], ptr [[B_ADDR]], align 4891// CHECK-32-EX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[B_ADDR]], align 4892// CHECK-32-EX-NEXT:    [[TMP1:%.*]] = load i32, ptr [[A_ADDR]], align 4893// CHECK-32-EX-NEXT:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 1894// CHECK-32-EX-NEXT:    store i32 [[ADD]], ptr [[A_ADDR]], align 4895// CHECK-32-EX-NEXT:    [[TMP2:%.*]] = load i16, ptr [[AA_ADDR]], align 2896// CHECK-32-EX-NEXT:    [[CONV:%.*]] = sext i16 [[TMP2]] to i32897// CHECK-32-EX-NEXT:    [[ADD1:%.*]] = add nsw i32 [[CONV]], 1898// CHECK-32-EX-NEXT:    [[CONV2:%.*]] = trunc i32 [[ADD1]] to i16899// CHECK-32-EX-NEXT:    store i16 [[CONV2]], ptr [[AA_ADDR]], align 2900// CHECK-32-EX-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x i32], ptr [[TMP0]], i32 0, i32 2901// CHECK-32-EX-NEXT:    [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX]], align 4902// CHECK-32-EX-NEXT:    [[ADD3:%.*]] = add nsw i32 [[TMP3]], 1903// CHECK-32-EX-NEXT:    store i32 [[ADD3]], ptr [[ARRAYIDX]], align 4904// CHECK-32-EX-NEXT:    ret void905//906