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 // RUN: %clang_cc1 -x c++ -O1 -disable-llvm-optzns -verify -fopenmp -internal-isystem %S/../Headers/Inputs/include -internal-isystem %S/../../lib/Headers/openmp_wrappers -include __clang_openmp_device_functions.h -triple powerpc64le-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm-bc %s -o %t-ppc-host.bc 3 // RUN: %clang_cc1 -x c++ -O1 -disable-llvm-optzns -verify -fopenmp -internal-isystem %S/../Headers/Inputs/include -internal-isystem %S/../../lib/Headers/openmp_wrappers -include __clang_openmp_device_functions.h -triple nvptx64-unknown-unknown -aux-triple powerpc64le-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - | FileCheck %s --check-prefix=CHECK1 4 // RUN: %clang_cc1 -x c++ -O1 -disable-llvm-optzns -verify -fopenmp -internal-isystem %S/../Headers/Inputs/include -internal-isystem %S/../../lib/Headers/openmp_wrappers -include __clang_openmp_device_functions.h -triple i386-unknown-unknown -fopenmp-targets=nvptx-nvidia-cuda -emit-llvm-bc %s -o %t-x86-host.bc 5 // RUN: %clang_cc1 -x c++ -O1 -disable-llvm-optzns -verify -fopenmp -internal-isystem %S/../Headers/Inputs/include -internal-isystem %S/../../lib/Headers/openmp_wrappers -include __clang_openmp_device_functions.h -triple nvptx64-unknown-unknown -aux-triple powerpc64le-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck %s --check-prefix=CHECK2 6 // RUN: %clang_cc1 -x c++ -O1 -disable-llvm-optzns -verify -fopenmp -internal-isystem %S/../Headers/Inputs/include -internal-isystem %S/../../lib/Headers/openmp_wrappers -include __clang_openmp_device_functions.h -fexceptions -fcxx-exceptions -aux-triple powerpc64le-unknown-unknown -triple nvptx64-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck %s --check-prefix=CHECK3 7 // expected-no-diagnostics 8 #ifndef HEADER 9 #define HEADER 10 11 #include <complex> 12 13 // Verify we do not add tbaa metadata to type punned memory operations: 14 15 16 17 template <typename T> 18 void complex_reduction() { 19 #pragma omp target teams distribute 20 for (int ib = 0; ib < 100; ib++) { 21 std::complex<T> partial_sum; 22 const int istart = ib * 4; 23 const int iend = (ib + 1) * 4; 24 #pragma omp parallel for reduction(+ \ 25 : partial_sum) 26 for (int i = istart; i < iend; i++) 27 partial_sum += std::complex<T>(i, i); 28 } 29 } 30 31 void test() { 32 complex_reduction<float>(); 33 complex_reduction<double>(); 34 } 35 #endif 36 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z17complex_reductionIfEvv_l19 37 // CHECK1-SAME: () #[[ATTR0:[0-9]+]] { 38 // CHECK1-NEXT: entry: 39 // CHECK1-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 40 // CHECK1-NEXT: [[DOTTHREADID_TEMP_:%.*]] = alloca i32, align 4 41 // CHECK1-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 42 // CHECK1-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* @[[GLOB1:[0-9]+]], i1 false, i1 true, i1 true) 43 // CHECK1-NEXT: [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1 44 // CHECK1-NEXT: br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]] 45 // CHECK1: user_code.entry: 46 // CHECK1-NEXT: [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 47 // CHECK1-NEXT: store i32 [[TMP1]], i32* [[DOTTHREADID_TEMP_]], align 4, !tbaa [[TBAA8:![0-9]+]] 48 // CHECK1-NEXT: call void @__omp_outlined__(i32* [[DOTTHREADID_TEMP_]], i32* [[DOTZERO_ADDR]]) #[[ATTR1:[0-9]+]] 49 // CHECK1-NEXT: call void @__kmpc_target_deinit(%struct.ident_t* @[[GLOB1]], i1 false, i1 true) 50 // CHECK1-NEXT: ret void 51 // CHECK1: worker.exit: 52 // CHECK1-NEXT: ret void 53 // 54 // 55 // CHECK1-LABEL: define {{[^@]+}}@__omp_outlined__ 56 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR0]] { 57 // CHECK1-NEXT: entry: 58 // CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 59 // CHECK1-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 60 // CHECK1-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 61 // CHECK1-NEXT: [[TMP:%.*]] = alloca i32, align 4 62 // CHECK1-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 63 // CHECK1-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 64 // CHECK1-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 65 // CHECK1-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 66 // CHECK1-NEXT: [[IB:%.*]] = alloca i32, align 4 67 // CHECK1-NEXT: [[REF_TMP:%.*]] = alloca float, align 4 68 // CHECK1-NEXT: [[REF_TMP2:%.*]] = alloca float, align 4 69 // CHECK1-NEXT: [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x i8*], align 8 70 // CHECK1-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12:![0-9]+]] 71 // CHECK1-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 72 // CHECK1-NEXT: [[ISTART:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 73 // CHECK1-NEXT: [[ISTART_ON_STACK:%.*]] = bitcast i8* [[ISTART]] to i32* 74 // CHECK1-NEXT: [[IEND:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 75 // CHECK1-NEXT: [[IEND_ON_STACK:%.*]] = bitcast i8* [[IEND]] to i32* 76 // CHECK1-NEXT: [[PARTIAL_SUM:%.*]] = call i8* @__kmpc_alloc_shared(i64 8) 77 // CHECK1-NEXT: [[PARTIAL_SUM_ON_STACK:%.*]] = bitcast i8* [[PARTIAL_SUM]] to %"class.std::complex"* 78 // CHECK1-NEXT: [[TMP0:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 79 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP0]]) #[[ATTR1]] 80 // CHECK1-NEXT: [[TMP1:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 81 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP1]]) #[[ATTR1]] 82 // CHECK1-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 83 // CHECK1-NEXT: [[TMP2:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 84 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP2]]) #[[ATTR1]] 85 // CHECK1-NEXT: store i32 99, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 86 // CHECK1-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 87 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 88 // CHECK1-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 89 // CHECK1-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 90 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 91 // CHECK1-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 92 // CHECK1-NEXT: [[TMP5:%.*]] = bitcast i32* [[IB]] to i8* 93 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP5]]) #[[ATTR1]] 94 // CHECK1-NEXT: [[TMP6:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 95 // CHECK1-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP6]], align 4, !tbaa [[TBAA8]] 96 // CHECK1-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP7]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 97 // CHECK1-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 98 // CHECK1-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP8]], 99 99 // CHECK1-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 100 // CHECK1: cond.true: 101 // CHECK1-NEXT: br label [[COND_END:%.*]] 102 // CHECK1: cond.false: 103 // CHECK1-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 104 // CHECK1-NEXT: br label [[COND_END]] 105 // CHECK1: cond.end: 106 // CHECK1-NEXT: [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP9]], [[COND_FALSE]] ] 107 // CHECK1-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 108 // CHECK1-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 109 // CHECK1-NEXT: store i32 [[TMP10]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 110 // CHECK1-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 111 // CHECK1: omp.inner.for.cond: 112 // CHECK1-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 113 // CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 114 // CHECK1-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP11]], [[TMP12]] 115 // CHECK1-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 116 // CHECK1: omp.inner.for.cond.cleanup: 117 // CHECK1-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 118 // CHECK1: omp.inner.for.body: 119 // CHECK1-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 120 // CHECK1-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP13]], 1 121 // CHECK1-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]] 122 // CHECK1-NEXT: store i32 [[ADD]], i32* [[IB]], align 4, !tbaa [[TBAA8]] 123 // CHECK1-NEXT: [[TMP14:%.*]] = bitcast float* [[REF_TMP]] to i8* 124 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP14]]) #[[ATTR1]] 125 // CHECK1-NEXT: store float 0.000000e+00, float* [[REF_TMP]], align 4, !tbaa [[TBAA14:![0-9]+]] 126 // CHECK1-NEXT: [[TMP15:%.*]] = bitcast float* [[REF_TMP2]] to i8* 127 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP15]]) #[[ATTR1]] 128 // CHECK1-NEXT: store float 0.000000e+00, float* [[REF_TMP2]], align 4, !tbaa [[TBAA14]] 129 // CHECK1-NEXT: call void @_ZNSt7complexIfEC1ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM_ON_STACK]], float* nonnull align 4 dereferenceable(4) [[REF_TMP]], float* nonnull align 4 dereferenceable(4) [[REF_TMP2]]) #[[ATTR8:[0-9]+]] 130 // CHECK1-NEXT: [[TMP16:%.*]] = bitcast float* [[REF_TMP2]] to i8* 131 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP16]]) #[[ATTR1]] 132 // CHECK1-NEXT: [[TMP17:%.*]] = bitcast float* [[REF_TMP]] to i8* 133 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP17]]) #[[ATTR1]] 134 // CHECK1-NEXT: [[TMP18:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 135 // CHECK1-NEXT: [[MUL3:%.*]] = mul nsw i32 [[TMP18]], 4 136 // CHECK1-NEXT: store i32 [[MUL3]], i32* [[ISTART_ON_STACK]], align 4, !tbaa [[TBAA8]] 137 // CHECK1-NEXT: [[TMP19:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 138 // CHECK1-NEXT: [[ADD4:%.*]] = add nsw i32 [[TMP19]], 1 139 // CHECK1-NEXT: [[MUL5:%.*]] = mul nsw i32 [[ADD4]], 4 140 // CHECK1-NEXT: store i32 [[MUL5]], i32* [[IEND_ON_STACK]], align 4, !tbaa [[TBAA8]] 141 // CHECK1-NEXT: [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 0 142 // CHECK1-NEXT: [[TMP21:%.*]] = bitcast i32* [[ISTART_ON_STACK]] to i8* 143 // CHECK1-NEXT: store i8* [[TMP21]], i8** [[TMP20]], align 8, !tbaa [[TBAA12]] 144 // CHECK1-NEXT: [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 1 145 // CHECK1-NEXT: [[TMP23:%.*]] = bitcast i32* [[IEND_ON_STACK]] to i8* 146 // CHECK1-NEXT: store i8* [[TMP23]], i8** [[TMP22]], align 8, !tbaa [[TBAA12]] 147 // CHECK1-NEXT: [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 2 148 // CHECK1-NEXT: [[TMP25:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM_ON_STACK]] to i8* 149 // CHECK1-NEXT: store i8* [[TMP25]], i8** [[TMP24]], align 8, !tbaa [[TBAA12]] 150 // CHECK1-NEXT: [[TMP26:%.*]] = bitcast [3 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8** 151 // CHECK1-NEXT: call void @__kmpc_parallel_51(%struct.ident_t* @[[GLOB1]], i32 [[TMP7]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*, i32*, i32*, %"class.std::complex"*)* @__omp_outlined__1 to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined__1_wrapper to i8*), i8** [[TMP26]], i64 3) 152 // CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 153 // CHECK1: omp.body.continue: 154 // CHECK1-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 155 // CHECK1: omp.inner.for.inc: 156 // CHECK1-NEXT: [[TMP27:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 157 // CHECK1-NEXT: [[ADD6:%.*]] = add nsw i32 [[TMP27]], 1 158 // CHECK1-NEXT: store i32 [[ADD6]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 159 // CHECK1-NEXT: br label [[OMP_INNER_FOR_COND]] 160 // CHECK1: omp.inner.for.end: 161 // CHECK1-NEXT: br label [[OMP_LOOP_EXIT:%.*]] 162 // CHECK1: omp.loop.exit: 163 // CHECK1-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB2]], i32 [[TMP7]]) 164 // CHECK1-NEXT: [[TMP28:%.*]] = bitcast i32* [[IB]] to i8* 165 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP28]]) #[[ATTR1]] 166 // CHECK1-NEXT: [[TMP29:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 167 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP29]]) #[[ATTR1]] 168 // CHECK1-NEXT: [[TMP30:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 169 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP30]]) #[[ATTR1]] 170 // CHECK1-NEXT: [[TMP31:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 171 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP31]]) #[[ATTR1]] 172 // CHECK1-NEXT: [[TMP32:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 173 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP32]]) #[[ATTR1]] 174 // CHECK1-NEXT: [[TMP33:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 175 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP33]]) #[[ATTR1]] 176 // CHECK1-NEXT: call void @__kmpc_free_shared(i8* [[PARTIAL_SUM]], i64 8) 177 // CHECK1-NEXT: call void @__kmpc_free_shared(i8* [[IEND]], i64 4) 178 // CHECK1-NEXT: call void @__kmpc_free_shared(i8* [[ISTART]], i64 4) 179 // CHECK1-NEXT: ret void 180 // 181 // 182 // CHECK1-LABEL: define {{[^@]+}}@_ZNSt7complexIfEC1ERKfS2_ 183 // CHECK1-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]], float* nonnull align 4 dereferenceable(4) [[__RE:%.*]], float* nonnull align 4 dereferenceable(4) [[__IM:%.*]]) unnamed_addr #[[ATTR3:[0-9]+]] comdat align 2 { 184 // CHECK1-NEXT: entry: 185 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 186 // CHECK1-NEXT: [[__RE_ADDR:%.*]] = alloca float*, align 8 187 // CHECK1-NEXT: [[__IM_ADDR:%.*]] = alloca float*, align 8 188 // CHECK1-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 189 // CHECK1-NEXT: store float* [[__RE]], float** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 190 // CHECK1-NEXT: store float* [[__IM]], float** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 191 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 192 // CHECK1-NEXT: [[TMP0:%.*]] = load float*, float** [[__RE_ADDR]], align 8 193 // CHECK1-NEXT: [[TMP1:%.*]] = load float*, float** [[__IM_ADDR]], align 8 194 // CHECK1-NEXT: call void @_ZNSt7complexIfEC2ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS1]], float* nonnull align 4 dereferenceable(4) [[TMP0]], float* nonnull align 4 dereferenceable(4) [[TMP1]]) #[[ATTR8]] 195 // CHECK1-NEXT: ret void 196 // 197 // 198 // CHECK1-LABEL: define {{[^@]+}}@__omp_outlined__1 199 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ISTART:%.*]], i32* nonnull align 4 dereferenceable(4) [[IEND:%.*]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM:%.*]]) #[[ATTR0]] { 200 // CHECK1-NEXT: entry: 201 // CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 202 // CHECK1-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 203 // CHECK1-NEXT: [[ISTART_ADDR:%.*]] = alloca i32*, align 8 204 // CHECK1-NEXT: [[IEND_ADDR:%.*]] = alloca i32*, align 8 205 // CHECK1-NEXT: [[PARTIAL_SUM_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 206 // CHECK1-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 207 // CHECK1-NEXT: [[TMP:%.*]] = alloca i32, align 4 208 // CHECK1-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4 209 // CHECK1-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4 210 // CHECK1-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4 211 // CHECK1-NEXT: [[I:%.*]] = alloca i32, align 4 212 // CHECK1-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 213 // CHECK1-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 214 // CHECK1-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 215 // CHECK1-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 216 // CHECK1-NEXT: [[PARTIAL_SUM5:%.*]] = alloca %"class.std::complex", align 4 217 // CHECK1-NEXT: [[REF_TMP:%.*]] = alloca float, align 4 218 // CHECK1-NEXT: [[REF_TMP6:%.*]] = alloca float, align 4 219 // CHECK1-NEXT: [[I7:%.*]] = alloca i32, align 4 220 // CHECK1-NEXT: [[REF_TMP14:%.*]] = alloca %"class.std::complex", align 4 221 // CHECK1-NEXT: [[REF_TMP15:%.*]] = alloca float, align 4 222 // CHECK1-NEXT: [[REF_TMP16:%.*]] = alloca float, align 4 223 // CHECK1-NEXT: [[DOTOMP_REDUCTION_RED_LIST:%.*]] = alloca [1 x i8*], align 8 224 // CHECK1-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12]] 225 // CHECK1-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 226 // CHECK1-NEXT: store i32* [[ISTART]], i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 227 // CHECK1-NEXT: store i32* [[IEND]], i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 228 // CHECK1-NEXT: store %"class.std::complex"* [[PARTIAL_SUM]], %"class.std::complex"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 229 // CHECK1-NEXT: [[TMP0:%.*]] = load i32*, i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 230 // CHECK1-NEXT: [[TMP1:%.*]] = load i32*, i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 231 // CHECK1-NEXT: [[TMP2:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 232 // CHECK1-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 233 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 234 // CHECK1-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 235 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 236 // CHECK1-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP0]], align 4, !tbaa [[TBAA8]] 237 // CHECK1-NEXT: store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 238 // CHECK1-NEXT: [[TMP6:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 239 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP6]]) #[[ATTR1]] 240 // CHECK1-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP1]], align 4, !tbaa [[TBAA8]] 241 // CHECK1-NEXT: store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 242 // CHECK1-NEXT: [[TMP8:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 243 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP8]]) #[[ATTR1]] 244 // CHECK1-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 245 // CHECK1-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 246 // CHECK1-NEXT: [[SUB:%.*]] = sub i32 [[TMP9]], [[TMP10]] 247 // CHECK1-NEXT: [[SUB3:%.*]] = sub i32 [[SUB]], 1 248 // CHECK1-NEXT: [[ADD:%.*]] = add i32 [[SUB3]], 1 249 // CHECK1-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], 1 250 // CHECK1-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1 251 // CHECK1-NEXT: store i32 [[SUB4]], i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 252 // CHECK1-NEXT: [[TMP11:%.*]] = bitcast i32* [[I]] to i8* 253 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP11]]) #[[ATTR1]] 254 // CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 255 // CHECK1-NEXT: store i32 [[TMP12]], i32* [[I]], align 4, !tbaa [[TBAA8]] 256 // CHECK1-NEXT: [[TMP13:%.*]] = bitcast i32* [[I]] to i8* 257 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP13]]) #[[ATTR1]] 258 // CHECK1-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 259 // CHECK1-NEXT: [[TMP15:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 260 // CHECK1-NEXT: [[CMP:%.*]] = icmp slt i32 [[TMP14]], [[TMP15]] 261 // CHECK1-NEXT: br i1 [[CMP]], label [[OMP_PRECOND_THEN:%.*]], label [[OMP_PRECOND_END:%.*]] 262 // CHECK1: omp.precond.then: 263 // CHECK1-NEXT: [[TMP16:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 264 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP16]]) #[[ATTR1]] 265 // CHECK1-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 266 // CHECK1-NEXT: [[TMP17:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 267 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP17]]) #[[ATTR1]] 268 // CHECK1-NEXT: [[TMP18:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 269 // CHECK1-NEXT: store i32 [[TMP18]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 270 // CHECK1-NEXT: [[TMP19:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 271 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP19]]) #[[ATTR1]] 272 // CHECK1-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 273 // CHECK1-NEXT: [[TMP20:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 274 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP20]]) #[[ATTR1]] 275 // CHECK1-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 276 // CHECK1-NEXT: [[TMP21:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM5]] to i8* 277 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP21]]) #[[ATTR1]] 278 // CHECK1-NEXT: [[TMP22:%.*]] = bitcast float* [[REF_TMP]] to i8* 279 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP22]]) #[[ATTR1]] 280 // CHECK1-NEXT: store float 0.000000e+00, float* [[REF_TMP]], align 4, !tbaa [[TBAA14]] 281 // CHECK1-NEXT: [[TMP23:%.*]] = bitcast float* [[REF_TMP6]] to i8* 282 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP23]]) #[[ATTR1]] 283 // CHECK1-NEXT: store float 0.000000e+00, float* [[REF_TMP6]], align 4, !tbaa [[TBAA14]] 284 // CHECK1-NEXT: call void @_ZNSt7complexIfEC1ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM5]], float* nonnull align 4 dereferenceable(4) [[REF_TMP]], float* nonnull align 4 dereferenceable(4) [[REF_TMP6]]) #[[ATTR8]] 285 // CHECK1-NEXT: [[TMP24:%.*]] = bitcast float* [[REF_TMP6]] to i8* 286 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP24]]) #[[ATTR1]] 287 // CHECK1-NEXT: [[TMP25:%.*]] = bitcast float* [[REF_TMP]] to i8* 288 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP25]]) #[[ATTR1]] 289 // CHECK1-NEXT: [[TMP26:%.*]] = bitcast i32* [[I7]] to i8* 290 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP26]]) #[[ATTR1]] 291 // CHECK1-NEXT: [[TMP27:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 292 // CHECK1-NEXT: [[TMP28:%.*]] = load i32, i32* [[TMP27]], align 4, !tbaa [[TBAA8]] 293 // CHECK1-NEXT: call void @__kmpc_for_static_init_4u(%struct.ident_t* @[[GLOB3:[0-9]+]], i32 [[TMP28]], i32 33, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 294 // CHECK1-NEXT: br label [[OMP_DISPATCH_COND:%.*]] 295 // CHECK1: omp.dispatch.cond: 296 // CHECK1-NEXT: [[TMP29:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 297 // CHECK1-NEXT: [[TMP30:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 298 // CHECK1-NEXT: [[CMP8:%.*]] = icmp ugt i32 [[TMP29]], [[TMP30]] 299 // CHECK1-NEXT: br i1 [[CMP8]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 300 // CHECK1: cond.true: 301 // CHECK1-NEXT: [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 302 // CHECK1-NEXT: br label [[COND_END:%.*]] 303 // CHECK1: cond.false: 304 // CHECK1-NEXT: [[TMP32:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 305 // CHECK1-NEXT: br label [[COND_END]] 306 // CHECK1: cond.end: 307 // CHECK1-NEXT: [[COND:%.*]] = phi i32 [ [[TMP31]], [[COND_TRUE]] ], [ [[TMP32]], [[COND_FALSE]] ] 308 // CHECK1-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 309 // CHECK1-NEXT: [[TMP33:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 310 // CHECK1-NEXT: store i32 [[TMP33]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 311 // CHECK1-NEXT: [[TMP34:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 312 // CHECK1-NEXT: [[TMP35:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 313 // CHECK1-NEXT: [[ADD9:%.*]] = add i32 [[TMP35]], 1 314 // CHECK1-NEXT: [[CMP10:%.*]] = icmp ult i32 [[TMP34]], [[ADD9]] 315 // CHECK1-NEXT: br i1 [[CMP10]], label [[OMP_DISPATCH_BODY:%.*]], label [[OMP_DISPATCH_CLEANUP:%.*]] 316 // CHECK1: omp.dispatch.cleanup: 317 // CHECK1-NEXT: br label [[OMP_DISPATCH_END:%.*]] 318 // CHECK1: omp.dispatch.body: 319 // CHECK1-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 320 // CHECK1: omp.inner.for.cond: 321 // CHECK1-NEXT: [[TMP36:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 322 // CHECK1-NEXT: [[TMP37:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 323 // CHECK1-NEXT: [[ADD11:%.*]] = add i32 [[TMP37]], 1 324 // CHECK1-NEXT: [[CMP12:%.*]] = icmp ult i32 [[TMP36]], [[ADD11]] 325 // CHECK1-NEXT: br i1 [[CMP12]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 326 // CHECK1: omp.inner.for.cond.cleanup: 327 // CHECK1-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 328 // CHECK1: omp.inner.for.body: 329 // CHECK1-NEXT: [[TMP38:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 330 // CHECK1-NEXT: [[TMP39:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 331 // CHECK1-NEXT: [[MUL:%.*]] = mul i32 [[TMP39]], 1 332 // CHECK1-NEXT: [[ADD13:%.*]] = add i32 [[TMP38]], [[MUL]] 333 // CHECK1-NEXT: store i32 [[ADD13]], i32* [[I7]], align 4, !tbaa [[TBAA8]] 334 // CHECK1-NEXT: [[TMP40:%.*]] = bitcast %"class.std::complex"* [[REF_TMP14]] to i8* 335 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP40]]) #[[ATTR1]] 336 // CHECK1-NEXT: [[TMP41:%.*]] = bitcast float* [[REF_TMP15]] to i8* 337 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP41]]) #[[ATTR1]] 338 // CHECK1-NEXT: [[TMP42:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 339 // CHECK1-NEXT: [[CONV:%.*]] = sitofp i32 [[TMP42]] to float 340 // CHECK1-NEXT: store float [[CONV]], float* [[REF_TMP15]], align 4, !tbaa [[TBAA14]] 341 // CHECK1-NEXT: [[TMP43:%.*]] = bitcast float* [[REF_TMP16]] to i8* 342 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP43]]) #[[ATTR1]] 343 // CHECK1-NEXT: [[TMP44:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 344 // CHECK1-NEXT: [[CONV17:%.*]] = sitofp i32 [[TMP44]] to float 345 // CHECK1-NEXT: store float [[CONV17]], float* [[REF_TMP16]], align 4, !tbaa [[TBAA14]] 346 // CHECK1-NEXT: call void @_ZNSt7complexIfEC1ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[REF_TMP14]], float* nonnull align 4 dereferenceable(4) [[REF_TMP15]], float* nonnull align 4 dereferenceable(4) [[REF_TMP16]]) #[[ATTR8]] 347 // CHECK1-NEXT: [[CALL:%.*]] = call nonnull align 4 dereferenceable(8) %"class.std::complex"* @_ZNSt7complexIfEpLIfEERS0_RKS_IT_E(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM5]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[REF_TMP14]]) #[[ATTR8]] 348 // CHECK1-NEXT: [[TMP45:%.*]] = bitcast float* [[REF_TMP16]] to i8* 349 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP45]]) #[[ATTR1]] 350 // CHECK1-NEXT: [[TMP46:%.*]] = bitcast float* [[REF_TMP15]] to i8* 351 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP46]]) #[[ATTR1]] 352 // CHECK1-NEXT: [[TMP47:%.*]] = bitcast %"class.std::complex"* [[REF_TMP14]] to i8* 353 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP47]]) #[[ATTR1]] 354 // CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 355 // CHECK1: omp.body.continue: 356 // CHECK1-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 357 // CHECK1: omp.inner.for.inc: 358 // CHECK1-NEXT: [[TMP48:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 359 // CHECK1-NEXT: [[ADD18:%.*]] = add i32 [[TMP48]], 1 360 // CHECK1-NEXT: store i32 [[ADD18]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 361 // CHECK1-NEXT: br label [[OMP_INNER_FOR_COND]] 362 // CHECK1: omp.inner.for.end: 363 // CHECK1-NEXT: br label [[OMP_DISPATCH_INC:%.*]] 364 // CHECK1: omp.dispatch.inc: 365 // CHECK1-NEXT: [[TMP49:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 366 // CHECK1-NEXT: [[TMP50:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 367 // CHECK1-NEXT: [[ADD19:%.*]] = add i32 [[TMP49]], [[TMP50]] 368 // CHECK1-NEXT: store i32 [[ADD19]], i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 369 // CHECK1-NEXT: [[TMP51:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 370 // CHECK1-NEXT: [[TMP52:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 371 // CHECK1-NEXT: [[ADD20:%.*]] = add i32 [[TMP51]], [[TMP52]] 372 // CHECK1-NEXT: store i32 [[ADD20]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 373 // CHECK1-NEXT: br label [[OMP_DISPATCH_COND]] 374 // CHECK1: omp.dispatch.end: 375 // CHECK1-NEXT: [[TMP53:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 376 // CHECK1-NEXT: [[TMP54:%.*]] = load i32, i32* [[TMP53]], align 4, !tbaa [[TBAA8]] 377 // CHECK1-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB3]], i32 [[TMP54]]) 378 // CHECK1-NEXT: [[TMP55:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 379 // CHECK1-NEXT: [[TMP56:%.*]] = load i32, i32* [[TMP55]], align 4, !tbaa [[TBAA8]] 380 // CHECK1-NEXT: [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]], i64 0, i64 0 381 // CHECK1-NEXT: [[TMP58:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM5]] to i8* 382 // CHECK1-NEXT: store i8* [[TMP58]], i8** [[TMP57]], align 8 383 // CHECK1-NEXT: [[TMP59:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]] to i8* 384 // CHECK1-NEXT: [[TMP60:%.*]] = call i32 @__kmpc_nvptx_parallel_reduce_nowait_v2(%struct.ident_t* @[[GLOB1]], i32 [[TMP56]], i32 1, i64 8, i8* [[TMP59]], void (i8*, i16, i16, i16)* @_omp_reduction_shuffle_and_reduce_func, void (i8*, i32)* @_omp_reduction_inter_warp_copy_func) 385 // CHECK1-NEXT: [[TMP61:%.*]] = icmp eq i32 [[TMP60]], 1 386 // CHECK1-NEXT: br i1 [[TMP61]], label [[DOTOMP_REDUCTION_THEN:%.*]], label [[DOTOMP_REDUCTION_DONE:%.*]] 387 // CHECK1: .omp.reduction.then: 388 // CHECK1-NEXT: [[CALL21:%.*]] = call nonnull align 4 dereferenceable(8) %"class.std::complex"* @_ZNSt7complexIfEpLIfEERS0_RKS_IT_E(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[TMP2]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM5]]) #[[ATTR8]] 389 // CHECK1-NEXT: call void @__kmpc_nvptx_end_reduce_nowait(i32 [[TMP56]]) 390 // CHECK1-NEXT: br label [[DOTOMP_REDUCTION_DONE]] 391 // CHECK1: .omp.reduction.done: 392 // CHECK1-NEXT: [[TMP62:%.*]] = bitcast i32* [[I7]] to i8* 393 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP62]]) #[[ATTR1]] 394 // CHECK1-NEXT: [[TMP63:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM5]] to i8* 395 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP63]]) #[[ATTR1]] 396 // CHECK1-NEXT: [[TMP64:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 397 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP64]]) #[[ATTR1]] 398 // CHECK1-NEXT: [[TMP65:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 399 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP65]]) #[[ATTR1]] 400 // CHECK1-NEXT: [[TMP66:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 401 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP66]]) #[[ATTR1]] 402 // CHECK1-NEXT: [[TMP67:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 403 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP67]]) #[[ATTR1]] 404 // CHECK1-NEXT: br label [[OMP_PRECOND_END]] 405 // CHECK1: omp.precond.end: 406 // CHECK1-NEXT: [[TMP68:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 407 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP68]]) #[[ATTR1]] 408 // CHECK1-NEXT: [[TMP69:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 409 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP69]]) #[[ATTR1]] 410 // CHECK1-NEXT: [[TMP70:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 411 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP70]]) #[[ATTR1]] 412 // CHECK1-NEXT: [[TMP71:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 413 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP71]]) #[[ATTR1]] 414 // CHECK1-NEXT: ret void 415 // 416 // 417 // CHECK1-LABEL: define {{[^@]+}}@_ZNSt7complexIfEpLIfEERS0_RKS_IT_E 418 // CHECK1-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[__C:%.*]]) #[[ATTR4:[0-9]+]] comdat align 2 { 419 // CHECK1-NEXT: entry: 420 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 421 // CHECK1-NEXT: [[__C_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 422 // CHECK1-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 423 // CHECK1-NEXT: store %"class.std::complex"* [[__C]], %"class.std::complex"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 424 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 425 // CHECK1-NEXT: [[TMP0:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 426 // CHECK1-NEXT: [[CALL:%.*]] = call float @_ZNKSt7complexIfE4realEv(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[TMP0]]) #[[ATTR8]] 427 // CHECK1-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 0 428 // CHECK1-NEXT: [[TMP1:%.*]] = load float, float* [[__RE_]], align 4, !tbaa [[TBAA16:![0-9]+]] 429 // CHECK1-NEXT: [[ADD:%.*]] = fadd float [[TMP1]], [[CALL]] 430 // CHECK1-NEXT: store float [[ADD]], float* [[__RE_]], align 4, !tbaa [[TBAA16]] 431 // CHECK1-NEXT: [[TMP2:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 432 // CHECK1-NEXT: [[CALL2:%.*]] = call float @_ZNKSt7complexIfE4imagEv(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[TMP2]]) #[[ATTR8]] 433 // CHECK1-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 1 434 // CHECK1-NEXT: [[TMP3:%.*]] = load float, float* [[__IM_]], align 4, !tbaa [[TBAA18:![0-9]+]] 435 // CHECK1-NEXT: [[ADD3:%.*]] = fadd float [[TMP3]], [[CALL2]] 436 // CHECK1-NEXT: store float [[ADD3]], float* [[__IM_]], align 4, !tbaa [[TBAA18]] 437 // CHECK1-NEXT: ret %"class.std::complex"* [[THIS1]] 438 // 439 // 440 // CHECK1-LABEL: define {{[^@]+}}@_omp_reduction_shuffle_and_reduce_func 441 // CHECK1-SAME: (i8* [[TMP0:%.*]], i16 signext [[TMP1:%.*]], i16 signext [[TMP2:%.*]], i16 signext [[TMP3:%.*]]) #[[ATTR0]] { 442 // CHECK1-NEXT: entry: 443 // CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 444 // CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca i16, align 2 445 // CHECK1-NEXT: [[DOTADDR2:%.*]] = alloca i16, align 2 446 // CHECK1-NEXT: [[DOTADDR3:%.*]] = alloca i16, align 2 447 // CHECK1-NEXT: [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST:%.*]] = alloca [1 x i8*], align 8 448 // CHECK1-NEXT: [[DOTOMP_REDUCTION_ELEMENT:%.*]] = alloca %"class.std::complex", align 4 449 // CHECK1-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 450 // CHECK1-NEXT: store i16 [[TMP1]], i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19:![0-9]+]] 451 // CHECK1-NEXT: store i16 [[TMP2]], i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 452 // CHECK1-NEXT: store i16 [[TMP3]], i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 453 // CHECK1-NEXT: [[TMP4:%.*]] = load i8*, i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 454 // CHECK1-NEXT: [[TMP5:%.*]] = bitcast i8* [[TMP4]] to [1 x i8*]* 455 // CHECK1-NEXT: [[TMP6:%.*]] = load i16, i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19]] 456 // CHECK1-NEXT: [[TMP7:%.*]] = load i16, i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 457 // CHECK1-NEXT: [[TMP8:%.*]] = load i16, i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 458 // CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 459 // CHECK1-NEXT: [[TMP10:%.*]] = load i8*, i8** [[TMP9]], align 8 460 // CHECK1-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 461 // CHECK1-NEXT: [[TMP12:%.*]] = bitcast i8* [[TMP10]] to %"class.std::complex"* 462 // CHECK1-NEXT: [[TMP13:%.*]] = getelementptr %"class.std::complex", %"class.std::complex"* [[TMP12]], i64 1 463 // CHECK1-NEXT: [[TMP14:%.*]] = bitcast %"class.std::complex"* [[TMP13]] to i8* 464 // CHECK1-NEXT: [[TMP15:%.*]] = bitcast %"class.std::complex"* [[TMP12]] to i64* 465 // CHECK1-NEXT: [[TMP16:%.*]] = bitcast %"class.std::complex"* [[DOTOMP_REDUCTION_ELEMENT]] to i64* 466 // CHECK1-NEXT: [[TMP17:%.*]] = load i64, i64* [[TMP15]], align 4 467 // CHECK1-NEXT: [[NVPTX_WARP_SIZE:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize() 468 // CHECK1-NEXT: [[TMP18:%.*]] = trunc i32 [[NVPTX_WARP_SIZE]] to i16 469 // CHECK1-NEXT: [[TMP19:%.*]] = call i64 @__kmpc_shuffle_int64(i64 [[TMP17]], i16 [[TMP7]], i16 [[TMP18]]) 470 // CHECK1-NEXT: store i64 [[TMP19]], i64* [[TMP16]], align 4 471 // CHECK1-NEXT: [[TMP20:%.*]] = getelementptr i64, i64* [[TMP15]], i64 1 472 // CHECK1-NEXT: [[TMP21:%.*]] = getelementptr i64, i64* [[TMP16]], i64 1 473 // CHECK1-NEXT: [[TMP22:%.*]] = bitcast %"class.std::complex"* [[DOTOMP_REDUCTION_ELEMENT]] to i8* 474 // CHECK1-NEXT: store i8* [[TMP22]], i8** [[TMP11]], align 8, !tbaa [[TBAA12]] 475 // CHECK1-NEXT: [[TMP23:%.*]] = icmp eq i16 [[TMP8]], 0 476 // CHECK1-NEXT: [[TMP24:%.*]] = icmp eq i16 [[TMP8]], 1 477 // CHECK1-NEXT: [[TMP25:%.*]] = icmp ult i16 [[TMP6]], [[TMP7]] 478 // CHECK1-NEXT: [[TMP26:%.*]] = and i1 [[TMP24]], [[TMP25]] 479 // CHECK1-NEXT: [[TMP27:%.*]] = icmp eq i16 [[TMP8]], 2 480 // CHECK1-NEXT: [[TMP28:%.*]] = and i16 [[TMP6]], 1 481 // CHECK1-NEXT: [[TMP29:%.*]] = icmp eq i16 [[TMP28]], 0 482 // CHECK1-NEXT: [[TMP30:%.*]] = and i1 [[TMP27]], [[TMP29]] 483 // CHECK1-NEXT: [[TMP31:%.*]] = icmp sgt i16 [[TMP7]], 0 484 // CHECK1-NEXT: [[TMP32:%.*]] = and i1 [[TMP30]], [[TMP31]] 485 // CHECK1-NEXT: [[TMP33:%.*]] = or i1 [[TMP23]], [[TMP26]] 486 // CHECK1-NEXT: [[TMP34:%.*]] = or i1 [[TMP33]], [[TMP32]] 487 // CHECK1-NEXT: br i1 [[TMP34]], label [[THEN:%.*]], label [[ELSE:%.*]] 488 // CHECK1: then: 489 // CHECK1-NEXT: [[TMP35:%.*]] = bitcast [1 x i8*]* [[TMP5]] to i8* 490 // CHECK1-NEXT: [[TMP36:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]] to i8* 491 // CHECK1-NEXT: call void @"_omp$reduction$reduction_func"(i8* [[TMP35]], i8* [[TMP36]]) #[[ATTR1]] 492 // CHECK1-NEXT: br label [[IFCONT:%.*]] 493 // CHECK1: else: 494 // CHECK1-NEXT: br label [[IFCONT]] 495 // CHECK1: ifcont: 496 // CHECK1-NEXT: [[TMP37:%.*]] = icmp eq i16 [[TMP8]], 1 497 // CHECK1-NEXT: [[TMP38:%.*]] = icmp uge i16 [[TMP6]], [[TMP7]] 498 // CHECK1-NEXT: [[TMP39:%.*]] = and i1 [[TMP37]], [[TMP38]] 499 // CHECK1-NEXT: br i1 [[TMP39]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 500 // CHECK1: then4: 501 // CHECK1-NEXT: [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 502 // CHECK1-NEXT: [[TMP41:%.*]] = load i8*, i8** [[TMP40]], align 8 503 // CHECK1-NEXT: [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 504 // CHECK1-NEXT: [[TMP43:%.*]] = load i8*, i8** [[TMP42]], align 8 505 // CHECK1-NEXT: [[TMP44:%.*]] = bitcast i8* [[TMP41]] to %"class.std::complex"* 506 // CHECK1-NEXT: [[TMP45:%.*]] = bitcast i8* [[TMP43]] to %"class.std::complex"* 507 // CHECK1-NEXT: [[TMP46:%.*]] = bitcast %"class.std::complex"* [[TMP45]] to i8* 508 // CHECK1-NEXT: [[TMP47:%.*]] = bitcast %"class.std::complex"* [[TMP44]] to i8* 509 // CHECK1-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP46]], i8* align 4 [[TMP47]], i64 8, i1 false), !tbaa.struct !21 510 // CHECK1-NEXT: br label [[IFCONT6:%.*]] 511 // CHECK1: else5: 512 // CHECK1-NEXT: br label [[IFCONT6]] 513 // CHECK1: ifcont6: 514 // CHECK1-NEXT: ret void 515 // 516 // 517 // CHECK1-LABEL: define {{[^@]+}}@_omp_reduction_inter_warp_copy_func 518 // CHECK1-SAME: (i8* [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 519 // CHECK1-NEXT: entry: 520 // CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 521 // CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 522 // CHECK1-NEXT: [[DOTCNT_ADDR:%.*]] = alloca i32, align 4 523 // CHECK1-NEXT: [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 524 // CHECK1-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 525 // CHECK1-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 526 // CHECK1-NEXT: [[NVPTX_TID:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 527 // CHECK1-NEXT: [[NVPTX_TID2:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 528 // CHECK1-NEXT: [[NVPTX_LANE_ID:%.*]] = and i32 [[NVPTX_TID2]], 31 529 // CHECK1-NEXT: [[NVPTX_TID3:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 530 // CHECK1-NEXT: [[NVPTX_WARP_ID:%.*]] = ashr i32 [[NVPTX_TID3]], 5 531 // CHECK1-NEXT: [[TMP3:%.*]] = load i8*, i8** [[DOTADDR]], align 8 532 // CHECK1-NEXT: [[TMP4:%.*]] = bitcast i8* [[TMP3]] to [1 x i8*]* 533 // CHECK1-NEXT: store i32 0, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 534 // CHECK1-NEXT: br label [[PRECOND:%.*]] 535 // CHECK1: precond: 536 // CHECK1-NEXT: [[TMP5:%.*]] = load i32, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 537 // CHECK1-NEXT: [[TMP6:%.*]] = icmp ult i32 [[TMP5]], 2 538 // CHECK1-NEXT: br i1 [[TMP6]], label [[BODY:%.*]], label [[EXIT:%.*]] 539 // CHECK1: body: 540 // CHECK1-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4:[0-9]+]], i32 [[TMP2]]) 541 // CHECK1-NEXT: [[WARP_MASTER:%.*]] = icmp eq i32 [[NVPTX_LANE_ID]], 0 542 // CHECK1-NEXT: br i1 [[WARP_MASTER]], label [[THEN:%.*]], label [[ELSE:%.*]] 543 // CHECK1: then: 544 // CHECK1-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 545 // CHECK1-NEXT: [[TMP8:%.*]] = load i8*, i8** [[TMP7]], align 8, !tbaa [[TBAA12]] 546 // CHECK1-NEXT: [[TMP9:%.*]] = bitcast i8* [[TMP8]] to i32* 547 // CHECK1-NEXT: [[TMP10:%.*]] = getelementptr i32, i32* [[TMP9]], i32 [[TMP5]] 548 // CHECK1-NEXT: [[TMP11:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_WARP_ID]] 549 // CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP10]], align 4 550 // CHECK1-NEXT: store volatile i32 [[TMP12]], i32 addrspace(3)* [[TMP11]], align 4 551 // CHECK1-NEXT: br label [[IFCONT:%.*]] 552 // CHECK1: else: 553 // CHECK1-NEXT: br label [[IFCONT]] 554 // CHECK1: ifcont: 555 // CHECK1-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4]], i32 [[TMP2]]) 556 // CHECK1-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 557 // CHECK1-NEXT: [[IS_ACTIVE_THREAD:%.*]] = icmp ult i32 [[NVPTX_TID]], [[TMP13]] 558 // CHECK1-NEXT: br i1 [[IS_ACTIVE_THREAD]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 559 // CHECK1: then4: 560 // CHECK1-NEXT: [[TMP14:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_TID]] 561 // CHECK1-NEXT: [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 562 // CHECK1-NEXT: [[TMP16:%.*]] = load i8*, i8** [[TMP15]], align 8, !tbaa [[TBAA12]] 563 // CHECK1-NEXT: [[TMP17:%.*]] = bitcast i8* [[TMP16]] to i32* 564 // CHECK1-NEXT: [[TMP18:%.*]] = getelementptr i32, i32* [[TMP17]], i32 [[TMP5]] 565 // CHECK1-NEXT: [[TMP19:%.*]] = load volatile i32, i32 addrspace(3)* [[TMP14]], align 4, !tbaa [[TBAA8]] 566 // CHECK1-NEXT: store i32 [[TMP19]], i32* [[TMP18]], align 4, !tbaa [[TBAA8]] 567 // CHECK1-NEXT: br label [[IFCONT6:%.*]] 568 // CHECK1: else5: 569 // CHECK1-NEXT: br label [[IFCONT6]] 570 // CHECK1: ifcont6: 571 // CHECK1-NEXT: [[TMP20:%.*]] = add nsw i32 [[TMP5]], 1 572 // CHECK1-NEXT: store i32 [[TMP20]], i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 573 // CHECK1-NEXT: br label [[PRECOND]] 574 // CHECK1: exit: 575 // CHECK1-NEXT: ret void 576 // 577 // 578 // CHECK1-LABEL: define {{[^@]+}}@__omp_outlined__1_wrapper 579 // CHECK1-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 580 // CHECK1-NEXT: entry: 581 // CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i16, align 2 582 // CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 583 // CHECK1-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 584 // CHECK1-NEXT: [[GLOBAL_ARGS:%.*]] = alloca i8**, align 8 585 // CHECK1-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 586 // CHECK1-NEXT: store i16 [[TMP0]], i16* [[DOTADDR]], align 2, !tbaa [[TBAA19]] 587 // CHECK1-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 588 // CHECK1-NEXT: call void @__kmpc_get_shared_variables(i8*** [[GLOBAL_ARGS]]) 589 // CHECK1-NEXT: [[TMP2:%.*]] = load i8**, i8*** [[GLOBAL_ARGS]], align 8 590 // CHECK1-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 0 591 // CHECK1-NEXT: [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32** 592 // CHECK1-NEXT: [[TMP5:%.*]] = load i32*, i32** [[TMP4]], align 8, !tbaa [[TBAA12]] 593 // CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 1 594 // CHECK1-NEXT: [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i32** 595 // CHECK1-NEXT: [[TMP8:%.*]] = load i32*, i32** [[TMP7]], align 8, !tbaa [[TBAA12]] 596 // CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 2 597 // CHECK1-NEXT: [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %"class.std::complex"** 598 // CHECK1-NEXT: [[TMP11:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[TMP10]], align 8, !tbaa [[TBAA12]] 599 // CHECK1-NEXT: call void @__omp_outlined__1(i32* [[DOTADDR1]], i32* [[DOTZERO_ADDR]], i32* [[TMP5]], i32* [[TMP8]], %"class.std::complex"* [[TMP11]]) #[[ATTR1]] 600 // CHECK1-NEXT: ret void 601 // 602 // 603 // CHECK1-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z17complex_reductionIdEvv_l19 604 // CHECK1-SAME: () #[[ATTR0]] { 605 // CHECK1-NEXT: entry: 606 // CHECK1-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 607 // CHECK1-NEXT: [[DOTTHREADID_TEMP_:%.*]] = alloca i32, align 4 608 // CHECK1-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 609 // CHECK1-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* @[[GLOB1]], i1 false, i1 true, i1 true) 610 // CHECK1-NEXT: [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1 611 // CHECK1-NEXT: br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]] 612 // CHECK1: user_code.entry: 613 // CHECK1-NEXT: [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 614 // CHECK1-NEXT: store i32 [[TMP1]], i32* [[DOTTHREADID_TEMP_]], align 4, !tbaa [[TBAA8]] 615 // CHECK1-NEXT: call void @__omp_outlined__2(i32* [[DOTTHREADID_TEMP_]], i32* [[DOTZERO_ADDR]]) #[[ATTR1]] 616 // CHECK1-NEXT: call void @__kmpc_target_deinit(%struct.ident_t* @[[GLOB1]], i1 false, i1 true) 617 // CHECK1-NEXT: ret void 618 // CHECK1: worker.exit: 619 // CHECK1-NEXT: ret void 620 // 621 // 622 // CHECK1-LABEL: define {{[^@]+}}@__omp_outlined__2 623 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR0]] { 624 // CHECK1-NEXT: entry: 625 // CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 626 // CHECK1-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 627 // CHECK1-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 628 // CHECK1-NEXT: [[TMP:%.*]] = alloca i32, align 4 629 // CHECK1-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 630 // CHECK1-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 631 // CHECK1-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 632 // CHECK1-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 633 // CHECK1-NEXT: [[IB:%.*]] = alloca i32, align 4 634 // CHECK1-NEXT: [[REF_TMP:%.*]] = alloca double, align 8 635 // CHECK1-NEXT: [[REF_TMP2:%.*]] = alloca double, align 8 636 // CHECK1-NEXT: [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x i8*], align 8 637 // CHECK1-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12]] 638 // CHECK1-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 639 // CHECK1-NEXT: [[ISTART:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 640 // CHECK1-NEXT: [[ISTART_ON_STACK:%.*]] = bitcast i8* [[ISTART]] to i32* 641 // CHECK1-NEXT: [[IEND:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 642 // CHECK1-NEXT: [[IEND_ON_STACK:%.*]] = bitcast i8* [[IEND]] to i32* 643 // CHECK1-NEXT: [[PARTIAL_SUM:%.*]] = call i8* @__kmpc_alloc_shared(i64 16) 644 // CHECK1-NEXT: [[PARTIAL_SUM_ON_STACK:%.*]] = bitcast i8* [[PARTIAL_SUM]] to %"class.std::complex.0"* 645 // CHECK1-NEXT: [[TMP0:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 646 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP0]]) #[[ATTR1]] 647 // CHECK1-NEXT: [[TMP1:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 648 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP1]]) #[[ATTR1]] 649 // CHECK1-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 650 // CHECK1-NEXT: [[TMP2:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 651 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP2]]) #[[ATTR1]] 652 // CHECK1-NEXT: store i32 99, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 653 // CHECK1-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 654 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 655 // CHECK1-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 656 // CHECK1-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 657 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 658 // CHECK1-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 659 // CHECK1-NEXT: [[TMP5:%.*]] = bitcast i32* [[IB]] to i8* 660 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP5]]) #[[ATTR1]] 661 // CHECK1-NEXT: [[TMP6:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 662 // CHECK1-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP6]], align 4, !tbaa [[TBAA8]] 663 // CHECK1-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP7]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 664 // CHECK1-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 665 // CHECK1-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP8]], 99 666 // CHECK1-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 667 // CHECK1: cond.true: 668 // CHECK1-NEXT: br label [[COND_END:%.*]] 669 // CHECK1: cond.false: 670 // CHECK1-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 671 // CHECK1-NEXT: br label [[COND_END]] 672 // CHECK1: cond.end: 673 // CHECK1-NEXT: [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP9]], [[COND_FALSE]] ] 674 // CHECK1-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 675 // CHECK1-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 676 // CHECK1-NEXT: store i32 [[TMP10]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 677 // CHECK1-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 678 // CHECK1: omp.inner.for.cond: 679 // CHECK1-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 680 // CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 681 // CHECK1-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP11]], [[TMP12]] 682 // CHECK1-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 683 // CHECK1: omp.inner.for.cond.cleanup: 684 // CHECK1-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 685 // CHECK1: omp.inner.for.body: 686 // CHECK1-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 687 // CHECK1-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP13]], 1 688 // CHECK1-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]] 689 // CHECK1-NEXT: store i32 [[ADD]], i32* [[IB]], align 4, !tbaa [[TBAA8]] 690 // CHECK1-NEXT: [[TMP14:%.*]] = bitcast double* [[REF_TMP]] to i8* 691 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP14]]) #[[ATTR1]] 692 // CHECK1-NEXT: store double 0.000000e+00, double* [[REF_TMP]], align 8, !tbaa [[TBAA22:![0-9]+]] 693 // CHECK1-NEXT: [[TMP15:%.*]] = bitcast double* [[REF_TMP2]] to i8* 694 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP15]]) #[[ATTR1]] 695 // CHECK1-NEXT: store double 0.000000e+00, double* [[REF_TMP2]], align 8, !tbaa [[TBAA22]] 696 // CHECK1-NEXT: call void @_ZNSt7complexIdEC1ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM_ON_STACK]], double* nonnull align 8 dereferenceable(8) [[REF_TMP]], double* nonnull align 8 dereferenceable(8) [[REF_TMP2]]) #[[ATTR8]] 697 // CHECK1-NEXT: [[TMP16:%.*]] = bitcast double* [[REF_TMP2]] to i8* 698 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP16]]) #[[ATTR1]] 699 // CHECK1-NEXT: [[TMP17:%.*]] = bitcast double* [[REF_TMP]] to i8* 700 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP17]]) #[[ATTR1]] 701 // CHECK1-NEXT: [[TMP18:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 702 // CHECK1-NEXT: [[MUL3:%.*]] = mul nsw i32 [[TMP18]], 4 703 // CHECK1-NEXT: store i32 [[MUL3]], i32* [[ISTART_ON_STACK]], align 4, !tbaa [[TBAA8]] 704 // CHECK1-NEXT: [[TMP19:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 705 // CHECK1-NEXT: [[ADD4:%.*]] = add nsw i32 [[TMP19]], 1 706 // CHECK1-NEXT: [[MUL5:%.*]] = mul nsw i32 [[ADD4]], 4 707 // CHECK1-NEXT: store i32 [[MUL5]], i32* [[IEND_ON_STACK]], align 4, !tbaa [[TBAA8]] 708 // CHECK1-NEXT: [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 0 709 // CHECK1-NEXT: [[TMP21:%.*]] = bitcast i32* [[ISTART_ON_STACK]] to i8* 710 // CHECK1-NEXT: store i8* [[TMP21]], i8** [[TMP20]], align 8, !tbaa [[TBAA12]] 711 // CHECK1-NEXT: [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 1 712 // CHECK1-NEXT: [[TMP23:%.*]] = bitcast i32* [[IEND_ON_STACK]] to i8* 713 // CHECK1-NEXT: store i8* [[TMP23]], i8** [[TMP22]], align 8, !tbaa [[TBAA12]] 714 // CHECK1-NEXT: [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 2 715 // CHECK1-NEXT: [[TMP25:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM_ON_STACK]] to i8* 716 // CHECK1-NEXT: store i8* [[TMP25]], i8** [[TMP24]], align 8, !tbaa [[TBAA12]] 717 // CHECK1-NEXT: [[TMP26:%.*]] = bitcast [3 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8** 718 // CHECK1-NEXT: call void @__kmpc_parallel_51(%struct.ident_t* @[[GLOB1]], i32 [[TMP7]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*, i32*, i32*, %"class.std::complex.0"*)* @__omp_outlined__3 to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined__3_wrapper to i8*), i8** [[TMP26]], i64 3) 719 // CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 720 // CHECK1: omp.body.continue: 721 // CHECK1-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 722 // CHECK1: omp.inner.for.inc: 723 // CHECK1-NEXT: [[TMP27:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 724 // CHECK1-NEXT: [[ADD6:%.*]] = add nsw i32 [[TMP27]], 1 725 // CHECK1-NEXT: store i32 [[ADD6]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 726 // CHECK1-NEXT: br label [[OMP_INNER_FOR_COND]] 727 // CHECK1: omp.inner.for.end: 728 // CHECK1-NEXT: br label [[OMP_LOOP_EXIT:%.*]] 729 // CHECK1: omp.loop.exit: 730 // CHECK1-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB2]], i32 [[TMP7]]) 731 // CHECK1-NEXT: [[TMP28:%.*]] = bitcast i32* [[IB]] to i8* 732 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP28]]) #[[ATTR1]] 733 // CHECK1-NEXT: [[TMP29:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 734 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP29]]) #[[ATTR1]] 735 // CHECK1-NEXT: [[TMP30:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 736 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP30]]) #[[ATTR1]] 737 // CHECK1-NEXT: [[TMP31:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 738 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP31]]) #[[ATTR1]] 739 // CHECK1-NEXT: [[TMP32:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 740 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP32]]) #[[ATTR1]] 741 // CHECK1-NEXT: [[TMP33:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 742 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP33]]) #[[ATTR1]] 743 // CHECK1-NEXT: call void @__kmpc_free_shared(i8* [[PARTIAL_SUM]], i64 16) 744 // CHECK1-NEXT: call void @__kmpc_free_shared(i8* [[IEND]], i64 4) 745 // CHECK1-NEXT: call void @__kmpc_free_shared(i8* [[ISTART]], i64 4) 746 // CHECK1-NEXT: ret void 747 // 748 // 749 // CHECK1-LABEL: define {{[^@]+}}@_ZNSt7complexIdEC1ERKdS2_ 750 // CHECK1-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]], double* nonnull align 8 dereferenceable(8) [[__RE:%.*]], double* nonnull align 8 dereferenceable(8) [[__IM:%.*]]) unnamed_addr #[[ATTR3]] comdat align 2 { 751 // CHECK1-NEXT: entry: 752 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 753 // CHECK1-NEXT: [[__RE_ADDR:%.*]] = alloca double*, align 8 754 // CHECK1-NEXT: [[__IM_ADDR:%.*]] = alloca double*, align 8 755 // CHECK1-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 756 // CHECK1-NEXT: store double* [[__RE]], double** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 757 // CHECK1-NEXT: store double* [[__IM]], double** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 758 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 759 // CHECK1-NEXT: [[TMP0:%.*]] = load double*, double** [[__RE_ADDR]], align 8 760 // CHECK1-NEXT: [[TMP1:%.*]] = load double*, double** [[__IM_ADDR]], align 8 761 // CHECK1-NEXT: call void @_ZNSt7complexIdEC2ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS1]], double* nonnull align 8 dereferenceable(8) [[TMP0]], double* nonnull align 8 dereferenceable(8) [[TMP1]]) #[[ATTR8]] 762 // CHECK1-NEXT: ret void 763 // 764 // 765 // CHECK1-LABEL: define {{[^@]+}}@__omp_outlined__3 766 // CHECK1-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ISTART:%.*]], i32* nonnull align 4 dereferenceable(4) [[IEND:%.*]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM:%.*]]) #[[ATTR0]] { 767 // CHECK1-NEXT: entry: 768 // CHECK1-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 769 // CHECK1-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 770 // CHECK1-NEXT: [[ISTART_ADDR:%.*]] = alloca i32*, align 8 771 // CHECK1-NEXT: [[IEND_ADDR:%.*]] = alloca i32*, align 8 772 // CHECK1-NEXT: [[PARTIAL_SUM_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 773 // CHECK1-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 774 // CHECK1-NEXT: [[TMP:%.*]] = alloca i32, align 4 775 // CHECK1-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4 776 // CHECK1-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4 777 // CHECK1-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4 778 // CHECK1-NEXT: [[I:%.*]] = alloca i32, align 4 779 // CHECK1-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 780 // CHECK1-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 781 // CHECK1-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 782 // CHECK1-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 783 // CHECK1-NEXT: [[PARTIAL_SUM5:%.*]] = alloca %"class.std::complex.0", align 8 784 // CHECK1-NEXT: [[REF_TMP:%.*]] = alloca double, align 8 785 // CHECK1-NEXT: [[REF_TMP6:%.*]] = alloca double, align 8 786 // CHECK1-NEXT: [[I7:%.*]] = alloca i32, align 4 787 // CHECK1-NEXT: [[REF_TMP14:%.*]] = alloca %"class.std::complex.0", align 8 788 // CHECK1-NEXT: [[REF_TMP15:%.*]] = alloca double, align 8 789 // CHECK1-NEXT: [[REF_TMP16:%.*]] = alloca double, align 8 790 // CHECK1-NEXT: [[DOTOMP_REDUCTION_RED_LIST:%.*]] = alloca [1 x i8*], align 8 791 // CHECK1-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12]] 792 // CHECK1-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 793 // CHECK1-NEXT: store i32* [[ISTART]], i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 794 // CHECK1-NEXT: store i32* [[IEND]], i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 795 // CHECK1-NEXT: store %"class.std::complex.0"* [[PARTIAL_SUM]], %"class.std::complex.0"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 796 // CHECK1-NEXT: [[TMP0:%.*]] = load i32*, i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 797 // CHECK1-NEXT: [[TMP1:%.*]] = load i32*, i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 798 // CHECK1-NEXT: [[TMP2:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 799 // CHECK1-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 800 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 801 // CHECK1-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 802 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 803 // CHECK1-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP0]], align 4, !tbaa [[TBAA8]] 804 // CHECK1-NEXT: store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 805 // CHECK1-NEXT: [[TMP6:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 806 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP6]]) #[[ATTR1]] 807 // CHECK1-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP1]], align 4, !tbaa [[TBAA8]] 808 // CHECK1-NEXT: store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 809 // CHECK1-NEXT: [[TMP8:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 810 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP8]]) #[[ATTR1]] 811 // CHECK1-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 812 // CHECK1-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 813 // CHECK1-NEXT: [[SUB:%.*]] = sub i32 [[TMP9]], [[TMP10]] 814 // CHECK1-NEXT: [[SUB3:%.*]] = sub i32 [[SUB]], 1 815 // CHECK1-NEXT: [[ADD:%.*]] = add i32 [[SUB3]], 1 816 // CHECK1-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], 1 817 // CHECK1-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1 818 // CHECK1-NEXT: store i32 [[SUB4]], i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 819 // CHECK1-NEXT: [[TMP11:%.*]] = bitcast i32* [[I]] to i8* 820 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP11]]) #[[ATTR1]] 821 // CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 822 // CHECK1-NEXT: store i32 [[TMP12]], i32* [[I]], align 4, !tbaa [[TBAA8]] 823 // CHECK1-NEXT: [[TMP13:%.*]] = bitcast i32* [[I]] to i8* 824 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP13]]) #[[ATTR1]] 825 // CHECK1-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 826 // CHECK1-NEXT: [[TMP15:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 827 // CHECK1-NEXT: [[CMP:%.*]] = icmp slt i32 [[TMP14]], [[TMP15]] 828 // CHECK1-NEXT: br i1 [[CMP]], label [[OMP_PRECOND_THEN:%.*]], label [[OMP_PRECOND_END:%.*]] 829 // CHECK1: omp.precond.then: 830 // CHECK1-NEXT: [[TMP16:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 831 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP16]]) #[[ATTR1]] 832 // CHECK1-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 833 // CHECK1-NEXT: [[TMP17:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 834 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP17]]) #[[ATTR1]] 835 // CHECK1-NEXT: [[TMP18:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 836 // CHECK1-NEXT: store i32 [[TMP18]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 837 // CHECK1-NEXT: [[TMP19:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 838 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP19]]) #[[ATTR1]] 839 // CHECK1-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 840 // CHECK1-NEXT: [[TMP20:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 841 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP20]]) #[[ATTR1]] 842 // CHECK1-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 843 // CHECK1-NEXT: [[TMP21:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM5]] to i8* 844 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[TMP21]]) #[[ATTR1]] 845 // CHECK1-NEXT: [[TMP22:%.*]] = bitcast double* [[REF_TMP]] to i8* 846 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP22]]) #[[ATTR1]] 847 // CHECK1-NEXT: store double 0.000000e+00, double* [[REF_TMP]], align 8, !tbaa [[TBAA22]] 848 // CHECK1-NEXT: [[TMP23:%.*]] = bitcast double* [[REF_TMP6]] to i8* 849 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP23]]) #[[ATTR1]] 850 // CHECK1-NEXT: store double 0.000000e+00, double* [[REF_TMP6]], align 8, !tbaa [[TBAA22]] 851 // CHECK1-NEXT: call void @_ZNSt7complexIdEC1ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM5]], double* nonnull align 8 dereferenceable(8) [[REF_TMP]], double* nonnull align 8 dereferenceable(8) [[REF_TMP6]]) #[[ATTR8]] 852 // CHECK1-NEXT: [[TMP24:%.*]] = bitcast double* [[REF_TMP6]] to i8* 853 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP24]]) #[[ATTR1]] 854 // CHECK1-NEXT: [[TMP25:%.*]] = bitcast double* [[REF_TMP]] to i8* 855 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP25]]) #[[ATTR1]] 856 // CHECK1-NEXT: [[TMP26:%.*]] = bitcast i32* [[I7]] to i8* 857 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP26]]) #[[ATTR1]] 858 // CHECK1-NEXT: [[TMP27:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 859 // CHECK1-NEXT: [[TMP28:%.*]] = load i32, i32* [[TMP27]], align 4, !tbaa [[TBAA8]] 860 // CHECK1-NEXT: call void @__kmpc_for_static_init_4u(%struct.ident_t* @[[GLOB3]], i32 [[TMP28]], i32 33, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 861 // CHECK1-NEXT: br label [[OMP_DISPATCH_COND:%.*]] 862 // CHECK1: omp.dispatch.cond: 863 // CHECK1-NEXT: [[TMP29:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 864 // CHECK1-NEXT: [[TMP30:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 865 // CHECK1-NEXT: [[CMP8:%.*]] = icmp ugt i32 [[TMP29]], [[TMP30]] 866 // CHECK1-NEXT: br i1 [[CMP8]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 867 // CHECK1: cond.true: 868 // CHECK1-NEXT: [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 869 // CHECK1-NEXT: br label [[COND_END:%.*]] 870 // CHECK1: cond.false: 871 // CHECK1-NEXT: [[TMP32:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 872 // CHECK1-NEXT: br label [[COND_END]] 873 // CHECK1: cond.end: 874 // CHECK1-NEXT: [[COND:%.*]] = phi i32 [ [[TMP31]], [[COND_TRUE]] ], [ [[TMP32]], [[COND_FALSE]] ] 875 // CHECK1-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 876 // CHECK1-NEXT: [[TMP33:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 877 // CHECK1-NEXT: store i32 [[TMP33]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 878 // CHECK1-NEXT: [[TMP34:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 879 // CHECK1-NEXT: [[TMP35:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 880 // CHECK1-NEXT: [[ADD9:%.*]] = add i32 [[TMP35]], 1 881 // CHECK1-NEXT: [[CMP10:%.*]] = icmp ult i32 [[TMP34]], [[ADD9]] 882 // CHECK1-NEXT: br i1 [[CMP10]], label [[OMP_DISPATCH_BODY:%.*]], label [[OMP_DISPATCH_CLEANUP:%.*]] 883 // CHECK1: omp.dispatch.cleanup: 884 // CHECK1-NEXT: br label [[OMP_DISPATCH_END:%.*]] 885 // CHECK1: omp.dispatch.body: 886 // CHECK1-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 887 // CHECK1: omp.inner.for.cond: 888 // CHECK1-NEXT: [[TMP36:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 889 // CHECK1-NEXT: [[TMP37:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 890 // CHECK1-NEXT: [[ADD11:%.*]] = add i32 [[TMP37]], 1 891 // CHECK1-NEXT: [[CMP12:%.*]] = icmp ult i32 [[TMP36]], [[ADD11]] 892 // CHECK1-NEXT: br i1 [[CMP12]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 893 // CHECK1: omp.inner.for.cond.cleanup: 894 // CHECK1-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 895 // CHECK1: omp.inner.for.body: 896 // CHECK1-NEXT: [[TMP38:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 897 // CHECK1-NEXT: [[TMP39:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 898 // CHECK1-NEXT: [[MUL:%.*]] = mul i32 [[TMP39]], 1 899 // CHECK1-NEXT: [[ADD13:%.*]] = add i32 [[TMP38]], [[MUL]] 900 // CHECK1-NEXT: store i32 [[ADD13]], i32* [[I7]], align 4, !tbaa [[TBAA8]] 901 // CHECK1-NEXT: [[TMP40:%.*]] = bitcast %"class.std::complex.0"* [[REF_TMP14]] to i8* 902 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[TMP40]]) #[[ATTR1]] 903 // CHECK1-NEXT: [[TMP41:%.*]] = bitcast double* [[REF_TMP15]] to i8* 904 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP41]]) #[[ATTR1]] 905 // CHECK1-NEXT: [[TMP42:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 906 // CHECK1-NEXT: [[CONV:%.*]] = sitofp i32 [[TMP42]] to double 907 // CHECK1-NEXT: store double [[CONV]], double* [[REF_TMP15]], align 8, !tbaa [[TBAA22]] 908 // CHECK1-NEXT: [[TMP43:%.*]] = bitcast double* [[REF_TMP16]] to i8* 909 // CHECK1-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP43]]) #[[ATTR1]] 910 // CHECK1-NEXT: [[TMP44:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 911 // CHECK1-NEXT: [[CONV17:%.*]] = sitofp i32 [[TMP44]] to double 912 // CHECK1-NEXT: store double [[CONV17]], double* [[REF_TMP16]], align 8, !tbaa [[TBAA22]] 913 // CHECK1-NEXT: call void @_ZNSt7complexIdEC1ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[REF_TMP14]], double* nonnull align 8 dereferenceable(8) [[REF_TMP15]], double* nonnull align 8 dereferenceable(8) [[REF_TMP16]]) #[[ATTR8]] 914 // CHECK1-NEXT: [[CALL:%.*]] = call nonnull align 8 dereferenceable(16) %"class.std::complex.0"* @_ZNSt7complexIdEpLIdEERS0_RKS_IT_E(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM5]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[REF_TMP14]]) #[[ATTR8]] 915 // CHECK1-NEXT: [[TMP45:%.*]] = bitcast double* [[REF_TMP16]] to i8* 916 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP45]]) #[[ATTR1]] 917 // CHECK1-NEXT: [[TMP46:%.*]] = bitcast double* [[REF_TMP15]] to i8* 918 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP46]]) #[[ATTR1]] 919 // CHECK1-NEXT: [[TMP47:%.*]] = bitcast %"class.std::complex.0"* [[REF_TMP14]] to i8* 920 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[TMP47]]) #[[ATTR1]] 921 // CHECK1-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 922 // CHECK1: omp.body.continue: 923 // CHECK1-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 924 // CHECK1: omp.inner.for.inc: 925 // CHECK1-NEXT: [[TMP48:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 926 // CHECK1-NEXT: [[ADD18:%.*]] = add i32 [[TMP48]], 1 927 // CHECK1-NEXT: store i32 [[ADD18]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 928 // CHECK1-NEXT: br label [[OMP_INNER_FOR_COND]] 929 // CHECK1: omp.inner.for.end: 930 // CHECK1-NEXT: br label [[OMP_DISPATCH_INC:%.*]] 931 // CHECK1: omp.dispatch.inc: 932 // CHECK1-NEXT: [[TMP49:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 933 // CHECK1-NEXT: [[TMP50:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 934 // CHECK1-NEXT: [[ADD19:%.*]] = add i32 [[TMP49]], [[TMP50]] 935 // CHECK1-NEXT: store i32 [[ADD19]], i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 936 // CHECK1-NEXT: [[TMP51:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 937 // CHECK1-NEXT: [[TMP52:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 938 // CHECK1-NEXT: [[ADD20:%.*]] = add i32 [[TMP51]], [[TMP52]] 939 // CHECK1-NEXT: store i32 [[ADD20]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 940 // CHECK1-NEXT: br label [[OMP_DISPATCH_COND]] 941 // CHECK1: omp.dispatch.end: 942 // CHECK1-NEXT: [[TMP53:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 943 // CHECK1-NEXT: [[TMP54:%.*]] = load i32, i32* [[TMP53]], align 4, !tbaa [[TBAA8]] 944 // CHECK1-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB3]], i32 [[TMP54]]) 945 // CHECK1-NEXT: [[TMP55:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 946 // CHECK1-NEXT: [[TMP56:%.*]] = load i32, i32* [[TMP55]], align 4, !tbaa [[TBAA8]] 947 // CHECK1-NEXT: [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]], i64 0, i64 0 948 // CHECK1-NEXT: [[TMP58:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM5]] to i8* 949 // CHECK1-NEXT: store i8* [[TMP58]], i8** [[TMP57]], align 8 950 // CHECK1-NEXT: [[TMP59:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]] to i8* 951 // CHECK1-NEXT: [[TMP60:%.*]] = call i32 @__kmpc_nvptx_parallel_reduce_nowait_v2(%struct.ident_t* @[[GLOB1]], i32 [[TMP56]], i32 1, i64 8, i8* [[TMP59]], void (i8*, i16, i16, i16)* @_omp_reduction_shuffle_and_reduce_func5, void (i8*, i32)* @_omp_reduction_inter_warp_copy_func6) 952 // CHECK1-NEXT: [[TMP61:%.*]] = icmp eq i32 [[TMP60]], 1 953 // CHECK1-NEXT: br i1 [[TMP61]], label [[DOTOMP_REDUCTION_THEN:%.*]], label [[DOTOMP_REDUCTION_DONE:%.*]] 954 // CHECK1: .omp.reduction.then: 955 // CHECK1-NEXT: [[CALL21:%.*]] = call nonnull align 8 dereferenceable(16) %"class.std::complex.0"* @_ZNSt7complexIdEpLIdEERS0_RKS_IT_E(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[TMP2]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM5]]) #[[ATTR8]] 956 // CHECK1-NEXT: call void @__kmpc_nvptx_end_reduce_nowait(i32 [[TMP56]]) 957 // CHECK1-NEXT: br label [[DOTOMP_REDUCTION_DONE]] 958 // CHECK1: .omp.reduction.done: 959 // CHECK1-NEXT: [[TMP62:%.*]] = bitcast i32* [[I7]] to i8* 960 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP62]]) #[[ATTR1]] 961 // CHECK1-NEXT: [[TMP63:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM5]] to i8* 962 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[TMP63]]) #[[ATTR1]] 963 // CHECK1-NEXT: [[TMP64:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 964 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP64]]) #[[ATTR1]] 965 // CHECK1-NEXT: [[TMP65:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 966 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP65]]) #[[ATTR1]] 967 // CHECK1-NEXT: [[TMP66:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 968 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP66]]) #[[ATTR1]] 969 // CHECK1-NEXT: [[TMP67:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 970 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP67]]) #[[ATTR1]] 971 // CHECK1-NEXT: br label [[OMP_PRECOND_END]] 972 // CHECK1: omp.precond.end: 973 // CHECK1-NEXT: [[TMP68:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 974 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP68]]) #[[ATTR1]] 975 // CHECK1-NEXT: [[TMP69:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 976 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP69]]) #[[ATTR1]] 977 // CHECK1-NEXT: [[TMP70:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 978 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP70]]) #[[ATTR1]] 979 // CHECK1-NEXT: [[TMP71:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 980 // CHECK1-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP71]]) #[[ATTR1]] 981 // CHECK1-NEXT: ret void 982 // 983 // 984 // CHECK1-LABEL: define {{[^@]+}}@_ZNSt7complexIdEpLIdEERS0_RKS_IT_E 985 // CHECK1-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[__C:%.*]]) #[[ATTR4]] comdat align 2 { 986 // CHECK1-NEXT: entry: 987 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 988 // CHECK1-NEXT: [[__C_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 989 // CHECK1-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 990 // CHECK1-NEXT: store %"class.std::complex.0"* [[__C]], %"class.std::complex.0"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 991 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 992 // CHECK1-NEXT: [[TMP0:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 993 // CHECK1-NEXT: [[CALL:%.*]] = call double @_ZNKSt7complexIdE4realEv(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[TMP0]]) #[[ATTR8]] 994 // CHECK1-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 0 995 // CHECK1-NEXT: [[TMP1:%.*]] = load double, double* [[__RE_]], align 8, !tbaa [[TBAA24:![0-9]+]] 996 // CHECK1-NEXT: [[ADD:%.*]] = fadd double [[TMP1]], [[CALL]] 997 // CHECK1-NEXT: store double [[ADD]], double* [[__RE_]], align 8, !tbaa [[TBAA24]] 998 // CHECK1-NEXT: [[TMP2:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 999 // CHECK1-NEXT: [[CALL2:%.*]] = call double @_ZNKSt7complexIdE4imagEv(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[TMP2]]) #[[ATTR8]] 1000 // CHECK1-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 1 1001 // CHECK1-NEXT: [[TMP3:%.*]] = load double, double* [[__IM_]], align 8, !tbaa [[TBAA26:![0-9]+]] 1002 // CHECK1-NEXT: [[ADD3:%.*]] = fadd double [[TMP3]], [[CALL2]] 1003 // CHECK1-NEXT: store double [[ADD3]], double* [[__IM_]], align 8, !tbaa [[TBAA26]] 1004 // CHECK1-NEXT: ret %"class.std::complex.0"* [[THIS1]] 1005 // 1006 // 1007 // CHECK1-LABEL: define {{[^@]+}}@_omp_reduction_shuffle_and_reduce_func5 1008 // CHECK1-SAME: (i8* [[TMP0:%.*]], i16 signext [[TMP1:%.*]], i16 signext [[TMP2:%.*]], i16 signext [[TMP3:%.*]]) #[[ATTR0]] { 1009 // CHECK1-NEXT: entry: 1010 // CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 1011 // CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca i16, align 2 1012 // CHECK1-NEXT: [[DOTADDR2:%.*]] = alloca i16, align 2 1013 // CHECK1-NEXT: [[DOTADDR3:%.*]] = alloca i16, align 2 1014 // CHECK1-NEXT: [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST:%.*]] = alloca [1 x i8*], align 8 1015 // CHECK1-NEXT: [[DOTOMP_REDUCTION_ELEMENT:%.*]] = alloca %"class.std::complex.0", align 8 1016 // CHECK1-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 1017 // CHECK1-NEXT: store i16 [[TMP1]], i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19]] 1018 // CHECK1-NEXT: store i16 [[TMP2]], i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 1019 // CHECK1-NEXT: store i16 [[TMP3]], i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 1020 // CHECK1-NEXT: [[TMP4:%.*]] = load i8*, i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 1021 // CHECK1-NEXT: [[TMP5:%.*]] = bitcast i8* [[TMP4]] to [1 x i8*]* 1022 // CHECK1-NEXT: [[TMP6:%.*]] = load i16, i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19]] 1023 // CHECK1-NEXT: [[TMP7:%.*]] = load i16, i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 1024 // CHECK1-NEXT: [[TMP8:%.*]] = load i16, i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 1025 // CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 1026 // CHECK1-NEXT: [[TMP10:%.*]] = load i8*, i8** [[TMP9]], align 8 1027 // CHECK1-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 1028 // CHECK1-NEXT: [[TMP12:%.*]] = bitcast i8* [[TMP10]] to %"class.std::complex.0"* 1029 // CHECK1-NEXT: [[TMP13:%.*]] = getelementptr %"class.std::complex.0", %"class.std::complex.0"* [[TMP12]], i64 1 1030 // CHECK1-NEXT: [[TMP14:%.*]] = bitcast %"class.std::complex.0"* [[TMP13]] to i8* 1031 // CHECK1-NEXT: [[TMP15:%.*]] = bitcast %"class.std::complex.0"* [[TMP12]] to i64* 1032 // CHECK1-NEXT: [[TMP16:%.*]] = bitcast %"class.std::complex.0"* [[DOTOMP_REDUCTION_ELEMENT]] to i64* 1033 // CHECK1-NEXT: br label [[DOTSHUFFLE_PRE_COND:%.*]] 1034 // CHECK1: .shuffle.pre_cond: 1035 // CHECK1-NEXT: [[TMP17:%.*]] = phi i64* [ [[TMP15]], [[ENTRY:%.*]] ], [ [[TMP28:%.*]], [[DOTSHUFFLE_THEN:%.*]] ] 1036 // CHECK1-NEXT: [[TMP18:%.*]] = phi i64* [ [[TMP16]], [[ENTRY]] ], [ [[TMP29:%.*]], [[DOTSHUFFLE_THEN]] ] 1037 // CHECK1-NEXT: [[TMP19:%.*]] = bitcast i64* [[TMP17]] to i8* 1038 // CHECK1-NEXT: [[TMP20:%.*]] = ptrtoint i8* [[TMP14]] to i64 1039 // CHECK1-NEXT: [[TMP21:%.*]] = ptrtoint i8* [[TMP19]] to i64 1040 // CHECK1-NEXT: [[TMP22:%.*]] = sub i64 [[TMP20]], [[TMP21]] 1041 // CHECK1-NEXT: [[TMP23:%.*]] = sdiv exact i64 [[TMP22]], ptrtoint (i8* getelementptr (i8, i8* null, i32 1) to i64) 1042 // CHECK1-NEXT: [[TMP24:%.*]] = icmp sgt i64 [[TMP23]], 7 1043 // CHECK1-NEXT: br i1 [[TMP24]], label [[DOTSHUFFLE_THEN]], label [[DOTSHUFFLE_EXIT:%.*]] 1044 // CHECK1: .shuffle.then: 1045 // CHECK1-NEXT: [[TMP25:%.*]] = load i64, i64* [[TMP17]], align 8 1046 // CHECK1-NEXT: [[NVPTX_WARP_SIZE:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize() 1047 // CHECK1-NEXT: [[TMP26:%.*]] = trunc i32 [[NVPTX_WARP_SIZE]] to i16 1048 // CHECK1-NEXT: [[TMP27:%.*]] = call i64 @__kmpc_shuffle_int64(i64 [[TMP25]], i16 [[TMP7]], i16 [[TMP26]]) 1049 // CHECK1-NEXT: store i64 [[TMP27]], i64* [[TMP18]], align 8 1050 // CHECK1-NEXT: [[TMP28]] = getelementptr i64, i64* [[TMP17]], i64 1 1051 // CHECK1-NEXT: [[TMP29]] = getelementptr i64, i64* [[TMP18]], i64 1 1052 // CHECK1-NEXT: br label [[DOTSHUFFLE_PRE_COND]] 1053 // CHECK1: .shuffle.exit: 1054 // CHECK1-NEXT: [[TMP30:%.*]] = bitcast %"class.std::complex.0"* [[DOTOMP_REDUCTION_ELEMENT]] to i8* 1055 // CHECK1-NEXT: store i8* [[TMP30]], i8** [[TMP11]], align 8, !tbaa [[TBAA12]] 1056 // CHECK1-NEXT: [[TMP31:%.*]] = icmp eq i16 [[TMP8]], 0 1057 // CHECK1-NEXT: [[TMP32:%.*]] = icmp eq i16 [[TMP8]], 1 1058 // CHECK1-NEXT: [[TMP33:%.*]] = icmp ult i16 [[TMP6]], [[TMP7]] 1059 // CHECK1-NEXT: [[TMP34:%.*]] = and i1 [[TMP32]], [[TMP33]] 1060 // CHECK1-NEXT: [[TMP35:%.*]] = icmp eq i16 [[TMP8]], 2 1061 // CHECK1-NEXT: [[TMP36:%.*]] = and i16 [[TMP6]], 1 1062 // CHECK1-NEXT: [[TMP37:%.*]] = icmp eq i16 [[TMP36]], 0 1063 // CHECK1-NEXT: [[TMP38:%.*]] = and i1 [[TMP35]], [[TMP37]] 1064 // CHECK1-NEXT: [[TMP39:%.*]] = icmp sgt i16 [[TMP7]], 0 1065 // CHECK1-NEXT: [[TMP40:%.*]] = and i1 [[TMP38]], [[TMP39]] 1066 // CHECK1-NEXT: [[TMP41:%.*]] = or i1 [[TMP31]], [[TMP34]] 1067 // CHECK1-NEXT: [[TMP42:%.*]] = or i1 [[TMP41]], [[TMP40]] 1068 // CHECK1-NEXT: br i1 [[TMP42]], label [[THEN:%.*]], label [[ELSE:%.*]] 1069 // CHECK1: then: 1070 // CHECK1-NEXT: [[TMP43:%.*]] = bitcast [1 x i8*]* [[TMP5]] to i8* 1071 // CHECK1-NEXT: [[TMP44:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]] to i8* 1072 // CHECK1-NEXT: call void @"_omp$reduction$reduction_func4"(i8* [[TMP43]], i8* [[TMP44]]) #[[ATTR1]] 1073 // CHECK1-NEXT: br label [[IFCONT:%.*]] 1074 // CHECK1: else: 1075 // CHECK1-NEXT: br label [[IFCONT]] 1076 // CHECK1: ifcont: 1077 // CHECK1-NEXT: [[TMP45:%.*]] = icmp eq i16 [[TMP8]], 1 1078 // CHECK1-NEXT: [[TMP46:%.*]] = icmp uge i16 [[TMP6]], [[TMP7]] 1079 // CHECK1-NEXT: [[TMP47:%.*]] = and i1 [[TMP45]], [[TMP46]] 1080 // CHECK1-NEXT: br i1 [[TMP47]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 1081 // CHECK1: then4: 1082 // CHECK1-NEXT: [[TMP48:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 1083 // CHECK1-NEXT: [[TMP49:%.*]] = load i8*, i8** [[TMP48]], align 8 1084 // CHECK1-NEXT: [[TMP50:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 1085 // CHECK1-NEXT: [[TMP51:%.*]] = load i8*, i8** [[TMP50]], align 8 1086 // CHECK1-NEXT: [[TMP52:%.*]] = bitcast i8* [[TMP49]] to %"class.std::complex.0"* 1087 // CHECK1-NEXT: [[TMP53:%.*]] = bitcast i8* [[TMP51]] to %"class.std::complex.0"* 1088 // CHECK1-NEXT: [[TMP54:%.*]] = bitcast %"class.std::complex.0"* [[TMP53]] to i8* 1089 // CHECK1-NEXT: [[TMP55:%.*]] = bitcast %"class.std::complex.0"* [[TMP52]] to i8* 1090 // CHECK1-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP54]], i8* align 8 [[TMP55]], i64 16, i1 false), !tbaa.struct !27 1091 // CHECK1-NEXT: br label [[IFCONT6:%.*]] 1092 // CHECK1: else5: 1093 // CHECK1-NEXT: br label [[IFCONT6]] 1094 // CHECK1: ifcont6: 1095 // CHECK1-NEXT: ret void 1096 // 1097 // 1098 // CHECK1-LABEL: define {{[^@]+}}@_omp_reduction_inter_warp_copy_func6 1099 // CHECK1-SAME: (i8* [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 1100 // CHECK1-NEXT: entry: 1101 // CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 1102 // CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 1103 // CHECK1-NEXT: [[DOTCNT_ADDR:%.*]] = alloca i32, align 4 1104 // CHECK1-NEXT: [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 1105 // CHECK1-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 1106 // CHECK1-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 1107 // CHECK1-NEXT: [[NVPTX_TID:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 1108 // CHECK1-NEXT: [[NVPTX_TID2:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 1109 // CHECK1-NEXT: [[NVPTX_LANE_ID:%.*]] = and i32 [[NVPTX_TID2]], 31 1110 // CHECK1-NEXT: [[NVPTX_TID3:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 1111 // CHECK1-NEXT: [[NVPTX_WARP_ID:%.*]] = ashr i32 [[NVPTX_TID3]], 5 1112 // CHECK1-NEXT: [[TMP3:%.*]] = load i8*, i8** [[DOTADDR]], align 8 1113 // CHECK1-NEXT: [[TMP4:%.*]] = bitcast i8* [[TMP3]] to [1 x i8*]* 1114 // CHECK1-NEXT: store i32 0, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 1115 // CHECK1-NEXT: br label [[PRECOND:%.*]] 1116 // CHECK1: precond: 1117 // CHECK1-NEXT: [[TMP5:%.*]] = load i32, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 1118 // CHECK1-NEXT: [[TMP6:%.*]] = icmp ult i32 [[TMP5]], 4 1119 // CHECK1-NEXT: br i1 [[TMP6]], label [[BODY:%.*]], label [[EXIT:%.*]] 1120 // CHECK1: body: 1121 // CHECK1-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4]], i32 [[TMP2]]) 1122 // CHECK1-NEXT: [[WARP_MASTER:%.*]] = icmp eq i32 [[NVPTX_LANE_ID]], 0 1123 // CHECK1-NEXT: br i1 [[WARP_MASTER]], label [[THEN:%.*]], label [[ELSE:%.*]] 1124 // CHECK1: then: 1125 // CHECK1-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 1126 // CHECK1-NEXT: [[TMP8:%.*]] = load i8*, i8** [[TMP7]], align 8, !tbaa [[TBAA12]] 1127 // CHECK1-NEXT: [[TMP9:%.*]] = bitcast i8* [[TMP8]] to i32* 1128 // CHECK1-NEXT: [[TMP10:%.*]] = getelementptr i32, i32* [[TMP9]], i32 [[TMP5]] 1129 // CHECK1-NEXT: [[TMP11:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_WARP_ID]] 1130 // CHECK1-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP10]], align 4 1131 // CHECK1-NEXT: store volatile i32 [[TMP12]], i32 addrspace(3)* [[TMP11]], align 4 1132 // CHECK1-NEXT: br label [[IFCONT:%.*]] 1133 // CHECK1: else: 1134 // CHECK1-NEXT: br label [[IFCONT]] 1135 // CHECK1: ifcont: 1136 // CHECK1-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4]], i32 [[TMP2]]) 1137 // CHECK1-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 1138 // CHECK1-NEXT: [[IS_ACTIVE_THREAD:%.*]] = icmp ult i32 [[NVPTX_TID]], [[TMP13]] 1139 // CHECK1-NEXT: br i1 [[IS_ACTIVE_THREAD]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 1140 // CHECK1: then4: 1141 // CHECK1-NEXT: [[TMP14:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_TID]] 1142 // CHECK1-NEXT: [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 1143 // CHECK1-NEXT: [[TMP16:%.*]] = load i8*, i8** [[TMP15]], align 8, !tbaa [[TBAA12]] 1144 // CHECK1-NEXT: [[TMP17:%.*]] = bitcast i8* [[TMP16]] to i32* 1145 // CHECK1-NEXT: [[TMP18:%.*]] = getelementptr i32, i32* [[TMP17]], i32 [[TMP5]] 1146 // CHECK1-NEXT: [[TMP19:%.*]] = load volatile i32, i32 addrspace(3)* [[TMP14]], align 4, !tbaa [[TBAA8]] 1147 // CHECK1-NEXT: store i32 [[TMP19]], i32* [[TMP18]], align 4, !tbaa [[TBAA8]] 1148 // CHECK1-NEXT: br label [[IFCONT6:%.*]] 1149 // CHECK1: else5: 1150 // CHECK1-NEXT: br label [[IFCONT6]] 1151 // CHECK1: ifcont6: 1152 // CHECK1-NEXT: [[TMP20:%.*]] = add nsw i32 [[TMP5]], 1 1153 // CHECK1-NEXT: store i32 [[TMP20]], i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 1154 // CHECK1-NEXT: br label [[PRECOND]] 1155 // CHECK1: exit: 1156 // CHECK1-NEXT: ret void 1157 // 1158 // 1159 // CHECK1-LABEL: define {{[^@]+}}@__omp_outlined__3_wrapper 1160 // CHECK1-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 1161 // CHECK1-NEXT: entry: 1162 // CHECK1-NEXT: [[DOTADDR:%.*]] = alloca i16, align 2 1163 // CHECK1-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 1164 // CHECK1-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 1165 // CHECK1-NEXT: [[GLOBAL_ARGS:%.*]] = alloca i8**, align 8 1166 // CHECK1-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 1167 // CHECK1-NEXT: store i16 [[TMP0]], i16* [[DOTADDR]], align 2, !tbaa [[TBAA19]] 1168 // CHECK1-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 1169 // CHECK1-NEXT: call void @__kmpc_get_shared_variables(i8*** [[GLOBAL_ARGS]]) 1170 // CHECK1-NEXT: [[TMP2:%.*]] = load i8**, i8*** [[GLOBAL_ARGS]], align 8 1171 // CHECK1-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 0 1172 // CHECK1-NEXT: [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32** 1173 // CHECK1-NEXT: [[TMP5:%.*]] = load i32*, i32** [[TMP4]], align 8, !tbaa [[TBAA12]] 1174 // CHECK1-NEXT: [[TMP6:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 1 1175 // CHECK1-NEXT: [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i32** 1176 // CHECK1-NEXT: [[TMP8:%.*]] = load i32*, i32** [[TMP7]], align 8, !tbaa [[TBAA12]] 1177 // CHECK1-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 2 1178 // CHECK1-NEXT: [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %"class.std::complex.0"** 1179 // CHECK1-NEXT: [[TMP11:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[TMP10]], align 8, !tbaa [[TBAA12]] 1180 // CHECK1-NEXT: call void @__omp_outlined__3(i32* [[DOTADDR1]], i32* [[DOTZERO_ADDR]], i32* [[TMP5]], i32* [[TMP8]], %"class.std::complex.0"* [[TMP11]]) #[[ATTR1]] 1181 // CHECK1-NEXT: ret void 1182 // 1183 // 1184 // CHECK1-LABEL: define {{[^@]+}}@_ZNSt7complexIfEC2ERKfS2_ 1185 // CHECK1-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]], float* nonnull align 4 dereferenceable(4) [[__RE:%.*]], float* nonnull align 4 dereferenceable(4) [[__IM:%.*]]) unnamed_addr #[[ATTR3]] comdat align 2 { 1186 // CHECK1-NEXT: entry: 1187 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 1188 // CHECK1-NEXT: [[__RE_ADDR:%.*]] = alloca float*, align 8 1189 // CHECK1-NEXT: [[__IM_ADDR:%.*]] = alloca float*, align 8 1190 // CHECK1-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 1191 // CHECK1-NEXT: store float* [[__RE]], float** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 1192 // CHECK1-NEXT: store float* [[__IM]], float** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 1193 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 1194 // CHECK1-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 0 1195 // CHECK1-NEXT: [[TMP0:%.*]] = load float*, float** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 1196 // CHECK1-NEXT: [[TMP1:%.*]] = load float, float* [[TMP0]], align 4, !tbaa [[TBAA14]] 1197 // CHECK1-NEXT: store float [[TMP1]], float* [[__RE_]], align 4, !tbaa [[TBAA16]] 1198 // CHECK1-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 1 1199 // CHECK1-NEXT: [[TMP2:%.*]] = load float*, float** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 1200 // CHECK1-NEXT: [[TMP3:%.*]] = load float, float* [[TMP2]], align 4, !tbaa [[TBAA14]] 1201 // CHECK1-NEXT: store float [[TMP3]], float* [[__IM_]], align 4, !tbaa [[TBAA18]] 1202 // CHECK1-NEXT: ret void 1203 // 1204 // 1205 // CHECK1-LABEL: define {{[^@]+}}@_ZNKSt7complexIfE4realEv 1206 // CHECK1-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 1207 // CHECK1-NEXT: entry: 1208 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 1209 // CHECK1-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 1210 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 1211 // CHECK1-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 0 1212 // CHECK1-NEXT: [[TMP0:%.*]] = load float, float* [[__RE_]], align 4, !tbaa [[TBAA16]] 1213 // CHECK1-NEXT: ret float [[TMP0]] 1214 // 1215 // 1216 // CHECK1-LABEL: define {{[^@]+}}@_ZNKSt7complexIfE4imagEv 1217 // CHECK1-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 1218 // CHECK1-NEXT: entry: 1219 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 1220 // CHECK1-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 1221 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 1222 // CHECK1-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 1 1223 // CHECK1-NEXT: [[TMP0:%.*]] = load float, float* [[__IM_]], align 4, !tbaa [[TBAA18]] 1224 // CHECK1-NEXT: ret float [[TMP0]] 1225 // 1226 // 1227 // CHECK1-LABEL: define {{[^@]+}}@_ZNSt7complexIdEC2ERKdS2_ 1228 // CHECK1-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]], double* nonnull align 8 dereferenceable(8) [[__RE:%.*]], double* nonnull align 8 dereferenceable(8) [[__IM:%.*]]) unnamed_addr #[[ATTR3]] comdat align 2 { 1229 // CHECK1-NEXT: entry: 1230 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 1231 // CHECK1-NEXT: [[__RE_ADDR:%.*]] = alloca double*, align 8 1232 // CHECK1-NEXT: [[__IM_ADDR:%.*]] = alloca double*, align 8 1233 // CHECK1-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 1234 // CHECK1-NEXT: store double* [[__RE]], double** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 1235 // CHECK1-NEXT: store double* [[__IM]], double** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 1236 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 1237 // CHECK1-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 0 1238 // CHECK1-NEXT: [[TMP0:%.*]] = load double*, double** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 1239 // CHECK1-NEXT: [[TMP1:%.*]] = load double, double* [[TMP0]], align 8, !tbaa [[TBAA22]] 1240 // CHECK1-NEXT: store double [[TMP1]], double* [[__RE_]], align 8, !tbaa [[TBAA24]] 1241 // CHECK1-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 1 1242 // CHECK1-NEXT: [[TMP2:%.*]] = load double*, double** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 1243 // CHECK1-NEXT: [[TMP3:%.*]] = load double, double* [[TMP2]], align 8, !tbaa [[TBAA22]] 1244 // CHECK1-NEXT: store double [[TMP3]], double* [[__IM_]], align 8, !tbaa [[TBAA26]] 1245 // CHECK1-NEXT: ret void 1246 // 1247 // 1248 // CHECK1-LABEL: define {{[^@]+}}@_ZNKSt7complexIdE4realEv 1249 // CHECK1-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 1250 // CHECK1-NEXT: entry: 1251 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 1252 // CHECK1-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 1253 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 1254 // CHECK1-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 0 1255 // CHECK1-NEXT: [[TMP0:%.*]] = load double, double* [[__RE_]], align 8, !tbaa [[TBAA24]] 1256 // CHECK1-NEXT: ret double [[TMP0]] 1257 // 1258 // 1259 // CHECK1-LABEL: define {{[^@]+}}@_ZNKSt7complexIdE4imagEv 1260 // CHECK1-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 1261 // CHECK1-NEXT: entry: 1262 // CHECK1-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 1263 // CHECK1-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 1264 // CHECK1-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 1265 // CHECK1-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 1 1266 // CHECK1-NEXT: [[TMP0:%.*]] = load double, double* [[__IM_]], align 8, !tbaa [[TBAA26]] 1267 // CHECK1-NEXT: ret double [[TMP0]] 1268 // 1269 // 1270 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z17complex_reductionIfEvv_l19 1271 // CHECK2-SAME: () #[[ATTR0:[0-9]+]] { 1272 // CHECK2-NEXT: entry: 1273 // CHECK2-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 1274 // CHECK2-NEXT: [[DOTTHREADID_TEMP_:%.*]] = alloca i32, align 4 1275 // CHECK2-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 1276 // CHECK2-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* @[[GLOB1:[0-9]+]], i1 false, i1 true, i1 true) 1277 // CHECK2-NEXT: [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1 1278 // CHECK2-NEXT: br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]] 1279 // CHECK2: user_code.entry: 1280 // CHECK2-NEXT: [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 1281 // CHECK2-NEXT: store i32 [[TMP1]], i32* [[DOTTHREADID_TEMP_]], align 4, !tbaa [[TBAA8:![0-9]+]] 1282 // CHECK2-NEXT: call void @__omp_outlined__(i32* [[DOTTHREADID_TEMP_]], i32* [[DOTZERO_ADDR]]) #[[ATTR1:[0-9]+]] 1283 // CHECK2-NEXT: call void @__kmpc_target_deinit(%struct.ident_t* @[[GLOB1]], i1 false, i1 true) 1284 // CHECK2-NEXT: ret void 1285 // CHECK2: worker.exit: 1286 // CHECK2-NEXT: ret void 1287 // 1288 // 1289 // CHECK2-LABEL: define {{[^@]+}}@__omp_outlined__ 1290 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR0]] { 1291 // CHECK2-NEXT: entry: 1292 // CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 1293 // CHECK2-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 1294 // CHECK2-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 1295 // CHECK2-NEXT: [[TMP:%.*]] = alloca i32, align 4 1296 // CHECK2-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 1297 // CHECK2-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 1298 // CHECK2-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 1299 // CHECK2-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 1300 // CHECK2-NEXT: [[IB:%.*]] = alloca i32, align 4 1301 // CHECK2-NEXT: [[REF_TMP:%.*]] = alloca float, align 4 1302 // CHECK2-NEXT: [[REF_TMP2:%.*]] = alloca float, align 4 1303 // CHECK2-NEXT: [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x i8*], align 8 1304 // CHECK2-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12:![0-9]+]] 1305 // CHECK2-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 1306 // CHECK2-NEXT: [[ISTART:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 1307 // CHECK2-NEXT: [[ISTART_ON_STACK:%.*]] = bitcast i8* [[ISTART]] to i32* 1308 // CHECK2-NEXT: [[IEND:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 1309 // CHECK2-NEXT: [[IEND_ON_STACK:%.*]] = bitcast i8* [[IEND]] to i32* 1310 // CHECK2-NEXT: [[PARTIAL_SUM:%.*]] = call i8* @__kmpc_alloc_shared(i64 8) 1311 // CHECK2-NEXT: [[PARTIAL_SUM_ON_STACK:%.*]] = bitcast i8* [[PARTIAL_SUM]] to %"class.std::complex"* 1312 // CHECK2-NEXT: [[TMP0:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 1313 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP0]]) #[[ATTR1]] 1314 // CHECK2-NEXT: [[TMP1:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 1315 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP1]]) #[[ATTR1]] 1316 // CHECK2-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 1317 // CHECK2-NEXT: [[TMP2:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 1318 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP2]]) #[[ATTR1]] 1319 // CHECK2-NEXT: store i32 99, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1320 // CHECK2-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 1321 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 1322 // CHECK2-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 1323 // CHECK2-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 1324 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 1325 // CHECK2-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 1326 // CHECK2-NEXT: [[TMP5:%.*]] = bitcast i32* [[IB]] to i8* 1327 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP5]]) #[[ATTR1]] 1328 // CHECK2-NEXT: [[TMP6:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 1329 // CHECK2-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP6]], align 4, !tbaa [[TBAA8]] 1330 // CHECK2-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP7]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 1331 // CHECK2-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1332 // CHECK2-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP8]], 99 1333 // CHECK2-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 1334 // CHECK2: cond.true: 1335 // CHECK2-NEXT: br label [[COND_END:%.*]] 1336 // CHECK2: cond.false: 1337 // CHECK2-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1338 // CHECK2-NEXT: br label [[COND_END]] 1339 // CHECK2: cond.end: 1340 // CHECK2-NEXT: [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP9]], [[COND_FALSE]] ] 1341 // CHECK2-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1342 // CHECK2-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 1343 // CHECK2-NEXT: store i32 [[TMP10]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1344 // CHECK2-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 1345 // CHECK2: omp.inner.for.cond: 1346 // CHECK2-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1347 // CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1348 // CHECK2-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP11]], [[TMP12]] 1349 // CHECK2-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 1350 // CHECK2: omp.inner.for.cond.cleanup: 1351 // CHECK2-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 1352 // CHECK2: omp.inner.for.body: 1353 // CHECK2-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1354 // CHECK2-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP13]], 1 1355 // CHECK2-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]] 1356 // CHECK2-NEXT: store i32 [[ADD]], i32* [[IB]], align 4, !tbaa [[TBAA8]] 1357 // CHECK2-NEXT: [[TMP14:%.*]] = bitcast float* [[REF_TMP]] to i8* 1358 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP14]]) #[[ATTR1]] 1359 // CHECK2-NEXT: store float 0.000000e+00, float* [[REF_TMP]], align 4, !tbaa [[TBAA14:![0-9]+]] 1360 // CHECK2-NEXT: [[TMP15:%.*]] = bitcast float* [[REF_TMP2]] to i8* 1361 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP15]]) #[[ATTR1]] 1362 // CHECK2-NEXT: store float 0.000000e+00, float* [[REF_TMP2]], align 4, !tbaa [[TBAA14]] 1363 // CHECK2-NEXT: call void @_ZNSt7complexIfEC1ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM_ON_STACK]], float* nonnull align 4 dereferenceable(4) [[REF_TMP]], float* nonnull align 4 dereferenceable(4) [[REF_TMP2]]) #[[ATTR8:[0-9]+]] 1364 // CHECK2-NEXT: [[TMP16:%.*]] = bitcast float* [[REF_TMP2]] to i8* 1365 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP16]]) #[[ATTR1]] 1366 // CHECK2-NEXT: [[TMP17:%.*]] = bitcast float* [[REF_TMP]] to i8* 1367 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP17]]) #[[ATTR1]] 1368 // CHECK2-NEXT: [[TMP18:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 1369 // CHECK2-NEXT: [[MUL3:%.*]] = mul nsw i32 [[TMP18]], 4 1370 // CHECK2-NEXT: store i32 [[MUL3]], i32* [[ISTART_ON_STACK]], align 4, !tbaa [[TBAA8]] 1371 // CHECK2-NEXT: [[TMP19:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 1372 // CHECK2-NEXT: [[ADD4:%.*]] = add nsw i32 [[TMP19]], 1 1373 // CHECK2-NEXT: [[MUL5:%.*]] = mul nsw i32 [[ADD4]], 4 1374 // CHECK2-NEXT: store i32 [[MUL5]], i32* [[IEND_ON_STACK]], align 4, !tbaa [[TBAA8]] 1375 // CHECK2-NEXT: [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 0 1376 // CHECK2-NEXT: [[TMP21:%.*]] = bitcast i32* [[ISTART_ON_STACK]] to i8* 1377 // CHECK2-NEXT: store i8* [[TMP21]], i8** [[TMP20]], align 8, !tbaa [[TBAA12]] 1378 // CHECK2-NEXT: [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 1 1379 // CHECK2-NEXT: [[TMP23:%.*]] = bitcast i32* [[IEND_ON_STACK]] to i8* 1380 // CHECK2-NEXT: store i8* [[TMP23]], i8** [[TMP22]], align 8, !tbaa [[TBAA12]] 1381 // CHECK2-NEXT: [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 2 1382 // CHECK2-NEXT: [[TMP25:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM_ON_STACK]] to i8* 1383 // CHECK2-NEXT: store i8* [[TMP25]], i8** [[TMP24]], align 8, !tbaa [[TBAA12]] 1384 // CHECK2-NEXT: [[TMP26:%.*]] = bitcast [3 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8** 1385 // CHECK2-NEXT: call void @__kmpc_parallel_51(%struct.ident_t* @[[GLOB1]], i32 [[TMP7]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*, i32*, i32*, %"class.std::complex"*)* @__omp_outlined__1 to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined__1_wrapper to i8*), i8** [[TMP26]], i64 3) 1386 // CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 1387 // CHECK2: omp.body.continue: 1388 // CHECK2-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 1389 // CHECK2: omp.inner.for.inc: 1390 // CHECK2-NEXT: [[TMP27:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1391 // CHECK2-NEXT: [[ADD6:%.*]] = add nsw i32 [[TMP27]], 1 1392 // CHECK2-NEXT: store i32 [[ADD6]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1393 // CHECK2-NEXT: br label [[OMP_INNER_FOR_COND]] 1394 // CHECK2: omp.inner.for.end: 1395 // CHECK2-NEXT: br label [[OMP_LOOP_EXIT:%.*]] 1396 // CHECK2: omp.loop.exit: 1397 // CHECK2-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB2]], i32 [[TMP7]]) 1398 // CHECK2-NEXT: [[TMP28:%.*]] = bitcast i32* [[IB]] to i8* 1399 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP28]]) #[[ATTR1]] 1400 // CHECK2-NEXT: [[TMP29:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 1401 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP29]]) #[[ATTR1]] 1402 // CHECK2-NEXT: [[TMP30:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 1403 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP30]]) #[[ATTR1]] 1404 // CHECK2-NEXT: [[TMP31:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 1405 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP31]]) #[[ATTR1]] 1406 // CHECK2-NEXT: [[TMP32:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 1407 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP32]]) #[[ATTR1]] 1408 // CHECK2-NEXT: [[TMP33:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 1409 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP33]]) #[[ATTR1]] 1410 // CHECK2-NEXT: call void @__kmpc_free_shared(i8* [[PARTIAL_SUM]], i64 8) 1411 // CHECK2-NEXT: call void @__kmpc_free_shared(i8* [[IEND]], i64 4) 1412 // CHECK2-NEXT: call void @__kmpc_free_shared(i8* [[ISTART]], i64 4) 1413 // CHECK2-NEXT: ret void 1414 // 1415 // 1416 // CHECK2-LABEL: define {{[^@]+}}@_ZNSt7complexIfEC1ERKfS2_ 1417 // CHECK2-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]], float* nonnull align 4 dereferenceable(4) [[__RE:%.*]], float* nonnull align 4 dereferenceable(4) [[__IM:%.*]]) unnamed_addr #[[ATTR3:[0-9]+]] comdat align 2 { 1418 // CHECK2-NEXT: entry: 1419 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 1420 // CHECK2-NEXT: [[__RE_ADDR:%.*]] = alloca float*, align 8 1421 // CHECK2-NEXT: [[__IM_ADDR:%.*]] = alloca float*, align 8 1422 // CHECK2-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 1423 // CHECK2-NEXT: store float* [[__RE]], float** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 1424 // CHECK2-NEXT: store float* [[__IM]], float** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 1425 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 1426 // CHECK2-NEXT: [[TMP0:%.*]] = load float*, float** [[__RE_ADDR]], align 8 1427 // CHECK2-NEXT: [[TMP1:%.*]] = load float*, float** [[__IM_ADDR]], align 8 1428 // CHECK2-NEXT: call void @_ZNSt7complexIfEC2ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS1]], float* nonnull align 4 dereferenceable(4) [[TMP0]], float* nonnull align 4 dereferenceable(4) [[TMP1]]) #[[ATTR8]] 1429 // CHECK2-NEXT: ret void 1430 // 1431 // 1432 // CHECK2-LABEL: define {{[^@]+}}@__omp_outlined__1 1433 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ISTART:%.*]], i32* nonnull align 4 dereferenceable(4) [[IEND:%.*]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM:%.*]]) #[[ATTR0]] { 1434 // CHECK2-NEXT: entry: 1435 // CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 1436 // CHECK2-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 1437 // CHECK2-NEXT: [[ISTART_ADDR:%.*]] = alloca i32*, align 8 1438 // CHECK2-NEXT: [[IEND_ADDR:%.*]] = alloca i32*, align 8 1439 // CHECK2-NEXT: [[PARTIAL_SUM_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 1440 // CHECK2-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 1441 // CHECK2-NEXT: [[TMP:%.*]] = alloca i32, align 4 1442 // CHECK2-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4 1443 // CHECK2-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4 1444 // CHECK2-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4 1445 // CHECK2-NEXT: [[I:%.*]] = alloca i32, align 4 1446 // CHECK2-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 1447 // CHECK2-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 1448 // CHECK2-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 1449 // CHECK2-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 1450 // CHECK2-NEXT: [[PARTIAL_SUM5:%.*]] = alloca %"class.std::complex", align 4 1451 // CHECK2-NEXT: [[REF_TMP:%.*]] = alloca float, align 4 1452 // CHECK2-NEXT: [[REF_TMP6:%.*]] = alloca float, align 4 1453 // CHECK2-NEXT: [[I7:%.*]] = alloca i32, align 4 1454 // CHECK2-NEXT: [[REF_TMP14:%.*]] = alloca %"class.std::complex", align 4 1455 // CHECK2-NEXT: [[REF_TMP15:%.*]] = alloca float, align 4 1456 // CHECK2-NEXT: [[REF_TMP16:%.*]] = alloca float, align 4 1457 // CHECK2-NEXT: [[DOTOMP_REDUCTION_RED_LIST:%.*]] = alloca [1 x i8*], align 8 1458 // CHECK2-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12]] 1459 // CHECK2-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 1460 // CHECK2-NEXT: store i32* [[ISTART]], i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 1461 // CHECK2-NEXT: store i32* [[IEND]], i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 1462 // CHECK2-NEXT: store %"class.std::complex"* [[PARTIAL_SUM]], %"class.std::complex"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 1463 // CHECK2-NEXT: [[TMP0:%.*]] = load i32*, i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 1464 // CHECK2-NEXT: [[TMP1:%.*]] = load i32*, i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 1465 // CHECK2-NEXT: [[TMP2:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 1466 // CHECK2-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 1467 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 1468 // CHECK2-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 1469 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 1470 // CHECK2-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP0]], align 4, !tbaa [[TBAA8]] 1471 // CHECK2-NEXT: store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 1472 // CHECK2-NEXT: [[TMP6:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 1473 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP6]]) #[[ATTR1]] 1474 // CHECK2-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP1]], align 4, !tbaa [[TBAA8]] 1475 // CHECK2-NEXT: store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 1476 // CHECK2-NEXT: [[TMP8:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 1477 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP8]]) #[[ATTR1]] 1478 // CHECK2-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 1479 // CHECK2-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 1480 // CHECK2-NEXT: [[SUB:%.*]] = sub i32 [[TMP9]], [[TMP10]] 1481 // CHECK2-NEXT: [[SUB3:%.*]] = sub i32 [[SUB]], 1 1482 // CHECK2-NEXT: [[ADD:%.*]] = add i32 [[SUB3]], 1 1483 // CHECK2-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], 1 1484 // CHECK2-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1 1485 // CHECK2-NEXT: store i32 [[SUB4]], i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 1486 // CHECK2-NEXT: [[TMP11:%.*]] = bitcast i32* [[I]] to i8* 1487 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP11]]) #[[ATTR1]] 1488 // CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 1489 // CHECK2-NEXT: store i32 [[TMP12]], i32* [[I]], align 4, !tbaa [[TBAA8]] 1490 // CHECK2-NEXT: [[TMP13:%.*]] = bitcast i32* [[I]] to i8* 1491 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP13]]) #[[ATTR1]] 1492 // CHECK2-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 1493 // CHECK2-NEXT: [[TMP15:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 1494 // CHECK2-NEXT: [[CMP:%.*]] = icmp slt i32 [[TMP14]], [[TMP15]] 1495 // CHECK2-NEXT: br i1 [[CMP]], label [[OMP_PRECOND_THEN:%.*]], label [[OMP_PRECOND_END:%.*]] 1496 // CHECK2: omp.precond.then: 1497 // CHECK2-NEXT: [[TMP16:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 1498 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP16]]) #[[ATTR1]] 1499 // CHECK2-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 1500 // CHECK2-NEXT: [[TMP17:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 1501 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP17]]) #[[ATTR1]] 1502 // CHECK2-NEXT: [[TMP18:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 1503 // CHECK2-NEXT: store i32 [[TMP18]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1504 // CHECK2-NEXT: [[TMP19:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 1505 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP19]]) #[[ATTR1]] 1506 // CHECK2-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 1507 // CHECK2-NEXT: [[TMP20:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 1508 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP20]]) #[[ATTR1]] 1509 // CHECK2-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 1510 // CHECK2-NEXT: [[TMP21:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM5]] to i8* 1511 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP21]]) #[[ATTR1]] 1512 // CHECK2-NEXT: [[TMP22:%.*]] = bitcast float* [[REF_TMP]] to i8* 1513 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP22]]) #[[ATTR1]] 1514 // CHECK2-NEXT: store float 0.000000e+00, float* [[REF_TMP]], align 4, !tbaa [[TBAA14]] 1515 // CHECK2-NEXT: [[TMP23:%.*]] = bitcast float* [[REF_TMP6]] to i8* 1516 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP23]]) #[[ATTR1]] 1517 // CHECK2-NEXT: store float 0.000000e+00, float* [[REF_TMP6]], align 4, !tbaa [[TBAA14]] 1518 // CHECK2-NEXT: call void @_ZNSt7complexIfEC1ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM5]], float* nonnull align 4 dereferenceable(4) [[REF_TMP]], float* nonnull align 4 dereferenceable(4) [[REF_TMP6]]) #[[ATTR8]] 1519 // CHECK2-NEXT: [[TMP24:%.*]] = bitcast float* [[REF_TMP6]] to i8* 1520 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP24]]) #[[ATTR1]] 1521 // CHECK2-NEXT: [[TMP25:%.*]] = bitcast float* [[REF_TMP]] to i8* 1522 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP25]]) #[[ATTR1]] 1523 // CHECK2-NEXT: [[TMP26:%.*]] = bitcast i32* [[I7]] to i8* 1524 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP26]]) #[[ATTR1]] 1525 // CHECK2-NEXT: [[TMP27:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 1526 // CHECK2-NEXT: [[TMP28:%.*]] = load i32, i32* [[TMP27]], align 4, !tbaa [[TBAA8]] 1527 // CHECK2-NEXT: call void @__kmpc_for_static_init_4u(%struct.ident_t* @[[GLOB3:[0-9]+]], i32 [[TMP28]], i32 33, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 1528 // CHECK2-NEXT: br label [[OMP_DISPATCH_COND:%.*]] 1529 // CHECK2: omp.dispatch.cond: 1530 // CHECK2-NEXT: [[TMP29:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1531 // CHECK2-NEXT: [[TMP30:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 1532 // CHECK2-NEXT: [[CMP8:%.*]] = icmp ugt i32 [[TMP29]], [[TMP30]] 1533 // CHECK2-NEXT: br i1 [[CMP8]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 1534 // CHECK2: cond.true: 1535 // CHECK2-NEXT: [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 1536 // CHECK2-NEXT: br label [[COND_END:%.*]] 1537 // CHECK2: cond.false: 1538 // CHECK2-NEXT: [[TMP32:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1539 // CHECK2-NEXT: br label [[COND_END]] 1540 // CHECK2: cond.end: 1541 // CHECK2-NEXT: [[COND:%.*]] = phi i32 [ [[TMP31]], [[COND_TRUE]] ], [ [[TMP32]], [[COND_FALSE]] ] 1542 // CHECK2-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1543 // CHECK2-NEXT: [[TMP33:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 1544 // CHECK2-NEXT: store i32 [[TMP33]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1545 // CHECK2-NEXT: [[TMP34:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1546 // CHECK2-NEXT: [[TMP35:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1547 // CHECK2-NEXT: [[ADD9:%.*]] = add i32 [[TMP35]], 1 1548 // CHECK2-NEXT: [[CMP10:%.*]] = icmp ult i32 [[TMP34]], [[ADD9]] 1549 // CHECK2-NEXT: br i1 [[CMP10]], label [[OMP_DISPATCH_BODY:%.*]], label [[OMP_DISPATCH_CLEANUP:%.*]] 1550 // CHECK2: omp.dispatch.cleanup: 1551 // CHECK2-NEXT: br label [[OMP_DISPATCH_END:%.*]] 1552 // CHECK2: omp.dispatch.body: 1553 // CHECK2-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 1554 // CHECK2: omp.inner.for.cond: 1555 // CHECK2-NEXT: [[TMP36:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1556 // CHECK2-NEXT: [[TMP37:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1557 // CHECK2-NEXT: [[ADD11:%.*]] = add i32 [[TMP37]], 1 1558 // CHECK2-NEXT: [[CMP12:%.*]] = icmp ult i32 [[TMP36]], [[ADD11]] 1559 // CHECK2-NEXT: br i1 [[CMP12]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 1560 // CHECK2: omp.inner.for.cond.cleanup: 1561 // CHECK2-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 1562 // CHECK2: omp.inner.for.body: 1563 // CHECK2-NEXT: [[TMP38:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 1564 // CHECK2-NEXT: [[TMP39:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1565 // CHECK2-NEXT: [[MUL:%.*]] = mul i32 [[TMP39]], 1 1566 // CHECK2-NEXT: [[ADD13:%.*]] = add i32 [[TMP38]], [[MUL]] 1567 // CHECK2-NEXT: store i32 [[ADD13]], i32* [[I7]], align 4, !tbaa [[TBAA8]] 1568 // CHECK2-NEXT: [[TMP40:%.*]] = bitcast %"class.std::complex"* [[REF_TMP14]] to i8* 1569 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP40]]) #[[ATTR1]] 1570 // CHECK2-NEXT: [[TMP41:%.*]] = bitcast float* [[REF_TMP15]] to i8* 1571 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP41]]) #[[ATTR1]] 1572 // CHECK2-NEXT: [[TMP42:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 1573 // CHECK2-NEXT: [[CONV:%.*]] = sitofp i32 [[TMP42]] to float 1574 // CHECK2-NEXT: store float [[CONV]], float* [[REF_TMP15]], align 4, !tbaa [[TBAA14]] 1575 // CHECK2-NEXT: [[TMP43:%.*]] = bitcast float* [[REF_TMP16]] to i8* 1576 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP43]]) #[[ATTR1]] 1577 // CHECK2-NEXT: [[TMP44:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 1578 // CHECK2-NEXT: [[CONV17:%.*]] = sitofp i32 [[TMP44]] to float 1579 // CHECK2-NEXT: store float [[CONV17]], float* [[REF_TMP16]], align 4, !tbaa [[TBAA14]] 1580 // CHECK2-NEXT: call void @_ZNSt7complexIfEC1ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[REF_TMP14]], float* nonnull align 4 dereferenceable(4) [[REF_TMP15]], float* nonnull align 4 dereferenceable(4) [[REF_TMP16]]) #[[ATTR8]] 1581 // CHECK2-NEXT: [[CALL:%.*]] = call nonnull align 4 dereferenceable(8) %"class.std::complex"* @_ZNSt7complexIfEpLIfEERS0_RKS_IT_E(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM5]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[REF_TMP14]]) #[[ATTR8]] 1582 // CHECK2-NEXT: [[TMP45:%.*]] = bitcast float* [[REF_TMP16]] to i8* 1583 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP45]]) #[[ATTR1]] 1584 // CHECK2-NEXT: [[TMP46:%.*]] = bitcast float* [[REF_TMP15]] to i8* 1585 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP46]]) #[[ATTR1]] 1586 // CHECK2-NEXT: [[TMP47:%.*]] = bitcast %"class.std::complex"* [[REF_TMP14]] to i8* 1587 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP47]]) #[[ATTR1]] 1588 // CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 1589 // CHECK2: omp.body.continue: 1590 // CHECK2-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 1591 // CHECK2: omp.inner.for.inc: 1592 // CHECK2-NEXT: [[TMP48:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1593 // CHECK2-NEXT: [[ADD18:%.*]] = add i32 [[TMP48]], 1 1594 // CHECK2-NEXT: store i32 [[ADD18]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1595 // CHECK2-NEXT: br label [[OMP_INNER_FOR_COND]] 1596 // CHECK2: omp.inner.for.end: 1597 // CHECK2-NEXT: br label [[OMP_DISPATCH_INC:%.*]] 1598 // CHECK2: omp.dispatch.inc: 1599 // CHECK2-NEXT: [[TMP49:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 1600 // CHECK2-NEXT: [[TMP50:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 1601 // CHECK2-NEXT: [[ADD19:%.*]] = add i32 [[TMP49]], [[TMP50]] 1602 // CHECK2-NEXT: store i32 [[ADD19]], i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 1603 // CHECK2-NEXT: [[TMP51:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1604 // CHECK2-NEXT: [[TMP52:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 1605 // CHECK2-NEXT: [[ADD20:%.*]] = add i32 [[TMP51]], [[TMP52]] 1606 // CHECK2-NEXT: store i32 [[ADD20]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1607 // CHECK2-NEXT: br label [[OMP_DISPATCH_COND]] 1608 // CHECK2: omp.dispatch.end: 1609 // CHECK2-NEXT: [[TMP53:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 1610 // CHECK2-NEXT: [[TMP54:%.*]] = load i32, i32* [[TMP53]], align 4, !tbaa [[TBAA8]] 1611 // CHECK2-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB3]], i32 [[TMP54]]) 1612 // CHECK2-NEXT: [[TMP55:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 1613 // CHECK2-NEXT: [[TMP56:%.*]] = load i32, i32* [[TMP55]], align 4, !tbaa [[TBAA8]] 1614 // CHECK2-NEXT: [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]], i64 0, i64 0 1615 // CHECK2-NEXT: [[TMP58:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM5]] to i8* 1616 // CHECK2-NEXT: store i8* [[TMP58]], i8** [[TMP57]], align 8 1617 // CHECK2-NEXT: [[TMP59:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]] to i8* 1618 // CHECK2-NEXT: [[TMP60:%.*]] = call i32 @__kmpc_nvptx_parallel_reduce_nowait_v2(%struct.ident_t* @[[GLOB1]], i32 [[TMP56]], i32 1, i64 8, i8* [[TMP59]], void (i8*, i16, i16, i16)* @_omp_reduction_shuffle_and_reduce_func, void (i8*, i32)* @_omp_reduction_inter_warp_copy_func) 1619 // CHECK2-NEXT: [[TMP61:%.*]] = icmp eq i32 [[TMP60]], 1 1620 // CHECK2-NEXT: br i1 [[TMP61]], label [[DOTOMP_REDUCTION_THEN:%.*]], label [[DOTOMP_REDUCTION_DONE:%.*]] 1621 // CHECK2: .omp.reduction.then: 1622 // CHECK2-NEXT: [[CALL21:%.*]] = call nonnull align 4 dereferenceable(8) %"class.std::complex"* @_ZNSt7complexIfEpLIfEERS0_RKS_IT_E(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[TMP2]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM5]]) #[[ATTR8]] 1623 // CHECK2-NEXT: call void @__kmpc_nvptx_end_reduce_nowait(i32 [[TMP56]]) 1624 // CHECK2-NEXT: br label [[DOTOMP_REDUCTION_DONE]] 1625 // CHECK2: .omp.reduction.done: 1626 // CHECK2-NEXT: [[TMP62:%.*]] = bitcast i32* [[I7]] to i8* 1627 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP62]]) #[[ATTR1]] 1628 // CHECK2-NEXT: [[TMP63:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM5]] to i8* 1629 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP63]]) #[[ATTR1]] 1630 // CHECK2-NEXT: [[TMP64:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 1631 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP64]]) #[[ATTR1]] 1632 // CHECK2-NEXT: [[TMP65:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 1633 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP65]]) #[[ATTR1]] 1634 // CHECK2-NEXT: [[TMP66:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 1635 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP66]]) #[[ATTR1]] 1636 // CHECK2-NEXT: [[TMP67:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 1637 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP67]]) #[[ATTR1]] 1638 // CHECK2-NEXT: br label [[OMP_PRECOND_END]] 1639 // CHECK2: omp.precond.end: 1640 // CHECK2-NEXT: [[TMP68:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 1641 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP68]]) #[[ATTR1]] 1642 // CHECK2-NEXT: [[TMP69:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 1643 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP69]]) #[[ATTR1]] 1644 // CHECK2-NEXT: [[TMP70:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 1645 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP70]]) #[[ATTR1]] 1646 // CHECK2-NEXT: [[TMP71:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 1647 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP71]]) #[[ATTR1]] 1648 // CHECK2-NEXT: ret void 1649 // 1650 // 1651 // CHECK2-LABEL: define {{[^@]+}}@_ZNSt7complexIfEpLIfEERS0_RKS_IT_E 1652 // CHECK2-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[__C:%.*]]) #[[ATTR4:[0-9]+]] comdat align 2 { 1653 // CHECK2-NEXT: entry: 1654 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 1655 // CHECK2-NEXT: [[__C_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 1656 // CHECK2-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 1657 // CHECK2-NEXT: store %"class.std::complex"* [[__C]], %"class.std::complex"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 1658 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 1659 // CHECK2-NEXT: [[TMP0:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 1660 // CHECK2-NEXT: [[CALL:%.*]] = call float @_ZNKSt7complexIfE4realEv(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[TMP0]]) #[[ATTR8]] 1661 // CHECK2-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 0 1662 // CHECK2-NEXT: [[TMP1:%.*]] = load float, float* [[__RE_]], align 4, !tbaa [[TBAA16:![0-9]+]] 1663 // CHECK2-NEXT: [[ADD:%.*]] = fadd float [[TMP1]], [[CALL]] 1664 // CHECK2-NEXT: store float [[ADD]], float* [[__RE_]], align 4, !tbaa [[TBAA16]] 1665 // CHECK2-NEXT: [[TMP2:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 1666 // CHECK2-NEXT: [[CALL2:%.*]] = call float @_ZNKSt7complexIfE4imagEv(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[TMP2]]) #[[ATTR8]] 1667 // CHECK2-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 1 1668 // CHECK2-NEXT: [[TMP3:%.*]] = load float, float* [[__IM_]], align 4, !tbaa [[TBAA18:![0-9]+]] 1669 // CHECK2-NEXT: [[ADD3:%.*]] = fadd float [[TMP3]], [[CALL2]] 1670 // CHECK2-NEXT: store float [[ADD3]], float* [[__IM_]], align 4, !tbaa [[TBAA18]] 1671 // CHECK2-NEXT: ret %"class.std::complex"* [[THIS1]] 1672 // 1673 // 1674 // CHECK2-LABEL: define {{[^@]+}}@_omp_reduction_shuffle_and_reduce_func 1675 // CHECK2-SAME: (i8* [[TMP0:%.*]], i16 signext [[TMP1:%.*]], i16 signext [[TMP2:%.*]], i16 signext [[TMP3:%.*]]) #[[ATTR0]] { 1676 // CHECK2-NEXT: entry: 1677 // CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 1678 // CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca i16, align 2 1679 // CHECK2-NEXT: [[DOTADDR2:%.*]] = alloca i16, align 2 1680 // CHECK2-NEXT: [[DOTADDR3:%.*]] = alloca i16, align 2 1681 // CHECK2-NEXT: [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST:%.*]] = alloca [1 x i8*], align 8 1682 // CHECK2-NEXT: [[DOTOMP_REDUCTION_ELEMENT:%.*]] = alloca %"class.std::complex", align 4 1683 // CHECK2-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 1684 // CHECK2-NEXT: store i16 [[TMP1]], i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19:![0-9]+]] 1685 // CHECK2-NEXT: store i16 [[TMP2]], i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 1686 // CHECK2-NEXT: store i16 [[TMP3]], i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 1687 // CHECK2-NEXT: [[TMP4:%.*]] = load i8*, i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 1688 // CHECK2-NEXT: [[TMP5:%.*]] = bitcast i8* [[TMP4]] to [1 x i8*]* 1689 // CHECK2-NEXT: [[TMP6:%.*]] = load i16, i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19]] 1690 // CHECK2-NEXT: [[TMP7:%.*]] = load i16, i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 1691 // CHECK2-NEXT: [[TMP8:%.*]] = load i16, i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 1692 // CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 1693 // CHECK2-NEXT: [[TMP10:%.*]] = load i8*, i8** [[TMP9]], align 8 1694 // CHECK2-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 1695 // CHECK2-NEXT: [[TMP12:%.*]] = bitcast i8* [[TMP10]] to %"class.std::complex"* 1696 // CHECK2-NEXT: [[TMP13:%.*]] = getelementptr %"class.std::complex", %"class.std::complex"* [[TMP12]], i64 1 1697 // CHECK2-NEXT: [[TMP14:%.*]] = bitcast %"class.std::complex"* [[TMP13]] to i8* 1698 // CHECK2-NEXT: [[TMP15:%.*]] = bitcast %"class.std::complex"* [[TMP12]] to i64* 1699 // CHECK2-NEXT: [[TMP16:%.*]] = bitcast %"class.std::complex"* [[DOTOMP_REDUCTION_ELEMENT]] to i64* 1700 // CHECK2-NEXT: [[TMP17:%.*]] = load i64, i64* [[TMP15]], align 4 1701 // CHECK2-NEXT: [[NVPTX_WARP_SIZE:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize() 1702 // CHECK2-NEXT: [[TMP18:%.*]] = trunc i32 [[NVPTX_WARP_SIZE]] to i16 1703 // CHECK2-NEXT: [[TMP19:%.*]] = call i64 @__kmpc_shuffle_int64(i64 [[TMP17]], i16 [[TMP7]], i16 [[TMP18]]) 1704 // CHECK2-NEXT: store i64 [[TMP19]], i64* [[TMP16]], align 4 1705 // CHECK2-NEXT: [[TMP20:%.*]] = getelementptr i64, i64* [[TMP15]], i64 1 1706 // CHECK2-NEXT: [[TMP21:%.*]] = getelementptr i64, i64* [[TMP16]], i64 1 1707 // CHECK2-NEXT: [[TMP22:%.*]] = bitcast %"class.std::complex"* [[DOTOMP_REDUCTION_ELEMENT]] to i8* 1708 // CHECK2-NEXT: store i8* [[TMP22]], i8** [[TMP11]], align 8, !tbaa [[TBAA12]] 1709 // CHECK2-NEXT: [[TMP23:%.*]] = icmp eq i16 [[TMP8]], 0 1710 // CHECK2-NEXT: [[TMP24:%.*]] = icmp eq i16 [[TMP8]], 1 1711 // CHECK2-NEXT: [[TMP25:%.*]] = icmp ult i16 [[TMP6]], [[TMP7]] 1712 // CHECK2-NEXT: [[TMP26:%.*]] = and i1 [[TMP24]], [[TMP25]] 1713 // CHECK2-NEXT: [[TMP27:%.*]] = icmp eq i16 [[TMP8]], 2 1714 // CHECK2-NEXT: [[TMP28:%.*]] = and i16 [[TMP6]], 1 1715 // CHECK2-NEXT: [[TMP29:%.*]] = icmp eq i16 [[TMP28]], 0 1716 // CHECK2-NEXT: [[TMP30:%.*]] = and i1 [[TMP27]], [[TMP29]] 1717 // CHECK2-NEXT: [[TMP31:%.*]] = icmp sgt i16 [[TMP7]], 0 1718 // CHECK2-NEXT: [[TMP32:%.*]] = and i1 [[TMP30]], [[TMP31]] 1719 // CHECK2-NEXT: [[TMP33:%.*]] = or i1 [[TMP23]], [[TMP26]] 1720 // CHECK2-NEXT: [[TMP34:%.*]] = or i1 [[TMP33]], [[TMP32]] 1721 // CHECK2-NEXT: br i1 [[TMP34]], label [[THEN:%.*]], label [[ELSE:%.*]] 1722 // CHECK2: then: 1723 // CHECK2-NEXT: [[TMP35:%.*]] = bitcast [1 x i8*]* [[TMP5]] to i8* 1724 // CHECK2-NEXT: [[TMP36:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]] to i8* 1725 // CHECK2-NEXT: call void @"_omp$reduction$reduction_func"(i8* [[TMP35]], i8* [[TMP36]]) #[[ATTR1]] 1726 // CHECK2-NEXT: br label [[IFCONT:%.*]] 1727 // CHECK2: else: 1728 // CHECK2-NEXT: br label [[IFCONT]] 1729 // CHECK2: ifcont: 1730 // CHECK2-NEXT: [[TMP37:%.*]] = icmp eq i16 [[TMP8]], 1 1731 // CHECK2-NEXT: [[TMP38:%.*]] = icmp uge i16 [[TMP6]], [[TMP7]] 1732 // CHECK2-NEXT: [[TMP39:%.*]] = and i1 [[TMP37]], [[TMP38]] 1733 // CHECK2-NEXT: br i1 [[TMP39]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 1734 // CHECK2: then4: 1735 // CHECK2-NEXT: [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 1736 // CHECK2-NEXT: [[TMP41:%.*]] = load i8*, i8** [[TMP40]], align 8 1737 // CHECK2-NEXT: [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 1738 // CHECK2-NEXT: [[TMP43:%.*]] = load i8*, i8** [[TMP42]], align 8 1739 // CHECK2-NEXT: [[TMP44:%.*]] = bitcast i8* [[TMP41]] to %"class.std::complex"* 1740 // CHECK2-NEXT: [[TMP45:%.*]] = bitcast i8* [[TMP43]] to %"class.std::complex"* 1741 // CHECK2-NEXT: [[TMP46:%.*]] = bitcast %"class.std::complex"* [[TMP45]] to i8* 1742 // CHECK2-NEXT: [[TMP47:%.*]] = bitcast %"class.std::complex"* [[TMP44]] to i8* 1743 // CHECK2-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP46]], i8* align 4 [[TMP47]], i64 8, i1 false), !tbaa.struct !21 1744 // CHECK2-NEXT: br label [[IFCONT6:%.*]] 1745 // CHECK2: else5: 1746 // CHECK2-NEXT: br label [[IFCONT6]] 1747 // CHECK2: ifcont6: 1748 // CHECK2-NEXT: ret void 1749 // 1750 // 1751 // CHECK2-LABEL: define {{[^@]+}}@_omp_reduction_inter_warp_copy_func 1752 // CHECK2-SAME: (i8* [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 1753 // CHECK2-NEXT: entry: 1754 // CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 1755 // CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 1756 // CHECK2-NEXT: [[DOTCNT_ADDR:%.*]] = alloca i32, align 4 1757 // CHECK2-NEXT: [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 1758 // CHECK2-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 1759 // CHECK2-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 1760 // CHECK2-NEXT: [[NVPTX_TID:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 1761 // CHECK2-NEXT: [[NVPTX_TID2:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 1762 // CHECK2-NEXT: [[NVPTX_LANE_ID:%.*]] = and i32 [[NVPTX_TID2]], 31 1763 // CHECK2-NEXT: [[NVPTX_TID3:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 1764 // CHECK2-NEXT: [[NVPTX_WARP_ID:%.*]] = ashr i32 [[NVPTX_TID3]], 5 1765 // CHECK2-NEXT: [[TMP3:%.*]] = load i8*, i8** [[DOTADDR]], align 8 1766 // CHECK2-NEXT: [[TMP4:%.*]] = bitcast i8* [[TMP3]] to [1 x i8*]* 1767 // CHECK2-NEXT: store i32 0, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 1768 // CHECK2-NEXT: br label [[PRECOND:%.*]] 1769 // CHECK2: precond: 1770 // CHECK2-NEXT: [[TMP5:%.*]] = load i32, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 1771 // CHECK2-NEXT: [[TMP6:%.*]] = icmp ult i32 [[TMP5]], 2 1772 // CHECK2-NEXT: br i1 [[TMP6]], label [[BODY:%.*]], label [[EXIT:%.*]] 1773 // CHECK2: body: 1774 // CHECK2-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4:[0-9]+]], i32 [[TMP2]]) 1775 // CHECK2-NEXT: [[WARP_MASTER:%.*]] = icmp eq i32 [[NVPTX_LANE_ID]], 0 1776 // CHECK2-NEXT: br i1 [[WARP_MASTER]], label [[THEN:%.*]], label [[ELSE:%.*]] 1777 // CHECK2: then: 1778 // CHECK2-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 1779 // CHECK2-NEXT: [[TMP8:%.*]] = load i8*, i8** [[TMP7]], align 8, !tbaa [[TBAA12]] 1780 // CHECK2-NEXT: [[TMP9:%.*]] = bitcast i8* [[TMP8]] to i32* 1781 // CHECK2-NEXT: [[TMP10:%.*]] = getelementptr i32, i32* [[TMP9]], i32 [[TMP5]] 1782 // CHECK2-NEXT: [[TMP11:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_WARP_ID]] 1783 // CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP10]], align 4 1784 // CHECK2-NEXT: store volatile i32 [[TMP12]], i32 addrspace(3)* [[TMP11]], align 4 1785 // CHECK2-NEXT: br label [[IFCONT:%.*]] 1786 // CHECK2: else: 1787 // CHECK2-NEXT: br label [[IFCONT]] 1788 // CHECK2: ifcont: 1789 // CHECK2-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4]], i32 [[TMP2]]) 1790 // CHECK2-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 1791 // CHECK2-NEXT: [[IS_ACTIVE_THREAD:%.*]] = icmp ult i32 [[NVPTX_TID]], [[TMP13]] 1792 // CHECK2-NEXT: br i1 [[IS_ACTIVE_THREAD]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 1793 // CHECK2: then4: 1794 // CHECK2-NEXT: [[TMP14:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_TID]] 1795 // CHECK2-NEXT: [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 1796 // CHECK2-NEXT: [[TMP16:%.*]] = load i8*, i8** [[TMP15]], align 8, !tbaa [[TBAA12]] 1797 // CHECK2-NEXT: [[TMP17:%.*]] = bitcast i8* [[TMP16]] to i32* 1798 // CHECK2-NEXT: [[TMP18:%.*]] = getelementptr i32, i32* [[TMP17]], i32 [[TMP5]] 1799 // CHECK2-NEXT: [[TMP19:%.*]] = load volatile i32, i32 addrspace(3)* [[TMP14]], align 4, !tbaa [[TBAA8]] 1800 // CHECK2-NEXT: store i32 [[TMP19]], i32* [[TMP18]], align 4, !tbaa [[TBAA8]] 1801 // CHECK2-NEXT: br label [[IFCONT6:%.*]] 1802 // CHECK2: else5: 1803 // CHECK2-NEXT: br label [[IFCONT6]] 1804 // CHECK2: ifcont6: 1805 // CHECK2-NEXT: [[TMP20:%.*]] = add nsw i32 [[TMP5]], 1 1806 // CHECK2-NEXT: store i32 [[TMP20]], i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 1807 // CHECK2-NEXT: br label [[PRECOND]] 1808 // CHECK2: exit: 1809 // CHECK2-NEXT: ret void 1810 // 1811 // 1812 // CHECK2-LABEL: define {{[^@]+}}@__omp_outlined__1_wrapper 1813 // CHECK2-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 1814 // CHECK2-NEXT: entry: 1815 // CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i16, align 2 1816 // CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 1817 // CHECK2-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 1818 // CHECK2-NEXT: [[GLOBAL_ARGS:%.*]] = alloca i8**, align 8 1819 // CHECK2-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 1820 // CHECK2-NEXT: store i16 [[TMP0]], i16* [[DOTADDR]], align 2, !tbaa [[TBAA19]] 1821 // CHECK2-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 1822 // CHECK2-NEXT: call void @__kmpc_get_shared_variables(i8*** [[GLOBAL_ARGS]]) 1823 // CHECK2-NEXT: [[TMP2:%.*]] = load i8**, i8*** [[GLOBAL_ARGS]], align 8 1824 // CHECK2-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 0 1825 // CHECK2-NEXT: [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32** 1826 // CHECK2-NEXT: [[TMP5:%.*]] = load i32*, i32** [[TMP4]], align 8, !tbaa [[TBAA12]] 1827 // CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 1 1828 // CHECK2-NEXT: [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i32** 1829 // CHECK2-NEXT: [[TMP8:%.*]] = load i32*, i32** [[TMP7]], align 8, !tbaa [[TBAA12]] 1830 // CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 2 1831 // CHECK2-NEXT: [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %"class.std::complex"** 1832 // CHECK2-NEXT: [[TMP11:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[TMP10]], align 8, !tbaa [[TBAA12]] 1833 // CHECK2-NEXT: call void @__omp_outlined__1(i32* [[DOTADDR1]], i32* [[DOTZERO_ADDR]], i32* [[TMP5]], i32* [[TMP8]], %"class.std::complex"* [[TMP11]]) #[[ATTR1]] 1834 // CHECK2-NEXT: ret void 1835 // 1836 // 1837 // CHECK2-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z17complex_reductionIdEvv_l19 1838 // CHECK2-SAME: () #[[ATTR0]] { 1839 // CHECK2-NEXT: entry: 1840 // CHECK2-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 1841 // CHECK2-NEXT: [[DOTTHREADID_TEMP_:%.*]] = alloca i32, align 4 1842 // CHECK2-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 1843 // CHECK2-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* @[[GLOB1]], i1 false, i1 true, i1 true) 1844 // CHECK2-NEXT: [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1 1845 // CHECK2-NEXT: br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]] 1846 // CHECK2: user_code.entry: 1847 // CHECK2-NEXT: [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 1848 // CHECK2-NEXT: store i32 [[TMP1]], i32* [[DOTTHREADID_TEMP_]], align 4, !tbaa [[TBAA8]] 1849 // CHECK2-NEXT: call void @__omp_outlined__2(i32* [[DOTTHREADID_TEMP_]], i32* [[DOTZERO_ADDR]]) #[[ATTR1]] 1850 // CHECK2-NEXT: call void @__kmpc_target_deinit(%struct.ident_t* @[[GLOB1]], i1 false, i1 true) 1851 // CHECK2-NEXT: ret void 1852 // CHECK2: worker.exit: 1853 // CHECK2-NEXT: ret void 1854 // 1855 // 1856 // CHECK2-LABEL: define {{[^@]+}}@__omp_outlined__2 1857 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR0]] { 1858 // CHECK2-NEXT: entry: 1859 // CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 1860 // CHECK2-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 1861 // CHECK2-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 1862 // CHECK2-NEXT: [[TMP:%.*]] = alloca i32, align 4 1863 // CHECK2-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 1864 // CHECK2-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 1865 // CHECK2-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 1866 // CHECK2-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 1867 // CHECK2-NEXT: [[IB:%.*]] = alloca i32, align 4 1868 // CHECK2-NEXT: [[REF_TMP:%.*]] = alloca double, align 8 1869 // CHECK2-NEXT: [[REF_TMP2:%.*]] = alloca double, align 8 1870 // CHECK2-NEXT: [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x i8*], align 8 1871 // CHECK2-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12]] 1872 // CHECK2-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 1873 // CHECK2-NEXT: [[ISTART:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 1874 // CHECK2-NEXT: [[ISTART_ON_STACK:%.*]] = bitcast i8* [[ISTART]] to i32* 1875 // CHECK2-NEXT: [[IEND:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 1876 // CHECK2-NEXT: [[IEND_ON_STACK:%.*]] = bitcast i8* [[IEND]] to i32* 1877 // CHECK2-NEXT: [[PARTIAL_SUM:%.*]] = call i8* @__kmpc_alloc_shared(i64 16) 1878 // CHECK2-NEXT: [[PARTIAL_SUM_ON_STACK:%.*]] = bitcast i8* [[PARTIAL_SUM]] to %"class.std::complex.0"* 1879 // CHECK2-NEXT: [[TMP0:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 1880 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP0]]) #[[ATTR1]] 1881 // CHECK2-NEXT: [[TMP1:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 1882 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP1]]) #[[ATTR1]] 1883 // CHECK2-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 1884 // CHECK2-NEXT: [[TMP2:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 1885 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP2]]) #[[ATTR1]] 1886 // CHECK2-NEXT: store i32 99, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1887 // CHECK2-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 1888 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 1889 // CHECK2-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 1890 // CHECK2-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 1891 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 1892 // CHECK2-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 1893 // CHECK2-NEXT: [[TMP5:%.*]] = bitcast i32* [[IB]] to i8* 1894 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP5]]) #[[ATTR1]] 1895 // CHECK2-NEXT: [[TMP6:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 1896 // CHECK2-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP6]], align 4, !tbaa [[TBAA8]] 1897 // CHECK2-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP7]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 1898 // CHECK2-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1899 // CHECK2-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP8]], 99 1900 // CHECK2-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 1901 // CHECK2: cond.true: 1902 // CHECK2-NEXT: br label [[COND_END:%.*]] 1903 // CHECK2: cond.false: 1904 // CHECK2-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1905 // CHECK2-NEXT: br label [[COND_END]] 1906 // CHECK2: cond.end: 1907 // CHECK2-NEXT: [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP9]], [[COND_FALSE]] ] 1908 // CHECK2-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1909 // CHECK2-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 1910 // CHECK2-NEXT: store i32 [[TMP10]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1911 // CHECK2-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 1912 // CHECK2: omp.inner.for.cond: 1913 // CHECK2-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1914 // CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 1915 // CHECK2-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP11]], [[TMP12]] 1916 // CHECK2-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 1917 // CHECK2: omp.inner.for.cond.cleanup: 1918 // CHECK2-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 1919 // CHECK2: omp.inner.for.body: 1920 // CHECK2-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1921 // CHECK2-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP13]], 1 1922 // CHECK2-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]] 1923 // CHECK2-NEXT: store i32 [[ADD]], i32* [[IB]], align 4, !tbaa [[TBAA8]] 1924 // CHECK2-NEXT: [[TMP14:%.*]] = bitcast double* [[REF_TMP]] to i8* 1925 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP14]]) #[[ATTR1]] 1926 // CHECK2-NEXT: store double 0.000000e+00, double* [[REF_TMP]], align 8, !tbaa [[TBAA22:![0-9]+]] 1927 // CHECK2-NEXT: [[TMP15:%.*]] = bitcast double* [[REF_TMP2]] to i8* 1928 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP15]]) #[[ATTR1]] 1929 // CHECK2-NEXT: store double 0.000000e+00, double* [[REF_TMP2]], align 8, !tbaa [[TBAA22]] 1930 // CHECK2-NEXT: call void @_ZNSt7complexIdEC1ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM_ON_STACK]], double* nonnull align 8 dereferenceable(8) [[REF_TMP]], double* nonnull align 8 dereferenceable(8) [[REF_TMP2]]) #[[ATTR8]] 1931 // CHECK2-NEXT: [[TMP16:%.*]] = bitcast double* [[REF_TMP2]] to i8* 1932 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP16]]) #[[ATTR1]] 1933 // CHECK2-NEXT: [[TMP17:%.*]] = bitcast double* [[REF_TMP]] to i8* 1934 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP17]]) #[[ATTR1]] 1935 // CHECK2-NEXT: [[TMP18:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 1936 // CHECK2-NEXT: [[MUL3:%.*]] = mul nsw i32 [[TMP18]], 4 1937 // CHECK2-NEXT: store i32 [[MUL3]], i32* [[ISTART_ON_STACK]], align 4, !tbaa [[TBAA8]] 1938 // CHECK2-NEXT: [[TMP19:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 1939 // CHECK2-NEXT: [[ADD4:%.*]] = add nsw i32 [[TMP19]], 1 1940 // CHECK2-NEXT: [[MUL5:%.*]] = mul nsw i32 [[ADD4]], 4 1941 // CHECK2-NEXT: store i32 [[MUL5]], i32* [[IEND_ON_STACK]], align 4, !tbaa [[TBAA8]] 1942 // CHECK2-NEXT: [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 0 1943 // CHECK2-NEXT: [[TMP21:%.*]] = bitcast i32* [[ISTART_ON_STACK]] to i8* 1944 // CHECK2-NEXT: store i8* [[TMP21]], i8** [[TMP20]], align 8, !tbaa [[TBAA12]] 1945 // CHECK2-NEXT: [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 1 1946 // CHECK2-NEXT: [[TMP23:%.*]] = bitcast i32* [[IEND_ON_STACK]] to i8* 1947 // CHECK2-NEXT: store i8* [[TMP23]], i8** [[TMP22]], align 8, !tbaa [[TBAA12]] 1948 // CHECK2-NEXT: [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 2 1949 // CHECK2-NEXT: [[TMP25:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM_ON_STACK]] to i8* 1950 // CHECK2-NEXT: store i8* [[TMP25]], i8** [[TMP24]], align 8, !tbaa [[TBAA12]] 1951 // CHECK2-NEXT: [[TMP26:%.*]] = bitcast [3 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8** 1952 // CHECK2-NEXT: call void @__kmpc_parallel_51(%struct.ident_t* @[[GLOB1]], i32 [[TMP7]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*, i32*, i32*, %"class.std::complex.0"*)* @__omp_outlined__3 to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined__3_wrapper to i8*), i8** [[TMP26]], i64 3) 1953 // CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 1954 // CHECK2: omp.body.continue: 1955 // CHECK2-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 1956 // CHECK2: omp.inner.for.inc: 1957 // CHECK2-NEXT: [[TMP27:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1958 // CHECK2-NEXT: [[ADD6:%.*]] = add nsw i32 [[TMP27]], 1 1959 // CHECK2-NEXT: store i32 [[ADD6]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 1960 // CHECK2-NEXT: br label [[OMP_INNER_FOR_COND]] 1961 // CHECK2: omp.inner.for.end: 1962 // CHECK2-NEXT: br label [[OMP_LOOP_EXIT:%.*]] 1963 // CHECK2: omp.loop.exit: 1964 // CHECK2-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB2]], i32 [[TMP7]]) 1965 // CHECK2-NEXT: [[TMP28:%.*]] = bitcast i32* [[IB]] to i8* 1966 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP28]]) #[[ATTR1]] 1967 // CHECK2-NEXT: [[TMP29:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 1968 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP29]]) #[[ATTR1]] 1969 // CHECK2-NEXT: [[TMP30:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 1970 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP30]]) #[[ATTR1]] 1971 // CHECK2-NEXT: [[TMP31:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 1972 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP31]]) #[[ATTR1]] 1973 // CHECK2-NEXT: [[TMP32:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 1974 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP32]]) #[[ATTR1]] 1975 // CHECK2-NEXT: [[TMP33:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 1976 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP33]]) #[[ATTR1]] 1977 // CHECK2-NEXT: call void @__kmpc_free_shared(i8* [[PARTIAL_SUM]], i64 16) 1978 // CHECK2-NEXT: call void @__kmpc_free_shared(i8* [[IEND]], i64 4) 1979 // CHECK2-NEXT: call void @__kmpc_free_shared(i8* [[ISTART]], i64 4) 1980 // CHECK2-NEXT: ret void 1981 // 1982 // 1983 // CHECK2-LABEL: define {{[^@]+}}@_ZNSt7complexIdEC1ERKdS2_ 1984 // CHECK2-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]], double* nonnull align 8 dereferenceable(8) [[__RE:%.*]], double* nonnull align 8 dereferenceable(8) [[__IM:%.*]]) unnamed_addr #[[ATTR3]] comdat align 2 { 1985 // CHECK2-NEXT: entry: 1986 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 1987 // CHECK2-NEXT: [[__RE_ADDR:%.*]] = alloca double*, align 8 1988 // CHECK2-NEXT: [[__IM_ADDR:%.*]] = alloca double*, align 8 1989 // CHECK2-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 1990 // CHECK2-NEXT: store double* [[__RE]], double** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 1991 // CHECK2-NEXT: store double* [[__IM]], double** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 1992 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 1993 // CHECK2-NEXT: [[TMP0:%.*]] = load double*, double** [[__RE_ADDR]], align 8 1994 // CHECK2-NEXT: [[TMP1:%.*]] = load double*, double** [[__IM_ADDR]], align 8 1995 // CHECK2-NEXT: call void @_ZNSt7complexIdEC2ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS1]], double* nonnull align 8 dereferenceable(8) [[TMP0]], double* nonnull align 8 dereferenceable(8) [[TMP1]]) #[[ATTR8]] 1996 // CHECK2-NEXT: ret void 1997 // 1998 // 1999 // CHECK2-LABEL: define {{[^@]+}}@__omp_outlined__3 2000 // CHECK2-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ISTART:%.*]], i32* nonnull align 4 dereferenceable(4) [[IEND:%.*]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM:%.*]]) #[[ATTR0]] { 2001 // CHECK2-NEXT: entry: 2002 // CHECK2-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 2003 // CHECK2-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 2004 // CHECK2-NEXT: [[ISTART_ADDR:%.*]] = alloca i32*, align 8 2005 // CHECK2-NEXT: [[IEND_ADDR:%.*]] = alloca i32*, align 8 2006 // CHECK2-NEXT: [[PARTIAL_SUM_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 2007 // CHECK2-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 2008 // CHECK2-NEXT: [[TMP:%.*]] = alloca i32, align 4 2009 // CHECK2-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4 2010 // CHECK2-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4 2011 // CHECK2-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4 2012 // CHECK2-NEXT: [[I:%.*]] = alloca i32, align 4 2013 // CHECK2-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 2014 // CHECK2-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 2015 // CHECK2-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 2016 // CHECK2-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 2017 // CHECK2-NEXT: [[PARTIAL_SUM5:%.*]] = alloca %"class.std::complex.0", align 8 2018 // CHECK2-NEXT: [[REF_TMP:%.*]] = alloca double, align 8 2019 // CHECK2-NEXT: [[REF_TMP6:%.*]] = alloca double, align 8 2020 // CHECK2-NEXT: [[I7:%.*]] = alloca i32, align 4 2021 // CHECK2-NEXT: [[REF_TMP14:%.*]] = alloca %"class.std::complex.0", align 8 2022 // CHECK2-NEXT: [[REF_TMP15:%.*]] = alloca double, align 8 2023 // CHECK2-NEXT: [[REF_TMP16:%.*]] = alloca double, align 8 2024 // CHECK2-NEXT: [[DOTOMP_REDUCTION_RED_LIST:%.*]] = alloca [1 x i8*], align 8 2025 // CHECK2-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12]] 2026 // CHECK2-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 2027 // CHECK2-NEXT: store i32* [[ISTART]], i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 2028 // CHECK2-NEXT: store i32* [[IEND]], i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 2029 // CHECK2-NEXT: store %"class.std::complex.0"* [[PARTIAL_SUM]], %"class.std::complex.0"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 2030 // CHECK2-NEXT: [[TMP0:%.*]] = load i32*, i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 2031 // CHECK2-NEXT: [[TMP1:%.*]] = load i32*, i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 2032 // CHECK2-NEXT: [[TMP2:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 2033 // CHECK2-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 2034 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 2035 // CHECK2-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 2036 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 2037 // CHECK2-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP0]], align 4, !tbaa [[TBAA8]] 2038 // CHECK2-NEXT: store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2039 // CHECK2-NEXT: [[TMP6:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 2040 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP6]]) #[[ATTR1]] 2041 // CHECK2-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP1]], align 4, !tbaa [[TBAA8]] 2042 // CHECK2-NEXT: store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 2043 // CHECK2-NEXT: [[TMP8:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 2044 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP8]]) #[[ATTR1]] 2045 // CHECK2-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 2046 // CHECK2-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2047 // CHECK2-NEXT: [[SUB:%.*]] = sub i32 [[TMP9]], [[TMP10]] 2048 // CHECK2-NEXT: [[SUB3:%.*]] = sub i32 [[SUB]], 1 2049 // CHECK2-NEXT: [[ADD:%.*]] = add i32 [[SUB3]], 1 2050 // CHECK2-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], 1 2051 // CHECK2-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1 2052 // CHECK2-NEXT: store i32 [[SUB4]], i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 2053 // CHECK2-NEXT: [[TMP11:%.*]] = bitcast i32* [[I]] to i8* 2054 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP11]]) #[[ATTR1]] 2055 // CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2056 // CHECK2-NEXT: store i32 [[TMP12]], i32* [[I]], align 4, !tbaa [[TBAA8]] 2057 // CHECK2-NEXT: [[TMP13:%.*]] = bitcast i32* [[I]] to i8* 2058 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP13]]) #[[ATTR1]] 2059 // CHECK2-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2060 // CHECK2-NEXT: [[TMP15:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 2061 // CHECK2-NEXT: [[CMP:%.*]] = icmp slt i32 [[TMP14]], [[TMP15]] 2062 // CHECK2-NEXT: br i1 [[CMP]], label [[OMP_PRECOND_THEN:%.*]], label [[OMP_PRECOND_END:%.*]] 2063 // CHECK2: omp.precond.then: 2064 // CHECK2-NEXT: [[TMP16:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 2065 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP16]]) #[[ATTR1]] 2066 // CHECK2-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2067 // CHECK2-NEXT: [[TMP17:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 2068 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP17]]) #[[ATTR1]] 2069 // CHECK2-NEXT: [[TMP18:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 2070 // CHECK2-NEXT: store i32 [[TMP18]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2071 // CHECK2-NEXT: [[TMP19:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 2072 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP19]]) #[[ATTR1]] 2073 // CHECK2-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 2074 // CHECK2-NEXT: [[TMP20:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 2075 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP20]]) #[[ATTR1]] 2076 // CHECK2-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 2077 // CHECK2-NEXT: [[TMP21:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM5]] to i8* 2078 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[TMP21]]) #[[ATTR1]] 2079 // CHECK2-NEXT: [[TMP22:%.*]] = bitcast double* [[REF_TMP]] to i8* 2080 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP22]]) #[[ATTR1]] 2081 // CHECK2-NEXT: store double 0.000000e+00, double* [[REF_TMP]], align 8, !tbaa [[TBAA22]] 2082 // CHECK2-NEXT: [[TMP23:%.*]] = bitcast double* [[REF_TMP6]] to i8* 2083 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP23]]) #[[ATTR1]] 2084 // CHECK2-NEXT: store double 0.000000e+00, double* [[REF_TMP6]], align 8, !tbaa [[TBAA22]] 2085 // CHECK2-NEXT: call void @_ZNSt7complexIdEC1ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM5]], double* nonnull align 8 dereferenceable(8) [[REF_TMP]], double* nonnull align 8 dereferenceable(8) [[REF_TMP6]]) #[[ATTR8]] 2086 // CHECK2-NEXT: [[TMP24:%.*]] = bitcast double* [[REF_TMP6]] to i8* 2087 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP24]]) #[[ATTR1]] 2088 // CHECK2-NEXT: [[TMP25:%.*]] = bitcast double* [[REF_TMP]] to i8* 2089 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP25]]) #[[ATTR1]] 2090 // CHECK2-NEXT: [[TMP26:%.*]] = bitcast i32* [[I7]] to i8* 2091 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP26]]) #[[ATTR1]] 2092 // CHECK2-NEXT: [[TMP27:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 2093 // CHECK2-NEXT: [[TMP28:%.*]] = load i32, i32* [[TMP27]], align 4, !tbaa [[TBAA8]] 2094 // CHECK2-NEXT: call void @__kmpc_for_static_init_4u(%struct.ident_t* @[[GLOB3]], i32 [[TMP28]], i32 33, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 2095 // CHECK2-NEXT: br label [[OMP_DISPATCH_COND:%.*]] 2096 // CHECK2: omp.dispatch.cond: 2097 // CHECK2-NEXT: [[TMP29:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2098 // CHECK2-NEXT: [[TMP30:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 2099 // CHECK2-NEXT: [[CMP8:%.*]] = icmp ugt i32 [[TMP29]], [[TMP30]] 2100 // CHECK2-NEXT: br i1 [[CMP8]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 2101 // CHECK2: cond.true: 2102 // CHECK2-NEXT: [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 2103 // CHECK2-NEXT: br label [[COND_END:%.*]] 2104 // CHECK2: cond.false: 2105 // CHECK2-NEXT: [[TMP32:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2106 // CHECK2-NEXT: br label [[COND_END]] 2107 // CHECK2: cond.end: 2108 // CHECK2-NEXT: [[COND:%.*]] = phi i32 [ [[TMP31]], [[COND_TRUE]] ], [ [[TMP32]], [[COND_FALSE]] ] 2109 // CHECK2-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2110 // CHECK2-NEXT: [[TMP33:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2111 // CHECK2-NEXT: store i32 [[TMP33]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2112 // CHECK2-NEXT: [[TMP34:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2113 // CHECK2-NEXT: [[TMP35:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2114 // CHECK2-NEXT: [[ADD9:%.*]] = add i32 [[TMP35]], 1 2115 // CHECK2-NEXT: [[CMP10:%.*]] = icmp ult i32 [[TMP34]], [[ADD9]] 2116 // CHECK2-NEXT: br i1 [[CMP10]], label [[OMP_DISPATCH_BODY:%.*]], label [[OMP_DISPATCH_CLEANUP:%.*]] 2117 // CHECK2: omp.dispatch.cleanup: 2118 // CHECK2-NEXT: br label [[OMP_DISPATCH_END:%.*]] 2119 // CHECK2: omp.dispatch.body: 2120 // CHECK2-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 2121 // CHECK2: omp.inner.for.cond: 2122 // CHECK2-NEXT: [[TMP36:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2123 // CHECK2-NEXT: [[TMP37:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2124 // CHECK2-NEXT: [[ADD11:%.*]] = add i32 [[TMP37]], 1 2125 // CHECK2-NEXT: [[CMP12:%.*]] = icmp ult i32 [[TMP36]], [[ADD11]] 2126 // CHECK2-NEXT: br i1 [[CMP12]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 2127 // CHECK2: omp.inner.for.cond.cleanup: 2128 // CHECK2-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 2129 // CHECK2: omp.inner.for.body: 2130 // CHECK2-NEXT: [[TMP38:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2131 // CHECK2-NEXT: [[TMP39:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2132 // CHECK2-NEXT: [[MUL:%.*]] = mul i32 [[TMP39]], 1 2133 // CHECK2-NEXT: [[ADD13:%.*]] = add i32 [[TMP38]], [[MUL]] 2134 // CHECK2-NEXT: store i32 [[ADD13]], i32* [[I7]], align 4, !tbaa [[TBAA8]] 2135 // CHECK2-NEXT: [[TMP40:%.*]] = bitcast %"class.std::complex.0"* [[REF_TMP14]] to i8* 2136 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[TMP40]]) #[[ATTR1]] 2137 // CHECK2-NEXT: [[TMP41:%.*]] = bitcast double* [[REF_TMP15]] to i8* 2138 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP41]]) #[[ATTR1]] 2139 // CHECK2-NEXT: [[TMP42:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 2140 // CHECK2-NEXT: [[CONV:%.*]] = sitofp i32 [[TMP42]] to double 2141 // CHECK2-NEXT: store double [[CONV]], double* [[REF_TMP15]], align 8, !tbaa [[TBAA22]] 2142 // CHECK2-NEXT: [[TMP43:%.*]] = bitcast double* [[REF_TMP16]] to i8* 2143 // CHECK2-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP43]]) #[[ATTR1]] 2144 // CHECK2-NEXT: [[TMP44:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 2145 // CHECK2-NEXT: [[CONV17:%.*]] = sitofp i32 [[TMP44]] to double 2146 // CHECK2-NEXT: store double [[CONV17]], double* [[REF_TMP16]], align 8, !tbaa [[TBAA22]] 2147 // CHECK2-NEXT: call void @_ZNSt7complexIdEC1ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[REF_TMP14]], double* nonnull align 8 dereferenceable(8) [[REF_TMP15]], double* nonnull align 8 dereferenceable(8) [[REF_TMP16]]) #[[ATTR8]] 2148 // CHECK2-NEXT: [[CALL:%.*]] = call nonnull align 8 dereferenceable(16) %"class.std::complex.0"* @_ZNSt7complexIdEpLIdEERS0_RKS_IT_E(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM5]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[REF_TMP14]]) #[[ATTR8]] 2149 // CHECK2-NEXT: [[TMP45:%.*]] = bitcast double* [[REF_TMP16]] to i8* 2150 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP45]]) #[[ATTR1]] 2151 // CHECK2-NEXT: [[TMP46:%.*]] = bitcast double* [[REF_TMP15]] to i8* 2152 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP46]]) #[[ATTR1]] 2153 // CHECK2-NEXT: [[TMP47:%.*]] = bitcast %"class.std::complex.0"* [[REF_TMP14]] to i8* 2154 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[TMP47]]) #[[ATTR1]] 2155 // CHECK2-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 2156 // CHECK2: omp.body.continue: 2157 // CHECK2-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 2158 // CHECK2: omp.inner.for.inc: 2159 // CHECK2-NEXT: [[TMP48:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2160 // CHECK2-NEXT: [[ADD18:%.*]] = add i32 [[TMP48]], 1 2161 // CHECK2-NEXT: store i32 [[ADD18]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2162 // CHECK2-NEXT: br label [[OMP_INNER_FOR_COND]] 2163 // CHECK2: omp.inner.for.end: 2164 // CHECK2-NEXT: br label [[OMP_DISPATCH_INC:%.*]] 2165 // CHECK2: omp.dispatch.inc: 2166 // CHECK2-NEXT: [[TMP49:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2167 // CHECK2-NEXT: [[TMP50:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 2168 // CHECK2-NEXT: [[ADD19:%.*]] = add i32 [[TMP49]], [[TMP50]] 2169 // CHECK2-NEXT: store i32 [[ADD19]], i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2170 // CHECK2-NEXT: [[TMP51:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2171 // CHECK2-NEXT: [[TMP52:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 2172 // CHECK2-NEXT: [[ADD20:%.*]] = add i32 [[TMP51]], [[TMP52]] 2173 // CHECK2-NEXT: store i32 [[ADD20]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2174 // CHECK2-NEXT: br label [[OMP_DISPATCH_COND]] 2175 // CHECK2: omp.dispatch.end: 2176 // CHECK2-NEXT: [[TMP53:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 2177 // CHECK2-NEXT: [[TMP54:%.*]] = load i32, i32* [[TMP53]], align 4, !tbaa [[TBAA8]] 2178 // CHECK2-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB3]], i32 [[TMP54]]) 2179 // CHECK2-NEXT: [[TMP55:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 2180 // CHECK2-NEXT: [[TMP56:%.*]] = load i32, i32* [[TMP55]], align 4, !tbaa [[TBAA8]] 2181 // CHECK2-NEXT: [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]], i64 0, i64 0 2182 // CHECK2-NEXT: [[TMP58:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM5]] to i8* 2183 // CHECK2-NEXT: store i8* [[TMP58]], i8** [[TMP57]], align 8 2184 // CHECK2-NEXT: [[TMP59:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]] to i8* 2185 // CHECK2-NEXT: [[TMP60:%.*]] = call i32 @__kmpc_nvptx_parallel_reduce_nowait_v2(%struct.ident_t* @[[GLOB1]], i32 [[TMP56]], i32 1, i64 8, i8* [[TMP59]], void (i8*, i16, i16, i16)* @_omp_reduction_shuffle_and_reduce_func5, void (i8*, i32)* @_omp_reduction_inter_warp_copy_func6) 2186 // CHECK2-NEXT: [[TMP61:%.*]] = icmp eq i32 [[TMP60]], 1 2187 // CHECK2-NEXT: br i1 [[TMP61]], label [[DOTOMP_REDUCTION_THEN:%.*]], label [[DOTOMP_REDUCTION_DONE:%.*]] 2188 // CHECK2: .omp.reduction.then: 2189 // CHECK2-NEXT: [[CALL21:%.*]] = call nonnull align 8 dereferenceable(16) %"class.std::complex.0"* @_ZNSt7complexIdEpLIdEERS0_RKS_IT_E(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[TMP2]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM5]]) #[[ATTR8]] 2190 // CHECK2-NEXT: call void @__kmpc_nvptx_end_reduce_nowait(i32 [[TMP56]]) 2191 // CHECK2-NEXT: br label [[DOTOMP_REDUCTION_DONE]] 2192 // CHECK2: .omp.reduction.done: 2193 // CHECK2-NEXT: [[TMP62:%.*]] = bitcast i32* [[I7]] to i8* 2194 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP62]]) #[[ATTR1]] 2195 // CHECK2-NEXT: [[TMP63:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM5]] to i8* 2196 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[TMP63]]) #[[ATTR1]] 2197 // CHECK2-NEXT: [[TMP64:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 2198 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP64]]) #[[ATTR1]] 2199 // CHECK2-NEXT: [[TMP65:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 2200 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP65]]) #[[ATTR1]] 2201 // CHECK2-NEXT: [[TMP66:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 2202 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP66]]) #[[ATTR1]] 2203 // CHECK2-NEXT: [[TMP67:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 2204 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP67]]) #[[ATTR1]] 2205 // CHECK2-NEXT: br label [[OMP_PRECOND_END]] 2206 // CHECK2: omp.precond.end: 2207 // CHECK2-NEXT: [[TMP68:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 2208 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP68]]) #[[ATTR1]] 2209 // CHECK2-NEXT: [[TMP69:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 2210 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP69]]) #[[ATTR1]] 2211 // CHECK2-NEXT: [[TMP70:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 2212 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP70]]) #[[ATTR1]] 2213 // CHECK2-NEXT: [[TMP71:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 2214 // CHECK2-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP71]]) #[[ATTR1]] 2215 // CHECK2-NEXT: ret void 2216 // 2217 // 2218 // CHECK2-LABEL: define {{[^@]+}}@_ZNSt7complexIdEpLIdEERS0_RKS_IT_E 2219 // CHECK2-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[__C:%.*]]) #[[ATTR4]] comdat align 2 { 2220 // CHECK2-NEXT: entry: 2221 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 2222 // CHECK2-NEXT: [[__C_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 2223 // CHECK2-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 2224 // CHECK2-NEXT: store %"class.std::complex.0"* [[__C]], %"class.std::complex.0"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 2225 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 2226 // CHECK2-NEXT: [[TMP0:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 2227 // CHECK2-NEXT: [[CALL:%.*]] = call double @_ZNKSt7complexIdE4realEv(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[TMP0]]) #[[ATTR8]] 2228 // CHECK2-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 0 2229 // CHECK2-NEXT: [[TMP1:%.*]] = load double, double* [[__RE_]], align 8, !tbaa [[TBAA24:![0-9]+]] 2230 // CHECK2-NEXT: [[ADD:%.*]] = fadd double [[TMP1]], [[CALL]] 2231 // CHECK2-NEXT: store double [[ADD]], double* [[__RE_]], align 8, !tbaa [[TBAA24]] 2232 // CHECK2-NEXT: [[TMP2:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 2233 // CHECK2-NEXT: [[CALL2:%.*]] = call double @_ZNKSt7complexIdE4imagEv(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[TMP2]]) #[[ATTR8]] 2234 // CHECK2-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 1 2235 // CHECK2-NEXT: [[TMP3:%.*]] = load double, double* [[__IM_]], align 8, !tbaa [[TBAA26:![0-9]+]] 2236 // CHECK2-NEXT: [[ADD3:%.*]] = fadd double [[TMP3]], [[CALL2]] 2237 // CHECK2-NEXT: store double [[ADD3]], double* [[__IM_]], align 8, !tbaa [[TBAA26]] 2238 // CHECK2-NEXT: ret %"class.std::complex.0"* [[THIS1]] 2239 // 2240 // 2241 // CHECK2-LABEL: define {{[^@]+}}@_omp_reduction_shuffle_and_reduce_func5 2242 // CHECK2-SAME: (i8* [[TMP0:%.*]], i16 signext [[TMP1:%.*]], i16 signext [[TMP2:%.*]], i16 signext [[TMP3:%.*]]) #[[ATTR0]] { 2243 // CHECK2-NEXT: entry: 2244 // CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 2245 // CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca i16, align 2 2246 // CHECK2-NEXT: [[DOTADDR2:%.*]] = alloca i16, align 2 2247 // CHECK2-NEXT: [[DOTADDR3:%.*]] = alloca i16, align 2 2248 // CHECK2-NEXT: [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST:%.*]] = alloca [1 x i8*], align 8 2249 // CHECK2-NEXT: [[DOTOMP_REDUCTION_ELEMENT:%.*]] = alloca %"class.std::complex.0", align 8 2250 // CHECK2-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 2251 // CHECK2-NEXT: store i16 [[TMP1]], i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19]] 2252 // CHECK2-NEXT: store i16 [[TMP2]], i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 2253 // CHECK2-NEXT: store i16 [[TMP3]], i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 2254 // CHECK2-NEXT: [[TMP4:%.*]] = load i8*, i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 2255 // CHECK2-NEXT: [[TMP5:%.*]] = bitcast i8* [[TMP4]] to [1 x i8*]* 2256 // CHECK2-NEXT: [[TMP6:%.*]] = load i16, i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19]] 2257 // CHECK2-NEXT: [[TMP7:%.*]] = load i16, i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 2258 // CHECK2-NEXT: [[TMP8:%.*]] = load i16, i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 2259 // CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 2260 // CHECK2-NEXT: [[TMP10:%.*]] = load i8*, i8** [[TMP9]], align 8 2261 // CHECK2-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 2262 // CHECK2-NEXT: [[TMP12:%.*]] = bitcast i8* [[TMP10]] to %"class.std::complex.0"* 2263 // CHECK2-NEXT: [[TMP13:%.*]] = getelementptr %"class.std::complex.0", %"class.std::complex.0"* [[TMP12]], i64 1 2264 // CHECK2-NEXT: [[TMP14:%.*]] = bitcast %"class.std::complex.0"* [[TMP13]] to i8* 2265 // CHECK2-NEXT: [[TMP15:%.*]] = bitcast %"class.std::complex.0"* [[TMP12]] to i64* 2266 // CHECK2-NEXT: [[TMP16:%.*]] = bitcast %"class.std::complex.0"* [[DOTOMP_REDUCTION_ELEMENT]] to i64* 2267 // CHECK2-NEXT: br label [[DOTSHUFFLE_PRE_COND:%.*]] 2268 // CHECK2: .shuffle.pre_cond: 2269 // CHECK2-NEXT: [[TMP17:%.*]] = phi i64* [ [[TMP15]], [[ENTRY:%.*]] ], [ [[TMP28:%.*]], [[DOTSHUFFLE_THEN:%.*]] ] 2270 // CHECK2-NEXT: [[TMP18:%.*]] = phi i64* [ [[TMP16]], [[ENTRY]] ], [ [[TMP29:%.*]], [[DOTSHUFFLE_THEN]] ] 2271 // CHECK2-NEXT: [[TMP19:%.*]] = bitcast i64* [[TMP17]] to i8* 2272 // CHECK2-NEXT: [[TMP20:%.*]] = ptrtoint i8* [[TMP14]] to i64 2273 // CHECK2-NEXT: [[TMP21:%.*]] = ptrtoint i8* [[TMP19]] to i64 2274 // CHECK2-NEXT: [[TMP22:%.*]] = sub i64 [[TMP20]], [[TMP21]] 2275 // CHECK2-NEXT: [[TMP23:%.*]] = sdiv exact i64 [[TMP22]], ptrtoint (i8* getelementptr (i8, i8* null, i32 1) to i64) 2276 // CHECK2-NEXT: [[TMP24:%.*]] = icmp sgt i64 [[TMP23]], 7 2277 // CHECK2-NEXT: br i1 [[TMP24]], label [[DOTSHUFFLE_THEN]], label [[DOTSHUFFLE_EXIT:%.*]] 2278 // CHECK2: .shuffle.then: 2279 // CHECK2-NEXT: [[TMP25:%.*]] = load i64, i64* [[TMP17]], align 8 2280 // CHECK2-NEXT: [[NVPTX_WARP_SIZE:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize() 2281 // CHECK2-NEXT: [[TMP26:%.*]] = trunc i32 [[NVPTX_WARP_SIZE]] to i16 2282 // CHECK2-NEXT: [[TMP27:%.*]] = call i64 @__kmpc_shuffle_int64(i64 [[TMP25]], i16 [[TMP7]], i16 [[TMP26]]) 2283 // CHECK2-NEXT: store i64 [[TMP27]], i64* [[TMP18]], align 8 2284 // CHECK2-NEXT: [[TMP28]] = getelementptr i64, i64* [[TMP17]], i64 1 2285 // CHECK2-NEXT: [[TMP29]] = getelementptr i64, i64* [[TMP18]], i64 1 2286 // CHECK2-NEXT: br label [[DOTSHUFFLE_PRE_COND]] 2287 // CHECK2: .shuffle.exit: 2288 // CHECK2-NEXT: [[TMP30:%.*]] = bitcast %"class.std::complex.0"* [[DOTOMP_REDUCTION_ELEMENT]] to i8* 2289 // CHECK2-NEXT: store i8* [[TMP30]], i8** [[TMP11]], align 8, !tbaa [[TBAA12]] 2290 // CHECK2-NEXT: [[TMP31:%.*]] = icmp eq i16 [[TMP8]], 0 2291 // CHECK2-NEXT: [[TMP32:%.*]] = icmp eq i16 [[TMP8]], 1 2292 // CHECK2-NEXT: [[TMP33:%.*]] = icmp ult i16 [[TMP6]], [[TMP7]] 2293 // CHECK2-NEXT: [[TMP34:%.*]] = and i1 [[TMP32]], [[TMP33]] 2294 // CHECK2-NEXT: [[TMP35:%.*]] = icmp eq i16 [[TMP8]], 2 2295 // CHECK2-NEXT: [[TMP36:%.*]] = and i16 [[TMP6]], 1 2296 // CHECK2-NEXT: [[TMP37:%.*]] = icmp eq i16 [[TMP36]], 0 2297 // CHECK2-NEXT: [[TMP38:%.*]] = and i1 [[TMP35]], [[TMP37]] 2298 // CHECK2-NEXT: [[TMP39:%.*]] = icmp sgt i16 [[TMP7]], 0 2299 // CHECK2-NEXT: [[TMP40:%.*]] = and i1 [[TMP38]], [[TMP39]] 2300 // CHECK2-NEXT: [[TMP41:%.*]] = or i1 [[TMP31]], [[TMP34]] 2301 // CHECK2-NEXT: [[TMP42:%.*]] = or i1 [[TMP41]], [[TMP40]] 2302 // CHECK2-NEXT: br i1 [[TMP42]], label [[THEN:%.*]], label [[ELSE:%.*]] 2303 // CHECK2: then: 2304 // CHECK2-NEXT: [[TMP43:%.*]] = bitcast [1 x i8*]* [[TMP5]] to i8* 2305 // CHECK2-NEXT: [[TMP44:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]] to i8* 2306 // CHECK2-NEXT: call void @"_omp$reduction$reduction_func4"(i8* [[TMP43]], i8* [[TMP44]]) #[[ATTR1]] 2307 // CHECK2-NEXT: br label [[IFCONT:%.*]] 2308 // CHECK2: else: 2309 // CHECK2-NEXT: br label [[IFCONT]] 2310 // CHECK2: ifcont: 2311 // CHECK2-NEXT: [[TMP45:%.*]] = icmp eq i16 [[TMP8]], 1 2312 // CHECK2-NEXT: [[TMP46:%.*]] = icmp uge i16 [[TMP6]], [[TMP7]] 2313 // CHECK2-NEXT: [[TMP47:%.*]] = and i1 [[TMP45]], [[TMP46]] 2314 // CHECK2-NEXT: br i1 [[TMP47]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 2315 // CHECK2: then4: 2316 // CHECK2-NEXT: [[TMP48:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 2317 // CHECK2-NEXT: [[TMP49:%.*]] = load i8*, i8** [[TMP48]], align 8 2318 // CHECK2-NEXT: [[TMP50:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 2319 // CHECK2-NEXT: [[TMP51:%.*]] = load i8*, i8** [[TMP50]], align 8 2320 // CHECK2-NEXT: [[TMP52:%.*]] = bitcast i8* [[TMP49]] to %"class.std::complex.0"* 2321 // CHECK2-NEXT: [[TMP53:%.*]] = bitcast i8* [[TMP51]] to %"class.std::complex.0"* 2322 // CHECK2-NEXT: [[TMP54:%.*]] = bitcast %"class.std::complex.0"* [[TMP53]] to i8* 2323 // CHECK2-NEXT: [[TMP55:%.*]] = bitcast %"class.std::complex.0"* [[TMP52]] to i8* 2324 // CHECK2-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP54]], i8* align 8 [[TMP55]], i64 16, i1 false), !tbaa.struct !27 2325 // CHECK2-NEXT: br label [[IFCONT6:%.*]] 2326 // CHECK2: else5: 2327 // CHECK2-NEXT: br label [[IFCONT6]] 2328 // CHECK2: ifcont6: 2329 // CHECK2-NEXT: ret void 2330 // 2331 // 2332 // CHECK2-LABEL: define {{[^@]+}}@_omp_reduction_inter_warp_copy_func6 2333 // CHECK2-SAME: (i8* [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 2334 // CHECK2-NEXT: entry: 2335 // CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 2336 // CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 2337 // CHECK2-NEXT: [[DOTCNT_ADDR:%.*]] = alloca i32, align 4 2338 // CHECK2-NEXT: [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 2339 // CHECK2-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 2340 // CHECK2-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 2341 // CHECK2-NEXT: [[NVPTX_TID:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 2342 // CHECK2-NEXT: [[NVPTX_TID2:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 2343 // CHECK2-NEXT: [[NVPTX_LANE_ID:%.*]] = and i32 [[NVPTX_TID2]], 31 2344 // CHECK2-NEXT: [[NVPTX_TID3:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 2345 // CHECK2-NEXT: [[NVPTX_WARP_ID:%.*]] = ashr i32 [[NVPTX_TID3]], 5 2346 // CHECK2-NEXT: [[TMP3:%.*]] = load i8*, i8** [[DOTADDR]], align 8 2347 // CHECK2-NEXT: [[TMP4:%.*]] = bitcast i8* [[TMP3]] to [1 x i8*]* 2348 // CHECK2-NEXT: store i32 0, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 2349 // CHECK2-NEXT: br label [[PRECOND:%.*]] 2350 // CHECK2: precond: 2351 // CHECK2-NEXT: [[TMP5:%.*]] = load i32, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 2352 // CHECK2-NEXT: [[TMP6:%.*]] = icmp ult i32 [[TMP5]], 4 2353 // CHECK2-NEXT: br i1 [[TMP6]], label [[BODY:%.*]], label [[EXIT:%.*]] 2354 // CHECK2: body: 2355 // CHECK2-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4]], i32 [[TMP2]]) 2356 // CHECK2-NEXT: [[WARP_MASTER:%.*]] = icmp eq i32 [[NVPTX_LANE_ID]], 0 2357 // CHECK2-NEXT: br i1 [[WARP_MASTER]], label [[THEN:%.*]], label [[ELSE:%.*]] 2358 // CHECK2: then: 2359 // CHECK2-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 2360 // CHECK2-NEXT: [[TMP8:%.*]] = load i8*, i8** [[TMP7]], align 8, !tbaa [[TBAA12]] 2361 // CHECK2-NEXT: [[TMP9:%.*]] = bitcast i8* [[TMP8]] to i32* 2362 // CHECK2-NEXT: [[TMP10:%.*]] = getelementptr i32, i32* [[TMP9]], i32 [[TMP5]] 2363 // CHECK2-NEXT: [[TMP11:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_WARP_ID]] 2364 // CHECK2-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP10]], align 4 2365 // CHECK2-NEXT: store volatile i32 [[TMP12]], i32 addrspace(3)* [[TMP11]], align 4 2366 // CHECK2-NEXT: br label [[IFCONT:%.*]] 2367 // CHECK2: else: 2368 // CHECK2-NEXT: br label [[IFCONT]] 2369 // CHECK2: ifcont: 2370 // CHECK2-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4]], i32 [[TMP2]]) 2371 // CHECK2-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 2372 // CHECK2-NEXT: [[IS_ACTIVE_THREAD:%.*]] = icmp ult i32 [[NVPTX_TID]], [[TMP13]] 2373 // CHECK2-NEXT: br i1 [[IS_ACTIVE_THREAD]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 2374 // CHECK2: then4: 2375 // CHECK2-NEXT: [[TMP14:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_TID]] 2376 // CHECK2-NEXT: [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 2377 // CHECK2-NEXT: [[TMP16:%.*]] = load i8*, i8** [[TMP15]], align 8, !tbaa [[TBAA12]] 2378 // CHECK2-NEXT: [[TMP17:%.*]] = bitcast i8* [[TMP16]] to i32* 2379 // CHECK2-NEXT: [[TMP18:%.*]] = getelementptr i32, i32* [[TMP17]], i32 [[TMP5]] 2380 // CHECK2-NEXT: [[TMP19:%.*]] = load volatile i32, i32 addrspace(3)* [[TMP14]], align 4, !tbaa [[TBAA8]] 2381 // CHECK2-NEXT: store i32 [[TMP19]], i32* [[TMP18]], align 4, !tbaa [[TBAA8]] 2382 // CHECK2-NEXT: br label [[IFCONT6:%.*]] 2383 // CHECK2: else5: 2384 // CHECK2-NEXT: br label [[IFCONT6]] 2385 // CHECK2: ifcont6: 2386 // CHECK2-NEXT: [[TMP20:%.*]] = add nsw i32 [[TMP5]], 1 2387 // CHECK2-NEXT: store i32 [[TMP20]], i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 2388 // CHECK2-NEXT: br label [[PRECOND]] 2389 // CHECK2: exit: 2390 // CHECK2-NEXT: ret void 2391 // 2392 // 2393 // CHECK2-LABEL: define {{[^@]+}}@__omp_outlined__3_wrapper 2394 // CHECK2-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 2395 // CHECK2-NEXT: entry: 2396 // CHECK2-NEXT: [[DOTADDR:%.*]] = alloca i16, align 2 2397 // CHECK2-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 2398 // CHECK2-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 2399 // CHECK2-NEXT: [[GLOBAL_ARGS:%.*]] = alloca i8**, align 8 2400 // CHECK2-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 2401 // CHECK2-NEXT: store i16 [[TMP0]], i16* [[DOTADDR]], align 2, !tbaa [[TBAA19]] 2402 // CHECK2-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 2403 // CHECK2-NEXT: call void @__kmpc_get_shared_variables(i8*** [[GLOBAL_ARGS]]) 2404 // CHECK2-NEXT: [[TMP2:%.*]] = load i8**, i8*** [[GLOBAL_ARGS]], align 8 2405 // CHECK2-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 0 2406 // CHECK2-NEXT: [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32** 2407 // CHECK2-NEXT: [[TMP5:%.*]] = load i32*, i32** [[TMP4]], align 8, !tbaa [[TBAA12]] 2408 // CHECK2-NEXT: [[TMP6:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 1 2409 // CHECK2-NEXT: [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i32** 2410 // CHECK2-NEXT: [[TMP8:%.*]] = load i32*, i32** [[TMP7]], align 8, !tbaa [[TBAA12]] 2411 // CHECK2-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 2 2412 // CHECK2-NEXT: [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %"class.std::complex.0"** 2413 // CHECK2-NEXT: [[TMP11:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[TMP10]], align 8, !tbaa [[TBAA12]] 2414 // CHECK2-NEXT: call void @__omp_outlined__3(i32* [[DOTADDR1]], i32* [[DOTZERO_ADDR]], i32* [[TMP5]], i32* [[TMP8]], %"class.std::complex.0"* [[TMP11]]) #[[ATTR1]] 2415 // CHECK2-NEXT: ret void 2416 // 2417 // 2418 // CHECK2-LABEL: define {{[^@]+}}@_ZNSt7complexIfEC2ERKfS2_ 2419 // CHECK2-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]], float* nonnull align 4 dereferenceable(4) [[__RE:%.*]], float* nonnull align 4 dereferenceable(4) [[__IM:%.*]]) unnamed_addr #[[ATTR3]] comdat align 2 { 2420 // CHECK2-NEXT: entry: 2421 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 2422 // CHECK2-NEXT: [[__RE_ADDR:%.*]] = alloca float*, align 8 2423 // CHECK2-NEXT: [[__IM_ADDR:%.*]] = alloca float*, align 8 2424 // CHECK2-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 2425 // CHECK2-NEXT: store float* [[__RE]], float** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 2426 // CHECK2-NEXT: store float* [[__IM]], float** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 2427 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 2428 // CHECK2-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 0 2429 // CHECK2-NEXT: [[TMP0:%.*]] = load float*, float** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 2430 // CHECK2-NEXT: [[TMP1:%.*]] = load float, float* [[TMP0]], align 4, !tbaa [[TBAA14]] 2431 // CHECK2-NEXT: store float [[TMP1]], float* [[__RE_]], align 4, !tbaa [[TBAA16]] 2432 // CHECK2-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 1 2433 // CHECK2-NEXT: [[TMP2:%.*]] = load float*, float** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 2434 // CHECK2-NEXT: [[TMP3:%.*]] = load float, float* [[TMP2]], align 4, !tbaa [[TBAA14]] 2435 // CHECK2-NEXT: store float [[TMP3]], float* [[__IM_]], align 4, !tbaa [[TBAA18]] 2436 // CHECK2-NEXT: ret void 2437 // 2438 // 2439 // CHECK2-LABEL: define {{[^@]+}}@_ZNKSt7complexIfE4realEv 2440 // CHECK2-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 2441 // CHECK2-NEXT: entry: 2442 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 2443 // CHECK2-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 2444 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 2445 // CHECK2-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 0 2446 // CHECK2-NEXT: [[TMP0:%.*]] = load float, float* [[__RE_]], align 4, !tbaa [[TBAA16]] 2447 // CHECK2-NEXT: ret float [[TMP0]] 2448 // 2449 // 2450 // CHECK2-LABEL: define {{[^@]+}}@_ZNKSt7complexIfE4imagEv 2451 // CHECK2-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 2452 // CHECK2-NEXT: entry: 2453 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 2454 // CHECK2-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 2455 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 2456 // CHECK2-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 1 2457 // CHECK2-NEXT: [[TMP0:%.*]] = load float, float* [[__IM_]], align 4, !tbaa [[TBAA18]] 2458 // CHECK2-NEXT: ret float [[TMP0]] 2459 // 2460 // 2461 // CHECK2-LABEL: define {{[^@]+}}@_ZNSt7complexIdEC2ERKdS2_ 2462 // CHECK2-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]], double* nonnull align 8 dereferenceable(8) [[__RE:%.*]], double* nonnull align 8 dereferenceable(8) [[__IM:%.*]]) unnamed_addr #[[ATTR3]] comdat align 2 { 2463 // CHECK2-NEXT: entry: 2464 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 2465 // CHECK2-NEXT: [[__RE_ADDR:%.*]] = alloca double*, align 8 2466 // CHECK2-NEXT: [[__IM_ADDR:%.*]] = alloca double*, align 8 2467 // CHECK2-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 2468 // CHECK2-NEXT: store double* [[__RE]], double** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 2469 // CHECK2-NEXT: store double* [[__IM]], double** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 2470 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 2471 // CHECK2-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 0 2472 // CHECK2-NEXT: [[TMP0:%.*]] = load double*, double** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 2473 // CHECK2-NEXT: [[TMP1:%.*]] = load double, double* [[TMP0]], align 8, !tbaa [[TBAA22]] 2474 // CHECK2-NEXT: store double [[TMP1]], double* [[__RE_]], align 8, !tbaa [[TBAA24]] 2475 // CHECK2-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 1 2476 // CHECK2-NEXT: [[TMP2:%.*]] = load double*, double** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 2477 // CHECK2-NEXT: [[TMP3:%.*]] = load double, double* [[TMP2]], align 8, !tbaa [[TBAA22]] 2478 // CHECK2-NEXT: store double [[TMP3]], double* [[__IM_]], align 8, !tbaa [[TBAA26]] 2479 // CHECK2-NEXT: ret void 2480 // 2481 // 2482 // CHECK2-LABEL: define {{[^@]+}}@_ZNKSt7complexIdE4realEv 2483 // CHECK2-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 2484 // CHECK2-NEXT: entry: 2485 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 2486 // CHECK2-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 2487 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 2488 // CHECK2-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 0 2489 // CHECK2-NEXT: [[TMP0:%.*]] = load double, double* [[__RE_]], align 8, !tbaa [[TBAA24]] 2490 // CHECK2-NEXT: ret double [[TMP0]] 2491 // 2492 // 2493 // CHECK2-LABEL: define {{[^@]+}}@_ZNKSt7complexIdE4imagEv 2494 // CHECK2-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 2495 // CHECK2-NEXT: entry: 2496 // CHECK2-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 2497 // CHECK2-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 2498 // CHECK2-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 2499 // CHECK2-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 1 2500 // CHECK2-NEXT: [[TMP0:%.*]] = load double, double* [[__IM_]], align 8, !tbaa [[TBAA26]] 2501 // CHECK2-NEXT: ret double [[TMP0]] 2502 // 2503 // 2504 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z17complex_reductionIfEvv_l19 2505 // CHECK3-SAME: () #[[ATTR0:[0-9]+]] { 2506 // CHECK3-NEXT: entry: 2507 // CHECK3-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 2508 // CHECK3-NEXT: [[DOTTHREADID_TEMP_:%.*]] = alloca i32, align 4 2509 // CHECK3-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 2510 // CHECK3-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* @[[GLOB1:[0-9]+]], i1 false, i1 true, i1 true) 2511 // CHECK3-NEXT: [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1 2512 // CHECK3-NEXT: br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]] 2513 // CHECK3: user_code.entry: 2514 // CHECK3-NEXT: [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 2515 // CHECK3-NEXT: store i32 [[TMP1]], i32* [[DOTTHREADID_TEMP_]], align 4, !tbaa [[TBAA8:![0-9]+]] 2516 // CHECK3-NEXT: call void @__omp_outlined__(i32* [[DOTTHREADID_TEMP_]], i32* [[DOTZERO_ADDR]]) #[[ATTR1:[0-9]+]] 2517 // CHECK3-NEXT: call void @__kmpc_target_deinit(%struct.ident_t* @[[GLOB1]], i1 false, i1 true) 2518 // CHECK3-NEXT: ret void 2519 // CHECK3: worker.exit: 2520 // CHECK3-NEXT: ret void 2521 // 2522 // 2523 // CHECK3-LABEL: define {{[^@]+}}@__omp_outlined__ 2524 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR0]] { 2525 // CHECK3-NEXT: entry: 2526 // CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 2527 // CHECK3-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 2528 // CHECK3-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 2529 // CHECK3-NEXT: [[TMP:%.*]] = alloca i32, align 4 2530 // CHECK3-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 2531 // CHECK3-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 2532 // CHECK3-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 2533 // CHECK3-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 2534 // CHECK3-NEXT: [[IB:%.*]] = alloca i32, align 4 2535 // CHECK3-NEXT: [[REF_TMP:%.*]] = alloca float, align 4 2536 // CHECK3-NEXT: [[REF_TMP2:%.*]] = alloca float, align 4 2537 // CHECK3-NEXT: [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x i8*], align 8 2538 // CHECK3-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12:![0-9]+]] 2539 // CHECK3-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 2540 // CHECK3-NEXT: [[ISTART:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 2541 // CHECK3-NEXT: [[ISTART_ON_STACK:%.*]] = bitcast i8* [[ISTART]] to i32* 2542 // CHECK3-NEXT: [[IEND:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 2543 // CHECK3-NEXT: [[IEND_ON_STACK:%.*]] = bitcast i8* [[IEND]] to i32* 2544 // CHECK3-NEXT: [[PARTIAL_SUM:%.*]] = call i8* @__kmpc_alloc_shared(i64 8) 2545 // CHECK3-NEXT: [[PARTIAL_SUM_ON_STACK:%.*]] = bitcast i8* [[PARTIAL_SUM]] to %"class.std::complex"* 2546 // CHECK3-NEXT: [[TMP0:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 2547 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP0]]) #[[ATTR1]] 2548 // CHECK3-NEXT: [[TMP1:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 2549 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP1]]) #[[ATTR1]] 2550 // CHECK3-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2551 // CHECK3-NEXT: [[TMP2:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 2552 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP2]]) #[[ATTR1]] 2553 // CHECK3-NEXT: store i32 99, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2554 // CHECK3-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 2555 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 2556 // CHECK3-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 2557 // CHECK3-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 2558 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 2559 // CHECK3-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 2560 // CHECK3-NEXT: [[TMP5:%.*]] = bitcast i32* [[IB]] to i8* 2561 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP5]]) #[[ATTR1]] 2562 // CHECK3-NEXT: [[TMP6:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 2563 // CHECK3-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP6]], align 4, !tbaa [[TBAA8]] 2564 // CHECK3-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2:[0-9]+]], i32 [[TMP7]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 2565 // CHECK3-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2566 // CHECK3-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP8]], 99 2567 // CHECK3-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 2568 // CHECK3: cond.true: 2569 // CHECK3-NEXT: br label [[COND_END:%.*]] 2570 // CHECK3: cond.false: 2571 // CHECK3-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2572 // CHECK3-NEXT: br label [[COND_END]] 2573 // CHECK3: cond.end: 2574 // CHECK3-NEXT: [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP9]], [[COND_FALSE]] ] 2575 // CHECK3-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2576 // CHECK3-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2577 // CHECK3-NEXT: store i32 [[TMP10]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2578 // CHECK3-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 2579 // CHECK3: omp.inner.for.cond: 2580 // CHECK3-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2581 // CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2582 // CHECK3-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP11]], [[TMP12]] 2583 // CHECK3-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 2584 // CHECK3: omp.inner.for.cond.cleanup: 2585 // CHECK3-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 2586 // CHECK3: omp.inner.for.body: 2587 // CHECK3-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2588 // CHECK3-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP13]], 1 2589 // CHECK3-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]] 2590 // CHECK3-NEXT: store i32 [[ADD]], i32* [[IB]], align 4, !tbaa [[TBAA8]] 2591 // CHECK3-NEXT: [[TMP14:%.*]] = bitcast float* [[REF_TMP]] to i8* 2592 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP14]]) #[[ATTR1]] 2593 // CHECK3-NEXT: store float 0.000000e+00, float* [[REF_TMP]], align 4, !tbaa [[TBAA14:![0-9]+]] 2594 // CHECK3-NEXT: [[TMP15:%.*]] = bitcast float* [[REF_TMP2]] to i8* 2595 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP15]]) #[[ATTR1]] 2596 // CHECK3-NEXT: store float 0.000000e+00, float* [[REF_TMP2]], align 4, !tbaa [[TBAA14]] 2597 // CHECK3-NEXT: call void @_ZNSt7complexIfEC1ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM_ON_STACK]], float* nonnull align 4 dereferenceable(4) [[REF_TMP]], float* nonnull align 4 dereferenceable(4) [[REF_TMP2]]) #[[ATTR8:[0-9]+]] 2598 // CHECK3-NEXT: [[TMP16:%.*]] = bitcast float* [[REF_TMP2]] to i8* 2599 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP16]]) #[[ATTR1]] 2600 // CHECK3-NEXT: [[TMP17:%.*]] = bitcast float* [[REF_TMP]] to i8* 2601 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP17]]) #[[ATTR1]] 2602 // CHECK3-NEXT: [[TMP18:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 2603 // CHECK3-NEXT: [[MUL3:%.*]] = mul nsw i32 [[TMP18]], 4 2604 // CHECK3-NEXT: store i32 [[MUL3]], i32* [[ISTART_ON_STACK]], align 4, !tbaa [[TBAA8]] 2605 // CHECK3-NEXT: [[TMP19:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 2606 // CHECK3-NEXT: [[ADD4:%.*]] = add nsw i32 [[TMP19]], 1 2607 // CHECK3-NEXT: [[MUL5:%.*]] = mul nsw i32 [[ADD4]], 4 2608 // CHECK3-NEXT: store i32 [[MUL5]], i32* [[IEND_ON_STACK]], align 4, !tbaa [[TBAA8]] 2609 // CHECK3-NEXT: [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 0 2610 // CHECK3-NEXT: [[TMP21:%.*]] = bitcast i32* [[ISTART_ON_STACK]] to i8* 2611 // CHECK3-NEXT: store i8* [[TMP21]], i8** [[TMP20]], align 8, !tbaa [[TBAA12]] 2612 // CHECK3-NEXT: [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 1 2613 // CHECK3-NEXT: [[TMP23:%.*]] = bitcast i32* [[IEND_ON_STACK]] to i8* 2614 // CHECK3-NEXT: store i8* [[TMP23]], i8** [[TMP22]], align 8, !tbaa [[TBAA12]] 2615 // CHECK3-NEXT: [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 2 2616 // CHECK3-NEXT: [[TMP25:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM_ON_STACK]] to i8* 2617 // CHECK3-NEXT: store i8* [[TMP25]], i8** [[TMP24]], align 8, !tbaa [[TBAA12]] 2618 // CHECK3-NEXT: [[TMP26:%.*]] = bitcast [3 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8** 2619 // CHECK3-NEXT: call void @__kmpc_parallel_51(%struct.ident_t* @[[GLOB1]], i32 [[TMP7]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*, i32*, i32*, %"class.std::complex"*)* @__omp_outlined__1 to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined__1_wrapper to i8*), i8** [[TMP26]], i64 3) 2620 // CHECK3-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 2621 // CHECK3: omp.body.continue: 2622 // CHECK3-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 2623 // CHECK3: omp.inner.for.inc: 2624 // CHECK3-NEXT: [[TMP27:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2625 // CHECK3-NEXT: [[ADD6:%.*]] = add nsw i32 [[TMP27]], 1 2626 // CHECK3-NEXT: store i32 [[ADD6]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2627 // CHECK3-NEXT: br label [[OMP_INNER_FOR_COND]] 2628 // CHECK3: omp.inner.for.end: 2629 // CHECK3-NEXT: br label [[OMP_LOOP_EXIT:%.*]] 2630 // CHECK3: omp.loop.exit: 2631 // CHECK3-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB2]], i32 [[TMP7]]) 2632 // CHECK3-NEXT: [[TMP28:%.*]] = bitcast i32* [[IB]] to i8* 2633 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP28]]) #[[ATTR1]] 2634 // CHECK3-NEXT: [[TMP29:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 2635 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP29]]) #[[ATTR1]] 2636 // CHECK3-NEXT: [[TMP30:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 2637 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP30]]) #[[ATTR1]] 2638 // CHECK3-NEXT: [[TMP31:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 2639 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP31]]) #[[ATTR1]] 2640 // CHECK3-NEXT: [[TMP32:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 2641 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP32]]) #[[ATTR1]] 2642 // CHECK3-NEXT: [[TMP33:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 2643 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP33]]) #[[ATTR1]] 2644 // CHECK3-NEXT: call void @__kmpc_free_shared(i8* [[PARTIAL_SUM]], i64 8) 2645 // CHECK3-NEXT: call void @__kmpc_free_shared(i8* [[IEND]], i64 4) 2646 // CHECK3-NEXT: call void @__kmpc_free_shared(i8* [[ISTART]], i64 4) 2647 // CHECK3-NEXT: ret void 2648 // 2649 // 2650 // CHECK3-LABEL: define {{[^@]+}}@_ZNSt7complexIfEC1ERKfS2_ 2651 // CHECK3-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]], float* nonnull align 4 dereferenceable(4) [[__RE:%.*]], float* nonnull align 4 dereferenceable(4) [[__IM:%.*]]) unnamed_addr #[[ATTR3:[0-9]+]] comdat align 2 { 2652 // CHECK3-NEXT: entry: 2653 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 2654 // CHECK3-NEXT: [[__RE_ADDR:%.*]] = alloca float*, align 8 2655 // CHECK3-NEXT: [[__IM_ADDR:%.*]] = alloca float*, align 8 2656 // CHECK3-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 2657 // CHECK3-NEXT: store float* [[__RE]], float** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 2658 // CHECK3-NEXT: store float* [[__IM]], float** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 2659 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 2660 // CHECK3-NEXT: [[TMP0:%.*]] = load float*, float** [[__RE_ADDR]], align 8 2661 // CHECK3-NEXT: [[TMP1:%.*]] = load float*, float** [[__IM_ADDR]], align 8 2662 // CHECK3-NEXT: call void @_ZNSt7complexIfEC2ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS1]], float* nonnull align 4 dereferenceable(4) [[TMP0]], float* nonnull align 4 dereferenceable(4) [[TMP1]]) #[[ATTR8]] 2663 // CHECK3-NEXT: ret void 2664 // 2665 // 2666 // CHECK3-LABEL: define {{[^@]+}}@__omp_outlined__1 2667 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ISTART:%.*]], i32* nonnull align 4 dereferenceable(4) [[IEND:%.*]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM:%.*]]) #[[ATTR0]] { 2668 // CHECK3-NEXT: entry: 2669 // CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 2670 // CHECK3-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 2671 // CHECK3-NEXT: [[ISTART_ADDR:%.*]] = alloca i32*, align 8 2672 // CHECK3-NEXT: [[IEND_ADDR:%.*]] = alloca i32*, align 8 2673 // CHECK3-NEXT: [[PARTIAL_SUM_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 2674 // CHECK3-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 2675 // CHECK3-NEXT: [[TMP:%.*]] = alloca i32, align 4 2676 // CHECK3-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4 2677 // CHECK3-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4 2678 // CHECK3-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4 2679 // CHECK3-NEXT: [[I:%.*]] = alloca i32, align 4 2680 // CHECK3-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 2681 // CHECK3-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 2682 // CHECK3-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 2683 // CHECK3-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 2684 // CHECK3-NEXT: [[PARTIAL_SUM5:%.*]] = alloca %"class.std::complex", align 4 2685 // CHECK3-NEXT: [[REF_TMP:%.*]] = alloca float, align 4 2686 // CHECK3-NEXT: [[REF_TMP6:%.*]] = alloca float, align 4 2687 // CHECK3-NEXT: [[I7:%.*]] = alloca i32, align 4 2688 // CHECK3-NEXT: [[REF_TMP14:%.*]] = alloca %"class.std::complex", align 4 2689 // CHECK3-NEXT: [[REF_TMP15:%.*]] = alloca float, align 4 2690 // CHECK3-NEXT: [[REF_TMP16:%.*]] = alloca float, align 4 2691 // CHECK3-NEXT: [[DOTOMP_REDUCTION_RED_LIST:%.*]] = alloca [1 x i8*], align 8 2692 // CHECK3-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12]] 2693 // CHECK3-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 2694 // CHECK3-NEXT: store i32* [[ISTART]], i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 2695 // CHECK3-NEXT: store i32* [[IEND]], i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 2696 // CHECK3-NEXT: store %"class.std::complex"* [[PARTIAL_SUM]], %"class.std::complex"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 2697 // CHECK3-NEXT: [[TMP0:%.*]] = load i32*, i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 2698 // CHECK3-NEXT: [[TMP1:%.*]] = load i32*, i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 2699 // CHECK3-NEXT: [[TMP2:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 2700 // CHECK3-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 2701 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 2702 // CHECK3-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 2703 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 2704 // CHECK3-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP0]], align 4, !tbaa [[TBAA8]] 2705 // CHECK3-NEXT: store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2706 // CHECK3-NEXT: [[TMP6:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 2707 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP6]]) #[[ATTR1]] 2708 // CHECK3-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP1]], align 4, !tbaa [[TBAA8]] 2709 // CHECK3-NEXT: store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 2710 // CHECK3-NEXT: [[TMP8:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 2711 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP8]]) #[[ATTR1]] 2712 // CHECK3-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 2713 // CHECK3-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2714 // CHECK3-NEXT: [[SUB:%.*]] = sub i32 [[TMP9]], [[TMP10]] 2715 // CHECK3-NEXT: [[SUB3:%.*]] = sub i32 [[SUB]], 1 2716 // CHECK3-NEXT: [[ADD:%.*]] = add i32 [[SUB3]], 1 2717 // CHECK3-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], 1 2718 // CHECK3-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1 2719 // CHECK3-NEXT: store i32 [[SUB4]], i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 2720 // CHECK3-NEXT: [[TMP11:%.*]] = bitcast i32* [[I]] to i8* 2721 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP11]]) #[[ATTR1]] 2722 // CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2723 // CHECK3-NEXT: store i32 [[TMP12]], i32* [[I]], align 4, !tbaa [[TBAA8]] 2724 // CHECK3-NEXT: [[TMP13:%.*]] = bitcast i32* [[I]] to i8* 2725 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP13]]) #[[ATTR1]] 2726 // CHECK3-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2727 // CHECK3-NEXT: [[TMP15:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 2728 // CHECK3-NEXT: [[CMP:%.*]] = icmp slt i32 [[TMP14]], [[TMP15]] 2729 // CHECK3-NEXT: br i1 [[CMP]], label [[OMP_PRECOND_THEN:%.*]], label [[OMP_PRECOND_END:%.*]] 2730 // CHECK3: omp.precond.then: 2731 // CHECK3-NEXT: [[TMP16:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 2732 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP16]]) #[[ATTR1]] 2733 // CHECK3-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2734 // CHECK3-NEXT: [[TMP17:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 2735 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP17]]) #[[ATTR1]] 2736 // CHECK3-NEXT: [[TMP18:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 2737 // CHECK3-NEXT: store i32 [[TMP18]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2738 // CHECK3-NEXT: [[TMP19:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 2739 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP19]]) #[[ATTR1]] 2740 // CHECK3-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 2741 // CHECK3-NEXT: [[TMP20:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 2742 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP20]]) #[[ATTR1]] 2743 // CHECK3-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 2744 // CHECK3-NEXT: [[TMP21:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM5]] to i8* 2745 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP21]]) #[[ATTR1]] 2746 // CHECK3-NEXT: [[TMP22:%.*]] = bitcast float* [[REF_TMP]] to i8* 2747 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP22]]) #[[ATTR1]] 2748 // CHECK3-NEXT: store float 0.000000e+00, float* [[REF_TMP]], align 4, !tbaa [[TBAA14]] 2749 // CHECK3-NEXT: [[TMP23:%.*]] = bitcast float* [[REF_TMP6]] to i8* 2750 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP23]]) #[[ATTR1]] 2751 // CHECK3-NEXT: store float 0.000000e+00, float* [[REF_TMP6]], align 4, !tbaa [[TBAA14]] 2752 // CHECK3-NEXT: call void @_ZNSt7complexIfEC1ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM5]], float* nonnull align 4 dereferenceable(4) [[REF_TMP]], float* nonnull align 4 dereferenceable(4) [[REF_TMP6]]) #[[ATTR8]] 2753 // CHECK3-NEXT: [[TMP24:%.*]] = bitcast float* [[REF_TMP6]] to i8* 2754 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP24]]) #[[ATTR1]] 2755 // CHECK3-NEXT: [[TMP25:%.*]] = bitcast float* [[REF_TMP]] to i8* 2756 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP25]]) #[[ATTR1]] 2757 // CHECK3-NEXT: [[TMP26:%.*]] = bitcast i32* [[I7]] to i8* 2758 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP26]]) #[[ATTR1]] 2759 // CHECK3-NEXT: [[TMP27:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 2760 // CHECK3-NEXT: [[TMP28:%.*]] = load i32, i32* [[TMP27]], align 4, !tbaa [[TBAA8]] 2761 // CHECK3-NEXT: call void @__kmpc_for_static_init_4u(%struct.ident_t* @[[GLOB3:[0-9]+]], i32 [[TMP28]], i32 33, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 2762 // CHECK3-NEXT: br label [[OMP_DISPATCH_COND:%.*]] 2763 // CHECK3: omp.dispatch.cond: 2764 // CHECK3-NEXT: [[TMP29:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2765 // CHECK3-NEXT: [[TMP30:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 2766 // CHECK3-NEXT: [[CMP8:%.*]] = icmp ugt i32 [[TMP29]], [[TMP30]] 2767 // CHECK3-NEXT: br i1 [[CMP8]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 2768 // CHECK3: cond.true: 2769 // CHECK3-NEXT: [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 2770 // CHECK3-NEXT: br label [[COND_END:%.*]] 2771 // CHECK3: cond.false: 2772 // CHECK3-NEXT: [[TMP32:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2773 // CHECK3-NEXT: br label [[COND_END]] 2774 // CHECK3: cond.end: 2775 // CHECK3-NEXT: [[COND:%.*]] = phi i32 [ [[TMP31]], [[COND_TRUE]] ], [ [[TMP32]], [[COND_FALSE]] ] 2776 // CHECK3-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2777 // CHECK3-NEXT: [[TMP33:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2778 // CHECK3-NEXT: store i32 [[TMP33]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2779 // CHECK3-NEXT: [[TMP34:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2780 // CHECK3-NEXT: [[TMP35:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2781 // CHECK3-NEXT: [[ADD9:%.*]] = add i32 [[TMP35]], 1 2782 // CHECK3-NEXT: [[CMP10:%.*]] = icmp ult i32 [[TMP34]], [[ADD9]] 2783 // CHECK3-NEXT: br i1 [[CMP10]], label [[OMP_DISPATCH_BODY:%.*]], label [[OMP_DISPATCH_CLEANUP:%.*]] 2784 // CHECK3: omp.dispatch.cleanup: 2785 // CHECK3-NEXT: br label [[OMP_DISPATCH_END:%.*]] 2786 // CHECK3: omp.dispatch.body: 2787 // CHECK3-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 2788 // CHECK3: omp.inner.for.cond: 2789 // CHECK3-NEXT: [[TMP36:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2790 // CHECK3-NEXT: [[TMP37:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2791 // CHECK3-NEXT: [[ADD11:%.*]] = add i32 [[TMP37]], 1 2792 // CHECK3-NEXT: [[CMP12:%.*]] = icmp ult i32 [[TMP36]], [[ADD11]] 2793 // CHECK3-NEXT: br i1 [[CMP12]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 2794 // CHECK3: omp.inner.for.cond.cleanup: 2795 // CHECK3-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 2796 // CHECK3: omp.inner.for.body: 2797 // CHECK3-NEXT: [[TMP38:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 2798 // CHECK3-NEXT: [[TMP39:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2799 // CHECK3-NEXT: [[MUL:%.*]] = mul i32 [[TMP39]], 1 2800 // CHECK3-NEXT: [[ADD13:%.*]] = add i32 [[TMP38]], [[MUL]] 2801 // CHECK3-NEXT: store i32 [[ADD13]], i32* [[I7]], align 4, !tbaa [[TBAA8]] 2802 // CHECK3-NEXT: [[TMP40:%.*]] = bitcast %"class.std::complex"* [[REF_TMP14]] to i8* 2803 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP40]]) #[[ATTR1]] 2804 // CHECK3-NEXT: [[TMP41:%.*]] = bitcast float* [[REF_TMP15]] to i8* 2805 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP41]]) #[[ATTR1]] 2806 // CHECK3-NEXT: [[TMP42:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 2807 // CHECK3-NEXT: [[CONV:%.*]] = sitofp i32 [[TMP42]] to float 2808 // CHECK3-NEXT: store float [[CONV]], float* [[REF_TMP15]], align 4, !tbaa [[TBAA14]] 2809 // CHECK3-NEXT: [[TMP43:%.*]] = bitcast float* [[REF_TMP16]] to i8* 2810 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP43]]) #[[ATTR1]] 2811 // CHECK3-NEXT: [[TMP44:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 2812 // CHECK3-NEXT: [[CONV17:%.*]] = sitofp i32 [[TMP44]] to float 2813 // CHECK3-NEXT: store float [[CONV17]], float* [[REF_TMP16]], align 4, !tbaa [[TBAA14]] 2814 // CHECK3-NEXT: call void @_ZNSt7complexIfEC1ERKfS2_(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[REF_TMP14]], float* nonnull align 4 dereferenceable(4) [[REF_TMP15]], float* nonnull align 4 dereferenceable(4) [[REF_TMP16]]) #[[ATTR8]] 2815 // CHECK3-NEXT: [[CALL:%.*]] = call nonnull align 4 dereferenceable(8) %"class.std::complex"* @_ZNSt7complexIfEpLIfEERS0_RKS_IT_E(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM5]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[REF_TMP14]]) #[[ATTR8]] 2816 // CHECK3-NEXT: [[TMP45:%.*]] = bitcast float* [[REF_TMP16]] to i8* 2817 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP45]]) #[[ATTR1]] 2818 // CHECK3-NEXT: [[TMP46:%.*]] = bitcast float* [[REF_TMP15]] to i8* 2819 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP46]]) #[[ATTR1]] 2820 // CHECK3-NEXT: [[TMP47:%.*]] = bitcast %"class.std::complex"* [[REF_TMP14]] to i8* 2821 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP47]]) #[[ATTR1]] 2822 // CHECK3-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 2823 // CHECK3: omp.body.continue: 2824 // CHECK3-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 2825 // CHECK3: omp.inner.for.inc: 2826 // CHECK3-NEXT: [[TMP48:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2827 // CHECK3-NEXT: [[ADD18:%.*]] = add i32 [[TMP48]], 1 2828 // CHECK3-NEXT: store i32 [[ADD18]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 2829 // CHECK3-NEXT: br label [[OMP_INNER_FOR_COND]] 2830 // CHECK3: omp.inner.for.end: 2831 // CHECK3-NEXT: br label [[OMP_DISPATCH_INC:%.*]] 2832 // CHECK3: omp.dispatch.inc: 2833 // CHECK3-NEXT: [[TMP49:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2834 // CHECK3-NEXT: [[TMP50:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 2835 // CHECK3-NEXT: [[ADD19:%.*]] = add i32 [[TMP49]], [[TMP50]] 2836 // CHECK3-NEXT: store i32 [[ADD19]], i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 2837 // CHECK3-NEXT: [[TMP51:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2838 // CHECK3-NEXT: [[TMP52:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 2839 // CHECK3-NEXT: [[ADD20:%.*]] = add i32 [[TMP51]], [[TMP52]] 2840 // CHECK3-NEXT: store i32 [[ADD20]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 2841 // CHECK3-NEXT: br label [[OMP_DISPATCH_COND]] 2842 // CHECK3: omp.dispatch.end: 2843 // CHECK3-NEXT: [[TMP53:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 2844 // CHECK3-NEXT: [[TMP54:%.*]] = load i32, i32* [[TMP53]], align 4, !tbaa [[TBAA8]] 2845 // CHECK3-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB3]], i32 [[TMP54]]) 2846 // CHECK3-NEXT: [[TMP55:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 2847 // CHECK3-NEXT: [[TMP56:%.*]] = load i32, i32* [[TMP55]], align 4, !tbaa [[TBAA8]] 2848 // CHECK3-NEXT: [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]], i64 0, i64 0 2849 // CHECK3-NEXT: [[TMP58:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM5]] to i8* 2850 // CHECK3-NEXT: store i8* [[TMP58]], i8** [[TMP57]], align 8 2851 // CHECK3-NEXT: [[TMP59:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]] to i8* 2852 // CHECK3-NEXT: [[TMP60:%.*]] = call i32 @__kmpc_nvptx_parallel_reduce_nowait_v2(%struct.ident_t* @[[GLOB1]], i32 [[TMP56]], i32 1, i64 8, i8* [[TMP59]], void (i8*, i16, i16, i16)* @_omp_reduction_shuffle_and_reduce_func, void (i8*, i32)* @_omp_reduction_inter_warp_copy_func) 2853 // CHECK3-NEXT: [[TMP61:%.*]] = icmp eq i32 [[TMP60]], 1 2854 // CHECK3-NEXT: br i1 [[TMP61]], label [[DOTOMP_REDUCTION_THEN:%.*]], label [[DOTOMP_REDUCTION_DONE:%.*]] 2855 // CHECK3: .omp.reduction.then: 2856 // CHECK3-NEXT: [[CALL21:%.*]] = call nonnull align 4 dereferenceable(8) %"class.std::complex"* @_ZNSt7complexIfEpLIfEERS0_RKS_IT_E(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[TMP2]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[PARTIAL_SUM5]]) #[[ATTR8]] 2857 // CHECK3-NEXT: call void @__kmpc_nvptx_end_reduce_nowait(i32 [[TMP56]]) 2858 // CHECK3-NEXT: br label [[DOTOMP_REDUCTION_DONE]] 2859 // CHECK3: .omp.reduction.done: 2860 // CHECK3-NEXT: [[TMP62:%.*]] = bitcast i32* [[I7]] to i8* 2861 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP62]]) #[[ATTR1]] 2862 // CHECK3-NEXT: [[TMP63:%.*]] = bitcast %"class.std::complex"* [[PARTIAL_SUM5]] to i8* 2863 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP63]]) #[[ATTR1]] 2864 // CHECK3-NEXT: [[TMP64:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 2865 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP64]]) #[[ATTR1]] 2866 // CHECK3-NEXT: [[TMP65:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 2867 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP65]]) #[[ATTR1]] 2868 // CHECK3-NEXT: [[TMP66:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 2869 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP66]]) #[[ATTR1]] 2870 // CHECK3-NEXT: [[TMP67:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 2871 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP67]]) #[[ATTR1]] 2872 // CHECK3-NEXT: br label [[OMP_PRECOND_END]] 2873 // CHECK3: omp.precond.end: 2874 // CHECK3-NEXT: [[TMP68:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 2875 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP68]]) #[[ATTR1]] 2876 // CHECK3-NEXT: [[TMP69:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 2877 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP69]]) #[[ATTR1]] 2878 // CHECK3-NEXT: [[TMP70:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 2879 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP70]]) #[[ATTR1]] 2880 // CHECK3-NEXT: [[TMP71:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 2881 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP71]]) #[[ATTR1]] 2882 // CHECK3-NEXT: ret void 2883 // 2884 // 2885 // CHECK3-LABEL: define {{[^@]+}}@_ZNSt7complexIfEpLIfEERS0_RKS_IT_E 2886 // CHECK3-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]], %"class.std::complex"* nonnull align 4 dereferenceable(8) [[__C:%.*]]) #[[ATTR4:[0-9]+]] comdat align 2 { 2887 // CHECK3-NEXT: entry: 2888 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 2889 // CHECK3-NEXT: [[__C_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 2890 // CHECK3-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 2891 // CHECK3-NEXT: store %"class.std::complex"* [[__C]], %"class.std::complex"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 2892 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 2893 // CHECK3-NEXT: [[TMP0:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 2894 // CHECK3-NEXT: [[CALL:%.*]] = call float @_ZNKSt7complexIfE4realEv(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[TMP0]]) #[[ATTR8]] 2895 // CHECK3-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 0 2896 // CHECK3-NEXT: [[TMP1:%.*]] = load float, float* [[__RE_]], align 4, !tbaa [[TBAA16:![0-9]+]] 2897 // CHECK3-NEXT: [[ADD:%.*]] = fadd float [[TMP1]], [[CALL]] 2898 // CHECK3-NEXT: store float [[ADD]], float* [[__RE_]], align 4, !tbaa [[TBAA16]] 2899 // CHECK3-NEXT: [[TMP2:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 2900 // CHECK3-NEXT: [[CALL2:%.*]] = call float @_ZNKSt7complexIfE4imagEv(%"class.std::complex"* nonnull align 4 dereferenceable(8) [[TMP2]]) #[[ATTR8]] 2901 // CHECK3-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 1 2902 // CHECK3-NEXT: [[TMP3:%.*]] = load float, float* [[__IM_]], align 4, !tbaa [[TBAA18:![0-9]+]] 2903 // CHECK3-NEXT: [[ADD3:%.*]] = fadd float [[TMP3]], [[CALL2]] 2904 // CHECK3-NEXT: store float [[ADD3]], float* [[__IM_]], align 4, !tbaa [[TBAA18]] 2905 // CHECK3-NEXT: ret %"class.std::complex"* [[THIS1]] 2906 // 2907 // 2908 // CHECK3-LABEL: define {{[^@]+}}@_omp_reduction_shuffle_and_reduce_func 2909 // CHECK3-SAME: (i8* [[TMP0:%.*]], i16 signext [[TMP1:%.*]], i16 signext [[TMP2:%.*]], i16 signext [[TMP3:%.*]]) #[[ATTR0]] { 2910 // CHECK3-NEXT: entry: 2911 // CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 2912 // CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca i16, align 2 2913 // CHECK3-NEXT: [[DOTADDR2:%.*]] = alloca i16, align 2 2914 // CHECK3-NEXT: [[DOTADDR3:%.*]] = alloca i16, align 2 2915 // CHECK3-NEXT: [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST:%.*]] = alloca [1 x i8*], align 8 2916 // CHECK3-NEXT: [[DOTOMP_REDUCTION_ELEMENT:%.*]] = alloca %"class.std::complex", align 4 2917 // CHECK3-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 2918 // CHECK3-NEXT: store i16 [[TMP1]], i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19:![0-9]+]] 2919 // CHECK3-NEXT: store i16 [[TMP2]], i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 2920 // CHECK3-NEXT: store i16 [[TMP3]], i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 2921 // CHECK3-NEXT: [[TMP4:%.*]] = load i8*, i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 2922 // CHECK3-NEXT: [[TMP5:%.*]] = bitcast i8* [[TMP4]] to [1 x i8*]* 2923 // CHECK3-NEXT: [[TMP6:%.*]] = load i16, i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19]] 2924 // CHECK3-NEXT: [[TMP7:%.*]] = load i16, i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 2925 // CHECK3-NEXT: [[TMP8:%.*]] = load i16, i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 2926 // CHECK3-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 2927 // CHECK3-NEXT: [[TMP10:%.*]] = load i8*, i8** [[TMP9]], align 8 2928 // CHECK3-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 2929 // CHECK3-NEXT: [[TMP12:%.*]] = bitcast i8* [[TMP10]] to %"class.std::complex"* 2930 // CHECK3-NEXT: [[TMP13:%.*]] = getelementptr %"class.std::complex", %"class.std::complex"* [[TMP12]], i64 1 2931 // CHECK3-NEXT: [[TMP14:%.*]] = bitcast %"class.std::complex"* [[TMP13]] to i8* 2932 // CHECK3-NEXT: [[TMP15:%.*]] = bitcast %"class.std::complex"* [[TMP12]] to i64* 2933 // CHECK3-NEXT: [[TMP16:%.*]] = bitcast %"class.std::complex"* [[DOTOMP_REDUCTION_ELEMENT]] to i64* 2934 // CHECK3-NEXT: [[TMP17:%.*]] = load i64, i64* [[TMP15]], align 4 2935 // CHECK3-NEXT: [[NVPTX_WARP_SIZE:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize() 2936 // CHECK3-NEXT: [[TMP18:%.*]] = trunc i32 [[NVPTX_WARP_SIZE]] to i16 2937 // CHECK3-NEXT: [[TMP19:%.*]] = call i64 @__kmpc_shuffle_int64(i64 [[TMP17]], i16 [[TMP7]], i16 [[TMP18]]) 2938 // CHECK3-NEXT: store i64 [[TMP19]], i64* [[TMP16]], align 4 2939 // CHECK3-NEXT: [[TMP20:%.*]] = getelementptr i64, i64* [[TMP15]], i64 1 2940 // CHECK3-NEXT: [[TMP21:%.*]] = getelementptr i64, i64* [[TMP16]], i64 1 2941 // CHECK3-NEXT: [[TMP22:%.*]] = bitcast %"class.std::complex"* [[DOTOMP_REDUCTION_ELEMENT]] to i8* 2942 // CHECK3-NEXT: store i8* [[TMP22]], i8** [[TMP11]], align 8, !tbaa [[TBAA12]] 2943 // CHECK3-NEXT: [[TMP23:%.*]] = icmp eq i16 [[TMP8]], 0 2944 // CHECK3-NEXT: [[TMP24:%.*]] = icmp eq i16 [[TMP8]], 1 2945 // CHECK3-NEXT: [[TMP25:%.*]] = icmp ult i16 [[TMP6]], [[TMP7]] 2946 // CHECK3-NEXT: [[TMP26:%.*]] = and i1 [[TMP24]], [[TMP25]] 2947 // CHECK3-NEXT: [[TMP27:%.*]] = icmp eq i16 [[TMP8]], 2 2948 // CHECK3-NEXT: [[TMP28:%.*]] = and i16 [[TMP6]], 1 2949 // CHECK3-NEXT: [[TMP29:%.*]] = icmp eq i16 [[TMP28]], 0 2950 // CHECK3-NEXT: [[TMP30:%.*]] = and i1 [[TMP27]], [[TMP29]] 2951 // CHECK3-NEXT: [[TMP31:%.*]] = icmp sgt i16 [[TMP7]], 0 2952 // CHECK3-NEXT: [[TMP32:%.*]] = and i1 [[TMP30]], [[TMP31]] 2953 // CHECK3-NEXT: [[TMP33:%.*]] = or i1 [[TMP23]], [[TMP26]] 2954 // CHECK3-NEXT: [[TMP34:%.*]] = or i1 [[TMP33]], [[TMP32]] 2955 // CHECK3-NEXT: br i1 [[TMP34]], label [[THEN:%.*]], label [[ELSE:%.*]] 2956 // CHECK3: then: 2957 // CHECK3-NEXT: [[TMP35:%.*]] = bitcast [1 x i8*]* [[TMP5]] to i8* 2958 // CHECK3-NEXT: [[TMP36:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]] to i8* 2959 // CHECK3-NEXT: call void @"_omp$reduction$reduction_func"(i8* [[TMP35]], i8* [[TMP36]]) #[[ATTR1]] 2960 // CHECK3-NEXT: br label [[IFCONT:%.*]] 2961 // CHECK3: else: 2962 // CHECK3-NEXT: br label [[IFCONT]] 2963 // CHECK3: ifcont: 2964 // CHECK3-NEXT: [[TMP37:%.*]] = icmp eq i16 [[TMP8]], 1 2965 // CHECK3-NEXT: [[TMP38:%.*]] = icmp uge i16 [[TMP6]], [[TMP7]] 2966 // CHECK3-NEXT: [[TMP39:%.*]] = and i1 [[TMP37]], [[TMP38]] 2967 // CHECK3-NEXT: br i1 [[TMP39]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 2968 // CHECK3: then4: 2969 // CHECK3-NEXT: [[TMP40:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 2970 // CHECK3-NEXT: [[TMP41:%.*]] = load i8*, i8** [[TMP40]], align 8 2971 // CHECK3-NEXT: [[TMP42:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 2972 // CHECK3-NEXT: [[TMP43:%.*]] = load i8*, i8** [[TMP42]], align 8 2973 // CHECK3-NEXT: [[TMP44:%.*]] = bitcast i8* [[TMP41]] to %"class.std::complex"* 2974 // CHECK3-NEXT: [[TMP45:%.*]] = bitcast i8* [[TMP43]] to %"class.std::complex"* 2975 // CHECK3-NEXT: [[TMP46:%.*]] = bitcast %"class.std::complex"* [[TMP45]] to i8* 2976 // CHECK3-NEXT: [[TMP47:%.*]] = bitcast %"class.std::complex"* [[TMP44]] to i8* 2977 // CHECK3-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP46]], i8* align 4 [[TMP47]], i64 8, i1 false), !tbaa.struct !21 2978 // CHECK3-NEXT: br label [[IFCONT6:%.*]] 2979 // CHECK3: else5: 2980 // CHECK3-NEXT: br label [[IFCONT6]] 2981 // CHECK3: ifcont6: 2982 // CHECK3-NEXT: ret void 2983 // 2984 // 2985 // CHECK3-LABEL: define {{[^@]+}}@_omp_reduction_inter_warp_copy_func 2986 // CHECK3-SAME: (i8* [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 2987 // CHECK3-NEXT: entry: 2988 // CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 2989 // CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 2990 // CHECK3-NEXT: [[DOTCNT_ADDR:%.*]] = alloca i32, align 4 2991 // CHECK3-NEXT: [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 2992 // CHECK3-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 2993 // CHECK3-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 2994 // CHECK3-NEXT: [[NVPTX_TID:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 2995 // CHECK3-NEXT: [[NVPTX_TID2:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 2996 // CHECK3-NEXT: [[NVPTX_LANE_ID:%.*]] = and i32 [[NVPTX_TID2]], 31 2997 // CHECK3-NEXT: [[NVPTX_TID3:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 2998 // CHECK3-NEXT: [[NVPTX_WARP_ID:%.*]] = ashr i32 [[NVPTX_TID3]], 5 2999 // CHECK3-NEXT: [[TMP3:%.*]] = load i8*, i8** [[DOTADDR]], align 8 3000 // CHECK3-NEXT: [[TMP4:%.*]] = bitcast i8* [[TMP3]] to [1 x i8*]* 3001 // CHECK3-NEXT: store i32 0, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 3002 // CHECK3-NEXT: br label [[PRECOND:%.*]] 3003 // CHECK3: precond: 3004 // CHECK3-NEXT: [[TMP5:%.*]] = load i32, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 3005 // CHECK3-NEXT: [[TMP6:%.*]] = icmp ult i32 [[TMP5]], 2 3006 // CHECK3-NEXT: br i1 [[TMP6]], label [[BODY:%.*]], label [[EXIT:%.*]] 3007 // CHECK3: body: 3008 // CHECK3-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4:[0-9]+]], i32 [[TMP2]]) 3009 // CHECK3-NEXT: [[WARP_MASTER:%.*]] = icmp eq i32 [[NVPTX_LANE_ID]], 0 3010 // CHECK3-NEXT: br i1 [[WARP_MASTER]], label [[THEN:%.*]], label [[ELSE:%.*]] 3011 // CHECK3: then: 3012 // CHECK3-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 3013 // CHECK3-NEXT: [[TMP8:%.*]] = load i8*, i8** [[TMP7]], align 8, !tbaa [[TBAA12]] 3014 // CHECK3-NEXT: [[TMP9:%.*]] = bitcast i8* [[TMP8]] to i32* 3015 // CHECK3-NEXT: [[TMP10:%.*]] = getelementptr i32, i32* [[TMP9]], i32 [[TMP5]] 3016 // CHECK3-NEXT: [[TMP11:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_WARP_ID]] 3017 // CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP10]], align 4 3018 // CHECK3-NEXT: store volatile i32 [[TMP12]], i32 addrspace(3)* [[TMP11]], align 4 3019 // CHECK3-NEXT: br label [[IFCONT:%.*]] 3020 // CHECK3: else: 3021 // CHECK3-NEXT: br label [[IFCONT]] 3022 // CHECK3: ifcont: 3023 // CHECK3-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4]], i32 [[TMP2]]) 3024 // CHECK3-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 3025 // CHECK3-NEXT: [[IS_ACTIVE_THREAD:%.*]] = icmp ult i32 [[NVPTX_TID]], [[TMP13]] 3026 // CHECK3-NEXT: br i1 [[IS_ACTIVE_THREAD]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 3027 // CHECK3: then4: 3028 // CHECK3-NEXT: [[TMP14:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_TID]] 3029 // CHECK3-NEXT: [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 3030 // CHECK3-NEXT: [[TMP16:%.*]] = load i8*, i8** [[TMP15]], align 8, !tbaa [[TBAA12]] 3031 // CHECK3-NEXT: [[TMP17:%.*]] = bitcast i8* [[TMP16]] to i32* 3032 // CHECK3-NEXT: [[TMP18:%.*]] = getelementptr i32, i32* [[TMP17]], i32 [[TMP5]] 3033 // CHECK3-NEXT: [[TMP19:%.*]] = load volatile i32, i32 addrspace(3)* [[TMP14]], align 4, !tbaa [[TBAA8]] 3034 // CHECK3-NEXT: store i32 [[TMP19]], i32* [[TMP18]], align 4, !tbaa [[TBAA8]] 3035 // CHECK3-NEXT: br label [[IFCONT6:%.*]] 3036 // CHECK3: else5: 3037 // CHECK3-NEXT: br label [[IFCONT6]] 3038 // CHECK3: ifcont6: 3039 // CHECK3-NEXT: [[TMP20:%.*]] = add nsw i32 [[TMP5]], 1 3040 // CHECK3-NEXT: store i32 [[TMP20]], i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 3041 // CHECK3-NEXT: br label [[PRECOND]] 3042 // CHECK3: exit: 3043 // CHECK3-NEXT: ret void 3044 // 3045 // 3046 // CHECK3-LABEL: define {{[^@]+}}@__omp_outlined__1_wrapper 3047 // CHECK3-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 3048 // CHECK3-NEXT: entry: 3049 // CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i16, align 2 3050 // CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 3051 // CHECK3-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 3052 // CHECK3-NEXT: [[GLOBAL_ARGS:%.*]] = alloca i8**, align 8 3053 // CHECK3-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 3054 // CHECK3-NEXT: store i16 [[TMP0]], i16* [[DOTADDR]], align 2, !tbaa [[TBAA19]] 3055 // CHECK3-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 3056 // CHECK3-NEXT: call void @__kmpc_get_shared_variables(i8*** [[GLOBAL_ARGS]]) 3057 // CHECK3-NEXT: [[TMP2:%.*]] = load i8**, i8*** [[GLOBAL_ARGS]], align 8 3058 // CHECK3-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 0 3059 // CHECK3-NEXT: [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32** 3060 // CHECK3-NEXT: [[TMP5:%.*]] = load i32*, i32** [[TMP4]], align 8, !tbaa [[TBAA12]] 3061 // CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 1 3062 // CHECK3-NEXT: [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i32** 3063 // CHECK3-NEXT: [[TMP8:%.*]] = load i32*, i32** [[TMP7]], align 8, !tbaa [[TBAA12]] 3064 // CHECK3-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 2 3065 // CHECK3-NEXT: [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %"class.std::complex"** 3066 // CHECK3-NEXT: [[TMP11:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[TMP10]], align 8, !tbaa [[TBAA12]] 3067 // CHECK3-NEXT: call void @__omp_outlined__1(i32* [[DOTADDR1]], i32* [[DOTZERO_ADDR]], i32* [[TMP5]], i32* [[TMP8]], %"class.std::complex"* [[TMP11]]) #[[ATTR1]] 3068 // CHECK3-NEXT: ret void 3069 // 3070 // 3071 // CHECK3-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}__Z17complex_reductionIdEvv_l19 3072 // CHECK3-SAME: () #[[ATTR0]] { 3073 // CHECK3-NEXT: entry: 3074 // CHECK3-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 3075 // CHECK3-NEXT: [[DOTTHREADID_TEMP_:%.*]] = alloca i32, align 4 3076 // CHECK3-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 3077 // CHECK3-NEXT: [[TMP0:%.*]] = call i32 @__kmpc_target_init(%struct.ident_t* @[[GLOB1]], i1 false, i1 true, i1 true) 3078 // CHECK3-NEXT: [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1 3079 // CHECK3-NEXT: br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]] 3080 // CHECK3: user_code.entry: 3081 // CHECK3-NEXT: [[TMP1:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 3082 // CHECK3-NEXT: store i32 [[TMP1]], i32* [[DOTTHREADID_TEMP_]], align 4, !tbaa [[TBAA8]] 3083 // CHECK3-NEXT: call void @__omp_outlined__2(i32* [[DOTTHREADID_TEMP_]], i32* [[DOTZERO_ADDR]]) #[[ATTR1]] 3084 // CHECK3-NEXT: call void @__kmpc_target_deinit(%struct.ident_t* @[[GLOB1]], i1 false, i1 true) 3085 // CHECK3-NEXT: ret void 3086 // CHECK3: worker.exit: 3087 // CHECK3-NEXT: ret void 3088 // 3089 // 3090 // CHECK3-LABEL: define {{[^@]+}}@__omp_outlined__2 3091 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]]) #[[ATTR0]] { 3092 // CHECK3-NEXT: entry: 3093 // CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 3094 // CHECK3-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 3095 // CHECK3-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 3096 // CHECK3-NEXT: [[TMP:%.*]] = alloca i32, align 4 3097 // CHECK3-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 3098 // CHECK3-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 3099 // CHECK3-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 3100 // CHECK3-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 3101 // CHECK3-NEXT: [[IB:%.*]] = alloca i32, align 4 3102 // CHECK3-NEXT: [[REF_TMP:%.*]] = alloca double, align 8 3103 // CHECK3-NEXT: [[REF_TMP2:%.*]] = alloca double, align 8 3104 // CHECK3-NEXT: [[CAPTURED_VARS_ADDRS:%.*]] = alloca [3 x i8*], align 8 3105 // CHECK3-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12]] 3106 // CHECK3-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 3107 // CHECK3-NEXT: [[ISTART:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 3108 // CHECK3-NEXT: [[ISTART_ON_STACK:%.*]] = bitcast i8* [[ISTART]] to i32* 3109 // CHECK3-NEXT: [[IEND:%.*]] = call i8* @__kmpc_alloc_shared(i64 4) 3110 // CHECK3-NEXT: [[IEND_ON_STACK:%.*]] = bitcast i8* [[IEND]] to i32* 3111 // CHECK3-NEXT: [[PARTIAL_SUM:%.*]] = call i8* @__kmpc_alloc_shared(i64 16) 3112 // CHECK3-NEXT: [[PARTIAL_SUM_ON_STACK:%.*]] = bitcast i8* [[PARTIAL_SUM]] to %"class.std::complex.0"* 3113 // CHECK3-NEXT: [[TMP0:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 3114 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP0]]) #[[ATTR1]] 3115 // CHECK3-NEXT: [[TMP1:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 3116 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP1]]) #[[ATTR1]] 3117 // CHECK3-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 3118 // CHECK3-NEXT: [[TMP2:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 3119 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP2]]) #[[ATTR1]] 3120 // CHECK3-NEXT: store i32 99, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3121 // CHECK3-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 3122 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 3123 // CHECK3-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 3124 // CHECK3-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 3125 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 3126 // CHECK3-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 3127 // CHECK3-NEXT: [[TMP5:%.*]] = bitcast i32* [[IB]] to i8* 3128 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP5]]) #[[ATTR1]] 3129 // CHECK3-NEXT: [[TMP6:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 3130 // CHECK3-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP6]], align 4, !tbaa [[TBAA8]] 3131 // CHECK3-NEXT: call void @__kmpc_for_static_init_4(%struct.ident_t* @[[GLOB2]], i32 [[TMP7]], i32 92, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 3132 // CHECK3-NEXT: [[TMP8:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3133 // CHECK3-NEXT: [[CMP:%.*]] = icmp sgt i32 [[TMP8]], 99 3134 // CHECK3-NEXT: br i1 [[CMP]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 3135 // CHECK3: cond.true: 3136 // CHECK3-NEXT: br label [[COND_END:%.*]] 3137 // CHECK3: cond.false: 3138 // CHECK3-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3139 // CHECK3-NEXT: br label [[COND_END]] 3140 // CHECK3: cond.end: 3141 // CHECK3-NEXT: [[COND:%.*]] = phi i32 [ 99, [[COND_TRUE]] ], [ [[TMP9]], [[COND_FALSE]] ] 3142 // CHECK3-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3143 // CHECK3-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 3144 // CHECK3-NEXT: store i32 [[TMP10]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3145 // CHECK3-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 3146 // CHECK3: omp.inner.for.cond: 3147 // CHECK3-NEXT: [[TMP11:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3148 // CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3149 // CHECK3-NEXT: [[CMP1:%.*]] = icmp sle i32 [[TMP11]], [[TMP12]] 3150 // CHECK3-NEXT: br i1 [[CMP1]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 3151 // CHECK3: omp.inner.for.cond.cleanup: 3152 // CHECK3-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 3153 // CHECK3: omp.inner.for.body: 3154 // CHECK3-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3155 // CHECK3-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP13]], 1 3156 // CHECK3-NEXT: [[ADD:%.*]] = add nsw i32 0, [[MUL]] 3157 // CHECK3-NEXT: store i32 [[ADD]], i32* [[IB]], align 4, !tbaa [[TBAA8]] 3158 // CHECK3-NEXT: [[TMP14:%.*]] = bitcast double* [[REF_TMP]] to i8* 3159 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP14]]) #[[ATTR1]] 3160 // CHECK3-NEXT: store double 0.000000e+00, double* [[REF_TMP]], align 8, !tbaa [[TBAA22:![0-9]+]] 3161 // CHECK3-NEXT: [[TMP15:%.*]] = bitcast double* [[REF_TMP2]] to i8* 3162 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP15]]) #[[ATTR1]] 3163 // CHECK3-NEXT: store double 0.000000e+00, double* [[REF_TMP2]], align 8, !tbaa [[TBAA22]] 3164 // CHECK3-NEXT: call void @_ZNSt7complexIdEC1ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM_ON_STACK]], double* nonnull align 8 dereferenceable(8) [[REF_TMP]], double* nonnull align 8 dereferenceable(8) [[REF_TMP2]]) #[[ATTR8]] 3165 // CHECK3-NEXT: [[TMP16:%.*]] = bitcast double* [[REF_TMP2]] to i8* 3166 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP16]]) #[[ATTR1]] 3167 // CHECK3-NEXT: [[TMP17:%.*]] = bitcast double* [[REF_TMP]] to i8* 3168 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP17]]) #[[ATTR1]] 3169 // CHECK3-NEXT: [[TMP18:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 3170 // CHECK3-NEXT: [[MUL3:%.*]] = mul nsw i32 [[TMP18]], 4 3171 // CHECK3-NEXT: store i32 [[MUL3]], i32* [[ISTART_ON_STACK]], align 4, !tbaa [[TBAA8]] 3172 // CHECK3-NEXT: [[TMP19:%.*]] = load i32, i32* [[IB]], align 4, !tbaa [[TBAA8]] 3173 // CHECK3-NEXT: [[ADD4:%.*]] = add nsw i32 [[TMP19]], 1 3174 // CHECK3-NEXT: [[MUL5:%.*]] = mul nsw i32 [[ADD4]], 4 3175 // CHECK3-NEXT: store i32 [[MUL5]], i32* [[IEND_ON_STACK]], align 4, !tbaa [[TBAA8]] 3176 // CHECK3-NEXT: [[TMP20:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 0 3177 // CHECK3-NEXT: [[TMP21:%.*]] = bitcast i32* [[ISTART_ON_STACK]] to i8* 3178 // CHECK3-NEXT: store i8* [[TMP21]], i8** [[TMP20]], align 8, !tbaa [[TBAA12]] 3179 // CHECK3-NEXT: [[TMP22:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 1 3180 // CHECK3-NEXT: [[TMP23:%.*]] = bitcast i32* [[IEND_ON_STACK]] to i8* 3181 // CHECK3-NEXT: store i8* [[TMP23]], i8** [[TMP22]], align 8, !tbaa [[TBAA12]] 3182 // CHECK3-NEXT: [[TMP24:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 2 3183 // CHECK3-NEXT: [[TMP25:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM_ON_STACK]] to i8* 3184 // CHECK3-NEXT: store i8* [[TMP25]], i8** [[TMP24]], align 8, !tbaa [[TBAA12]] 3185 // CHECK3-NEXT: [[TMP26:%.*]] = bitcast [3 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8** 3186 // CHECK3-NEXT: call void @__kmpc_parallel_51(%struct.ident_t* @[[GLOB1]], i32 [[TMP7]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*, i32*, i32*, %"class.std::complex.0"*)* @__omp_outlined__3 to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined__3_wrapper to i8*), i8** [[TMP26]], i64 3) 3187 // CHECK3-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 3188 // CHECK3: omp.body.continue: 3189 // CHECK3-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 3190 // CHECK3: omp.inner.for.inc: 3191 // CHECK3-NEXT: [[TMP27:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3192 // CHECK3-NEXT: [[ADD6:%.*]] = add nsw i32 [[TMP27]], 1 3193 // CHECK3-NEXT: store i32 [[ADD6]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3194 // CHECK3-NEXT: br label [[OMP_INNER_FOR_COND]] 3195 // CHECK3: omp.inner.for.end: 3196 // CHECK3-NEXT: br label [[OMP_LOOP_EXIT:%.*]] 3197 // CHECK3: omp.loop.exit: 3198 // CHECK3-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB2]], i32 [[TMP7]]) 3199 // CHECK3-NEXT: [[TMP28:%.*]] = bitcast i32* [[IB]] to i8* 3200 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP28]]) #[[ATTR1]] 3201 // CHECK3-NEXT: [[TMP29:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 3202 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP29]]) #[[ATTR1]] 3203 // CHECK3-NEXT: [[TMP30:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 3204 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP30]]) #[[ATTR1]] 3205 // CHECK3-NEXT: [[TMP31:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 3206 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP31]]) #[[ATTR1]] 3207 // CHECK3-NEXT: [[TMP32:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 3208 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP32]]) #[[ATTR1]] 3209 // CHECK3-NEXT: [[TMP33:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 3210 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP33]]) #[[ATTR1]] 3211 // CHECK3-NEXT: call void @__kmpc_free_shared(i8* [[PARTIAL_SUM]], i64 16) 3212 // CHECK3-NEXT: call void @__kmpc_free_shared(i8* [[IEND]], i64 4) 3213 // CHECK3-NEXT: call void @__kmpc_free_shared(i8* [[ISTART]], i64 4) 3214 // CHECK3-NEXT: ret void 3215 // 3216 // 3217 // CHECK3-LABEL: define {{[^@]+}}@_ZNSt7complexIdEC1ERKdS2_ 3218 // CHECK3-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]], double* nonnull align 8 dereferenceable(8) [[__RE:%.*]], double* nonnull align 8 dereferenceable(8) [[__IM:%.*]]) unnamed_addr #[[ATTR3]] comdat align 2 { 3219 // CHECK3-NEXT: entry: 3220 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 3221 // CHECK3-NEXT: [[__RE_ADDR:%.*]] = alloca double*, align 8 3222 // CHECK3-NEXT: [[__IM_ADDR:%.*]] = alloca double*, align 8 3223 // CHECK3-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 3224 // CHECK3-NEXT: store double* [[__RE]], double** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 3225 // CHECK3-NEXT: store double* [[__IM]], double** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 3226 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 3227 // CHECK3-NEXT: [[TMP0:%.*]] = load double*, double** [[__RE_ADDR]], align 8 3228 // CHECK3-NEXT: [[TMP1:%.*]] = load double*, double** [[__IM_ADDR]], align 8 3229 // CHECK3-NEXT: call void @_ZNSt7complexIdEC2ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS1]], double* nonnull align 8 dereferenceable(8) [[TMP0]], double* nonnull align 8 dereferenceable(8) [[TMP1]]) #[[ATTR8]] 3230 // CHECK3-NEXT: ret void 3231 // 3232 // 3233 // CHECK3-LABEL: define {{[^@]+}}@__omp_outlined__3 3234 // CHECK3-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], i32* nonnull align 4 dereferenceable(4) [[ISTART:%.*]], i32* nonnull align 4 dereferenceable(4) [[IEND:%.*]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM:%.*]]) #[[ATTR0]] { 3235 // CHECK3-NEXT: entry: 3236 // CHECK3-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 3237 // CHECK3-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 3238 // CHECK3-NEXT: [[ISTART_ADDR:%.*]] = alloca i32*, align 8 3239 // CHECK3-NEXT: [[IEND_ADDR:%.*]] = alloca i32*, align 8 3240 // CHECK3-NEXT: [[PARTIAL_SUM_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 3241 // CHECK3-NEXT: [[DOTOMP_IV:%.*]] = alloca i32, align 4 3242 // CHECK3-NEXT: [[TMP:%.*]] = alloca i32, align 4 3243 // CHECK3-NEXT: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4 3244 // CHECK3-NEXT: [[DOTCAPTURE_EXPR_1:%.*]] = alloca i32, align 4 3245 // CHECK3-NEXT: [[DOTCAPTURE_EXPR_2:%.*]] = alloca i32, align 4 3246 // CHECK3-NEXT: [[I:%.*]] = alloca i32, align 4 3247 // CHECK3-NEXT: [[DOTOMP_LB:%.*]] = alloca i32, align 4 3248 // CHECK3-NEXT: [[DOTOMP_UB:%.*]] = alloca i32, align 4 3249 // CHECK3-NEXT: [[DOTOMP_STRIDE:%.*]] = alloca i32, align 4 3250 // CHECK3-NEXT: [[DOTOMP_IS_LAST:%.*]] = alloca i32, align 4 3251 // CHECK3-NEXT: [[PARTIAL_SUM5:%.*]] = alloca %"class.std::complex.0", align 8 3252 // CHECK3-NEXT: [[REF_TMP:%.*]] = alloca double, align 8 3253 // CHECK3-NEXT: [[REF_TMP6:%.*]] = alloca double, align 8 3254 // CHECK3-NEXT: [[I7:%.*]] = alloca i32, align 4 3255 // CHECK3-NEXT: [[REF_TMP14:%.*]] = alloca %"class.std::complex.0", align 8 3256 // CHECK3-NEXT: [[REF_TMP15:%.*]] = alloca double, align 8 3257 // CHECK3-NEXT: [[REF_TMP16:%.*]] = alloca double, align 8 3258 // CHECK3-NEXT: [[DOTOMP_REDUCTION_RED_LIST:%.*]] = alloca [1 x i8*], align 8 3259 // CHECK3-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8, !tbaa [[TBAA12]] 3260 // CHECK3-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8, !tbaa [[TBAA12]] 3261 // CHECK3-NEXT: store i32* [[ISTART]], i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 3262 // CHECK3-NEXT: store i32* [[IEND]], i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 3263 // CHECK3-NEXT: store %"class.std::complex.0"* [[PARTIAL_SUM]], %"class.std::complex.0"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 3264 // CHECK3-NEXT: [[TMP0:%.*]] = load i32*, i32** [[ISTART_ADDR]], align 8, !tbaa [[TBAA12]] 3265 // CHECK3-NEXT: [[TMP1:%.*]] = load i32*, i32** [[IEND_ADDR]], align 8, !tbaa [[TBAA12]] 3266 // CHECK3-NEXT: [[TMP2:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[PARTIAL_SUM_ADDR]], align 8, !tbaa [[TBAA12]] 3267 // CHECK3-NEXT: [[TMP3:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 3268 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR1]] 3269 // CHECK3-NEXT: [[TMP4:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 3270 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR1]] 3271 // CHECK3-NEXT: [[TMP5:%.*]] = load i32, i32* [[TMP0]], align 4, !tbaa [[TBAA8]] 3272 // CHECK3-NEXT: store i32 [[TMP5]], i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 3273 // CHECK3-NEXT: [[TMP6:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 3274 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP6]]) #[[ATTR1]] 3275 // CHECK3-NEXT: [[TMP7:%.*]] = load i32, i32* [[TMP1]], align 4, !tbaa [[TBAA8]] 3276 // CHECK3-NEXT: store i32 [[TMP7]], i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 3277 // CHECK3-NEXT: [[TMP8:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 3278 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP8]]) #[[ATTR1]] 3279 // CHECK3-NEXT: [[TMP9:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 3280 // CHECK3-NEXT: [[TMP10:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 3281 // CHECK3-NEXT: [[SUB:%.*]] = sub i32 [[TMP9]], [[TMP10]] 3282 // CHECK3-NEXT: [[SUB3:%.*]] = sub i32 [[SUB]], 1 3283 // CHECK3-NEXT: [[ADD:%.*]] = add i32 [[SUB3]], 1 3284 // CHECK3-NEXT: [[DIV:%.*]] = udiv i32 [[ADD]], 1 3285 // CHECK3-NEXT: [[SUB4:%.*]] = sub i32 [[DIV]], 1 3286 // CHECK3-NEXT: store i32 [[SUB4]], i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 3287 // CHECK3-NEXT: [[TMP11:%.*]] = bitcast i32* [[I]] to i8* 3288 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP11]]) #[[ATTR1]] 3289 // CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 3290 // CHECK3-NEXT: store i32 [[TMP12]], i32* [[I]], align 4, !tbaa [[TBAA8]] 3291 // CHECK3-NEXT: [[TMP13:%.*]] = bitcast i32* [[I]] to i8* 3292 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP13]]) #[[ATTR1]] 3293 // CHECK3-NEXT: [[TMP14:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 3294 // CHECK3-NEXT: [[TMP15:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_1]], align 4, !tbaa [[TBAA8]] 3295 // CHECK3-NEXT: [[CMP:%.*]] = icmp slt i32 [[TMP14]], [[TMP15]] 3296 // CHECK3-NEXT: br i1 [[CMP]], label [[OMP_PRECOND_THEN:%.*]], label [[OMP_PRECOND_END:%.*]] 3297 // CHECK3: omp.precond.then: 3298 // CHECK3-NEXT: [[TMP16:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 3299 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP16]]) #[[ATTR1]] 3300 // CHECK3-NEXT: store i32 0, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 3301 // CHECK3-NEXT: [[TMP17:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 3302 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP17]]) #[[ATTR1]] 3303 // CHECK3-NEXT: [[TMP18:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 3304 // CHECK3-NEXT: store i32 [[TMP18]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3305 // CHECK3-NEXT: [[TMP19:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 3306 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP19]]) #[[ATTR1]] 3307 // CHECK3-NEXT: store i32 1, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 3308 // CHECK3-NEXT: [[TMP20:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 3309 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP20]]) #[[ATTR1]] 3310 // CHECK3-NEXT: store i32 0, i32* [[DOTOMP_IS_LAST]], align 4, !tbaa [[TBAA8]] 3311 // CHECK3-NEXT: [[TMP21:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM5]] to i8* 3312 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[TMP21]]) #[[ATTR1]] 3313 // CHECK3-NEXT: [[TMP22:%.*]] = bitcast double* [[REF_TMP]] to i8* 3314 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP22]]) #[[ATTR1]] 3315 // CHECK3-NEXT: store double 0.000000e+00, double* [[REF_TMP]], align 8, !tbaa [[TBAA22]] 3316 // CHECK3-NEXT: [[TMP23:%.*]] = bitcast double* [[REF_TMP6]] to i8* 3317 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP23]]) #[[ATTR1]] 3318 // CHECK3-NEXT: store double 0.000000e+00, double* [[REF_TMP6]], align 8, !tbaa [[TBAA22]] 3319 // CHECK3-NEXT: call void @_ZNSt7complexIdEC1ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM5]], double* nonnull align 8 dereferenceable(8) [[REF_TMP]], double* nonnull align 8 dereferenceable(8) [[REF_TMP6]]) #[[ATTR8]] 3320 // CHECK3-NEXT: [[TMP24:%.*]] = bitcast double* [[REF_TMP6]] to i8* 3321 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP24]]) #[[ATTR1]] 3322 // CHECK3-NEXT: [[TMP25:%.*]] = bitcast double* [[REF_TMP]] to i8* 3323 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP25]]) #[[ATTR1]] 3324 // CHECK3-NEXT: [[TMP26:%.*]] = bitcast i32* [[I7]] to i8* 3325 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP26]]) #[[ATTR1]] 3326 // CHECK3-NEXT: [[TMP27:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 3327 // CHECK3-NEXT: [[TMP28:%.*]] = load i32, i32* [[TMP27]], align 4, !tbaa [[TBAA8]] 3328 // CHECK3-NEXT: call void @__kmpc_for_static_init_4u(%struct.ident_t* @[[GLOB3]], i32 [[TMP28]], i32 33, i32* [[DOTOMP_IS_LAST]], i32* [[DOTOMP_LB]], i32* [[DOTOMP_UB]], i32* [[DOTOMP_STRIDE]], i32 1, i32 1) 3329 // CHECK3-NEXT: br label [[OMP_DISPATCH_COND:%.*]] 3330 // CHECK3: omp.dispatch.cond: 3331 // CHECK3-NEXT: [[TMP29:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3332 // CHECK3-NEXT: [[TMP30:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 3333 // CHECK3-NEXT: [[CMP8:%.*]] = icmp ugt i32 [[TMP29]], [[TMP30]] 3334 // CHECK3-NEXT: br i1 [[CMP8]], label [[COND_TRUE:%.*]], label [[COND_FALSE:%.*]] 3335 // CHECK3: cond.true: 3336 // CHECK3-NEXT: [[TMP31:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_2]], align 4, !tbaa [[TBAA8]] 3337 // CHECK3-NEXT: br label [[COND_END:%.*]] 3338 // CHECK3: cond.false: 3339 // CHECK3-NEXT: [[TMP32:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3340 // CHECK3-NEXT: br label [[COND_END]] 3341 // CHECK3: cond.end: 3342 // CHECK3-NEXT: [[COND:%.*]] = phi i32 [ [[TMP31]], [[COND_TRUE]] ], [ [[TMP32]], [[COND_FALSE]] ] 3343 // CHECK3-NEXT: store i32 [[COND]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3344 // CHECK3-NEXT: [[TMP33:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 3345 // CHECK3-NEXT: store i32 [[TMP33]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3346 // CHECK3-NEXT: [[TMP34:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3347 // CHECK3-NEXT: [[TMP35:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3348 // CHECK3-NEXT: [[ADD9:%.*]] = add i32 [[TMP35]], 1 3349 // CHECK3-NEXT: [[CMP10:%.*]] = icmp ult i32 [[TMP34]], [[ADD9]] 3350 // CHECK3-NEXT: br i1 [[CMP10]], label [[OMP_DISPATCH_BODY:%.*]], label [[OMP_DISPATCH_CLEANUP:%.*]] 3351 // CHECK3: omp.dispatch.cleanup: 3352 // CHECK3-NEXT: br label [[OMP_DISPATCH_END:%.*]] 3353 // CHECK3: omp.dispatch.body: 3354 // CHECK3-NEXT: br label [[OMP_INNER_FOR_COND:%.*]] 3355 // CHECK3: omp.inner.for.cond: 3356 // CHECK3-NEXT: [[TMP36:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3357 // CHECK3-NEXT: [[TMP37:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3358 // CHECK3-NEXT: [[ADD11:%.*]] = add i32 [[TMP37]], 1 3359 // CHECK3-NEXT: [[CMP12:%.*]] = icmp ult i32 [[TMP36]], [[ADD11]] 3360 // CHECK3-NEXT: br i1 [[CMP12]], label [[OMP_INNER_FOR_BODY:%.*]], label [[OMP_INNER_FOR_COND_CLEANUP:%.*]] 3361 // CHECK3: omp.inner.for.cond.cleanup: 3362 // CHECK3-NEXT: br label [[OMP_INNER_FOR_END:%.*]] 3363 // CHECK3: omp.inner.for.body: 3364 // CHECK3-NEXT: [[TMP38:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4, !tbaa [[TBAA8]] 3365 // CHECK3-NEXT: [[TMP39:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3366 // CHECK3-NEXT: [[MUL:%.*]] = mul i32 [[TMP39]], 1 3367 // CHECK3-NEXT: [[ADD13:%.*]] = add i32 [[TMP38]], [[MUL]] 3368 // CHECK3-NEXT: store i32 [[ADD13]], i32* [[I7]], align 4, !tbaa [[TBAA8]] 3369 // CHECK3-NEXT: [[TMP40:%.*]] = bitcast %"class.std::complex.0"* [[REF_TMP14]] to i8* 3370 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[TMP40]]) #[[ATTR1]] 3371 // CHECK3-NEXT: [[TMP41:%.*]] = bitcast double* [[REF_TMP15]] to i8* 3372 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP41]]) #[[ATTR1]] 3373 // CHECK3-NEXT: [[TMP42:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 3374 // CHECK3-NEXT: [[CONV:%.*]] = sitofp i32 [[TMP42]] to double 3375 // CHECK3-NEXT: store double [[CONV]], double* [[REF_TMP15]], align 8, !tbaa [[TBAA22]] 3376 // CHECK3-NEXT: [[TMP43:%.*]] = bitcast double* [[REF_TMP16]] to i8* 3377 // CHECK3-NEXT: call void @llvm.lifetime.start.p0i8(i64 8, i8* [[TMP43]]) #[[ATTR1]] 3378 // CHECK3-NEXT: [[TMP44:%.*]] = load i32, i32* [[I7]], align 4, !tbaa [[TBAA8]] 3379 // CHECK3-NEXT: [[CONV17:%.*]] = sitofp i32 [[TMP44]] to double 3380 // CHECK3-NEXT: store double [[CONV17]], double* [[REF_TMP16]], align 8, !tbaa [[TBAA22]] 3381 // CHECK3-NEXT: call void @_ZNSt7complexIdEC1ERKdS2_(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[REF_TMP14]], double* nonnull align 8 dereferenceable(8) [[REF_TMP15]], double* nonnull align 8 dereferenceable(8) [[REF_TMP16]]) #[[ATTR8]] 3382 // CHECK3-NEXT: [[CALL:%.*]] = call nonnull align 8 dereferenceable(16) %"class.std::complex.0"* @_ZNSt7complexIdEpLIdEERS0_RKS_IT_E(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM5]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[REF_TMP14]]) #[[ATTR8]] 3383 // CHECK3-NEXT: [[TMP45:%.*]] = bitcast double* [[REF_TMP16]] to i8* 3384 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP45]]) #[[ATTR1]] 3385 // CHECK3-NEXT: [[TMP46:%.*]] = bitcast double* [[REF_TMP15]] to i8* 3386 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 8, i8* [[TMP46]]) #[[ATTR1]] 3387 // CHECK3-NEXT: [[TMP47:%.*]] = bitcast %"class.std::complex.0"* [[REF_TMP14]] to i8* 3388 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[TMP47]]) #[[ATTR1]] 3389 // CHECK3-NEXT: br label [[OMP_BODY_CONTINUE:%.*]] 3390 // CHECK3: omp.body.continue: 3391 // CHECK3-NEXT: br label [[OMP_INNER_FOR_INC:%.*]] 3392 // CHECK3: omp.inner.for.inc: 3393 // CHECK3-NEXT: [[TMP48:%.*]] = load i32, i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3394 // CHECK3-NEXT: [[ADD18:%.*]] = add i32 [[TMP48]], 1 3395 // CHECK3-NEXT: store i32 [[ADD18]], i32* [[DOTOMP_IV]], align 4, !tbaa [[TBAA8]] 3396 // CHECK3-NEXT: br label [[OMP_INNER_FOR_COND]] 3397 // CHECK3: omp.inner.for.end: 3398 // CHECK3-NEXT: br label [[OMP_DISPATCH_INC:%.*]] 3399 // CHECK3: omp.dispatch.inc: 3400 // CHECK3-NEXT: [[TMP49:%.*]] = load i32, i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 3401 // CHECK3-NEXT: [[TMP50:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 3402 // CHECK3-NEXT: [[ADD19:%.*]] = add i32 [[TMP49]], [[TMP50]] 3403 // CHECK3-NEXT: store i32 [[ADD19]], i32* [[DOTOMP_LB]], align 4, !tbaa [[TBAA8]] 3404 // CHECK3-NEXT: [[TMP51:%.*]] = load i32, i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3405 // CHECK3-NEXT: [[TMP52:%.*]] = load i32, i32* [[DOTOMP_STRIDE]], align 4, !tbaa [[TBAA8]] 3406 // CHECK3-NEXT: [[ADD20:%.*]] = add i32 [[TMP51]], [[TMP52]] 3407 // CHECK3-NEXT: store i32 [[ADD20]], i32* [[DOTOMP_UB]], align 4, !tbaa [[TBAA8]] 3408 // CHECK3-NEXT: br label [[OMP_DISPATCH_COND]] 3409 // CHECK3: omp.dispatch.end: 3410 // CHECK3-NEXT: [[TMP53:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 3411 // CHECK3-NEXT: [[TMP54:%.*]] = load i32, i32* [[TMP53]], align 4, !tbaa [[TBAA8]] 3412 // CHECK3-NEXT: call void @__kmpc_for_static_fini(%struct.ident_t* @[[GLOB3]], i32 [[TMP54]]) 3413 // CHECK3-NEXT: [[TMP55:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8 3414 // CHECK3-NEXT: [[TMP56:%.*]] = load i32, i32* [[TMP55]], align 4, !tbaa [[TBAA8]] 3415 // CHECK3-NEXT: [[TMP57:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]], i64 0, i64 0 3416 // CHECK3-NEXT: [[TMP58:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM5]] to i8* 3417 // CHECK3-NEXT: store i8* [[TMP58]], i8** [[TMP57]], align 8 3418 // CHECK3-NEXT: [[TMP59:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_RED_LIST]] to i8* 3419 // CHECK3-NEXT: [[TMP60:%.*]] = call i32 @__kmpc_nvptx_parallel_reduce_nowait_v2(%struct.ident_t* @[[GLOB1]], i32 [[TMP56]], i32 1, i64 8, i8* [[TMP59]], void (i8*, i16, i16, i16)* @_omp_reduction_shuffle_and_reduce_func5, void (i8*, i32)* @_omp_reduction_inter_warp_copy_func6) 3420 // CHECK3-NEXT: [[TMP61:%.*]] = icmp eq i32 [[TMP60]], 1 3421 // CHECK3-NEXT: br i1 [[TMP61]], label [[DOTOMP_REDUCTION_THEN:%.*]], label [[DOTOMP_REDUCTION_DONE:%.*]] 3422 // CHECK3: .omp.reduction.then: 3423 // CHECK3-NEXT: [[CALL21:%.*]] = call nonnull align 8 dereferenceable(16) %"class.std::complex.0"* @_ZNSt7complexIdEpLIdEERS0_RKS_IT_E(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[TMP2]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[PARTIAL_SUM5]]) #[[ATTR8]] 3424 // CHECK3-NEXT: call void @__kmpc_nvptx_end_reduce_nowait(i32 [[TMP56]]) 3425 // CHECK3-NEXT: br label [[DOTOMP_REDUCTION_DONE]] 3426 // CHECK3: .omp.reduction.done: 3427 // CHECK3-NEXT: [[TMP62:%.*]] = bitcast i32* [[I7]] to i8* 3428 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP62]]) #[[ATTR1]] 3429 // CHECK3-NEXT: [[TMP63:%.*]] = bitcast %"class.std::complex.0"* [[PARTIAL_SUM5]] to i8* 3430 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[TMP63]]) #[[ATTR1]] 3431 // CHECK3-NEXT: [[TMP64:%.*]] = bitcast i32* [[DOTOMP_IS_LAST]] to i8* 3432 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP64]]) #[[ATTR1]] 3433 // CHECK3-NEXT: [[TMP65:%.*]] = bitcast i32* [[DOTOMP_STRIDE]] to i8* 3434 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP65]]) #[[ATTR1]] 3435 // CHECK3-NEXT: [[TMP66:%.*]] = bitcast i32* [[DOTOMP_UB]] to i8* 3436 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP66]]) #[[ATTR1]] 3437 // CHECK3-NEXT: [[TMP67:%.*]] = bitcast i32* [[DOTOMP_LB]] to i8* 3438 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP67]]) #[[ATTR1]] 3439 // CHECK3-NEXT: br label [[OMP_PRECOND_END]] 3440 // CHECK3: omp.precond.end: 3441 // CHECK3-NEXT: [[TMP68:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_2]] to i8* 3442 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP68]]) #[[ATTR1]] 3443 // CHECK3-NEXT: [[TMP69:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_1]] to i8* 3444 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP69]]) #[[ATTR1]] 3445 // CHECK3-NEXT: [[TMP70:%.*]] = bitcast i32* [[DOTCAPTURE_EXPR_]] to i8* 3446 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP70]]) #[[ATTR1]] 3447 // CHECK3-NEXT: [[TMP71:%.*]] = bitcast i32* [[DOTOMP_IV]] to i8* 3448 // CHECK3-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP71]]) #[[ATTR1]] 3449 // CHECK3-NEXT: ret void 3450 // 3451 // 3452 // CHECK3-LABEL: define {{[^@]+}}@_ZNSt7complexIdEpLIdEERS0_RKS_IT_E 3453 // CHECK3-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]], %"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[__C:%.*]]) #[[ATTR4]] comdat align 2 { 3454 // CHECK3-NEXT: entry: 3455 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 3456 // CHECK3-NEXT: [[__C_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 3457 // CHECK3-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 3458 // CHECK3-NEXT: store %"class.std::complex.0"* [[__C]], %"class.std::complex.0"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 3459 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 3460 // CHECK3-NEXT: [[TMP0:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 3461 // CHECK3-NEXT: [[CALL:%.*]] = call double @_ZNKSt7complexIdE4realEv(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[TMP0]]) #[[ATTR8]] 3462 // CHECK3-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 0 3463 // CHECK3-NEXT: [[TMP1:%.*]] = load double, double* [[__RE_]], align 8, !tbaa [[TBAA24:![0-9]+]] 3464 // CHECK3-NEXT: [[ADD:%.*]] = fadd double [[TMP1]], [[CALL]] 3465 // CHECK3-NEXT: store double [[ADD]], double* [[__RE_]], align 8, !tbaa [[TBAA24]] 3466 // CHECK3-NEXT: [[TMP2:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[__C_ADDR]], align 8, !tbaa [[TBAA12]] 3467 // CHECK3-NEXT: [[CALL2:%.*]] = call double @_ZNKSt7complexIdE4imagEv(%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[TMP2]]) #[[ATTR8]] 3468 // CHECK3-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 1 3469 // CHECK3-NEXT: [[TMP3:%.*]] = load double, double* [[__IM_]], align 8, !tbaa [[TBAA26:![0-9]+]] 3470 // CHECK3-NEXT: [[ADD3:%.*]] = fadd double [[TMP3]], [[CALL2]] 3471 // CHECK3-NEXT: store double [[ADD3]], double* [[__IM_]], align 8, !tbaa [[TBAA26]] 3472 // CHECK3-NEXT: ret %"class.std::complex.0"* [[THIS1]] 3473 // 3474 // 3475 // CHECK3-LABEL: define {{[^@]+}}@_omp_reduction_shuffle_and_reduce_func5 3476 // CHECK3-SAME: (i8* [[TMP0:%.*]], i16 signext [[TMP1:%.*]], i16 signext [[TMP2:%.*]], i16 signext [[TMP3:%.*]]) #[[ATTR0]] { 3477 // CHECK3-NEXT: entry: 3478 // CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 3479 // CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca i16, align 2 3480 // CHECK3-NEXT: [[DOTADDR2:%.*]] = alloca i16, align 2 3481 // CHECK3-NEXT: [[DOTADDR3:%.*]] = alloca i16, align 2 3482 // CHECK3-NEXT: [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST:%.*]] = alloca [1 x i8*], align 8 3483 // CHECK3-NEXT: [[DOTOMP_REDUCTION_ELEMENT:%.*]] = alloca %"class.std::complex.0", align 8 3484 // CHECK3-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 3485 // CHECK3-NEXT: store i16 [[TMP1]], i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19]] 3486 // CHECK3-NEXT: store i16 [[TMP2]], i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 3487 // CHECK3-NEXT: store i16 [[TMP3]], i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 3488 // CHECK3-NEXT: [[TMP4:%.*]] = load i8*, i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 3489 // CHECK3-NEXT: [[TMP5:%.*]] = bitcast i8* [[TMP4]] to [1 x i8*]* 3490 // CHECK3-NEXT: [[TMP6:%.*]] = load i16, i16* [[DOTADDR1]], align 2, !tbaa [[TBAA19]] 3491 // CHECK3-NEXT: [[TMP7:%.*]] = load i16, i16* [[DOTADDR2]], align 2, !tbaa [[TBAA19]] 3492 // CHECK3-NEXT: [[TMP8:%.*]] = load i16, i16* [[DOTADDR3]], align 2, !tbaa [[TBAA19]] 3493 // CHECK3-NEXT: [[TMP9:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 3494 // CHECK3-NEXT: [[TMP10:%.*]] = load i8*, i8** [[TMP9]], align 8 3495 // CHECK3-NEXT: [[TMP11:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 3496 // CHECK3-NEXT: [[TMP12:%.*]] = bitcast i8* [[TMP10]] to %"class.std::complex.0"* 3497 // CHECK3-NEXT: [[TMP13:%.*]] = getelementptr %"class.std::complex.0", %"class.std::complex.0"* [[TMP12]], i64 1 3498 // CHECK3-NEXT: [[TMP14:%.*]] = bitcast %"class.std::complex.0"* [[TMP13]] to i8* 3499 // CHECK3-NEXT: [[TMP15:%.*]] = bitcast %"class.std::complex.0"* [[TMP12]] to i64* 3500 // CHECK3-NEXT: [[TMP16:%.*]] = bitcast %"class.std::complex.0"* [[DOTOMP_REDUCTION_ELEMENT]] to i64* 3501 // CHECK3-NEXT: br label [[DOTSHUFFLE_PRE_COND:%.*]] 3502 // CHECK3: .shuffle.pre_cond: 3503 // CHECK3-NEXT: [[TMP17:%.*]] = phi i64* [ [[TMP15]], [[ENTRY:%.*]] ], [ [[TMP28:%.*]], [[DOTSHUFFLE_THEN:%.*]] ] 3504 // CHECK3-NEXT: [[TMP18:%.*]] = phi i64* [ [[TMP16]], [[ENTRY]] ], [ [[TMP29:%.*]], [[DOTSHUFFLE_THEN]] ] 3505 // CHECK3-NEXT: [[TMP19:%.*]] = bitcast i64* [[TMP17]] to i8* 3506 // CHECK3-NEXT: [[TMP20:%.*]] = ptrtoint i8* [[TMP14]] to i64 3507 // CHECK3-NEXT: [[TMP21:%.*]] = ptrtoint i8* [[TMP19]] to i64 3508 // CHECK3-NEXT: [[TMP22:%.*]] = sub i64 [[TMP20]], [[TMP21]] 3509 // CHECK3-NEXT: [[TMP23:%.*]] = sdiv exact i64 [[TMP22]], ptrtoint (i8* getelementptr (i8, i8* null, i32 1) to i64) 3510 // CHECK3-NEXT: [[TMP24:%.*]] = icmp sgt i64 [[TMP23]], 7 3511 // CHECK3-NEXT: br i1 [[TMP24]], label [[DOTSHUFFLE_THEN]], label [[DOTSHUFFLE_EXIT:%.*]] 3512 // CHECK3: .shuffle.then: 3513 // CHECK3-NEXT: [[TMP25:%.*]] = load i64, i64* [[TMP17]], align 8 3514 // CHECK3-NEXT: [[NVPTX_WARP_SIZE:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.warpsize() 3515 // CHECK3-NEXT: [[TMP26:%.*]] = trunc i32 [[NVPTX_WARP_SIZE]] to i16 3516 // CHECK3-NEXT: [[TMP27:%.*]] = call i64 @__kmpc_shuffle_int64(i64 [[TMP25]], i16 [[TMP7]], i16 [[TMP26]]) 3517 // CHECK3-NEXT: store i64 [[TMP27]], i64* [[TMP18]], align 8 3518 // CHECK3-NEXT: [[TMP28]] = getelementptr i64, i64* [[TMP17]], i64 1 3519 // CHECK3-NEXT: [[TMP29]] = getelementptr i64, i64* [[TMP18]], i64 1 3520 // CHECK3-NEXT: br label [[DOTSHUFFLE_PRE_COND]] 3521 // CHECK3: .shuffle.exit: 3522 // CHECK3-NEXT: [[TMP30:%.*]] = bitcast %"class.std::complex.0"* [[DOTOMP_REDUCTION_ELEMENT]] to i8* 3523 // CHECK3-NEXT: store i8* [[TMP30]], i8** [[TMP11]], align 8, !tbaa [[TBAA12]] 3524 // CHECK3-NEXT: [[TMP31:%.*]] = icmp eq i16 [[TMP8]], 0 3525 // CHECK3-NEXT: [[TMP32:%.*]] = icmp eq i16 [[TMP8]], 1 3526 // CHECK3-NEXT: [[TMP33:%.*]] = icmp ult i16 [[TMP6]], [[TMP7]] 3527 // CHECK3-NEXT: [[TMP34:%.*]] = and i1 [[TMP32]], [[TMP33]] 3528 // CHECK3-NEXT: [[TMP35:%.*]] = icmp eq i16 [[TMP8]], 2 3529 // CHECK3-NEXT: [[TMP36:%.*]] = and i16 [[TMP6]], 1 3530 // CHECK3-NEXT: [[TMP37:%.*]] = icmp eq i16 [[TMP36]], 0 3531 // CHECK3-NEXT: [[TMP38:%.*]] = and i1 [[TMP35]], [[TMP37]] 3532 // CHECK3-NEXT: [[TMP39:%.*]] = icmp sgt i16 [[TMP7]], 0 3533 // CHECK3-NEXT: [[TMP40:%.*]] = and i1 [[TMP38]], [[TMP39]] 3534 // CHECK3-NEXT: [[TMP41:%.*]] = or i1 [[TMP31]], [[TMP34]] 3535 // CHECK3-NEXT: [[TMP42:%.*]] = or i1 [[TMP41]], [[TMP40]] 3536 // CHECK3-NEXT: br i1 [[TMP42]], label [[THEN:%.*]], label [[ELSE:%.*]] 3537 // CHECK3: then: 3538 // CHECK3-NEXT: [[TMP43:%.*]] = bitcast [1 x i8*]* [[TMP5]] to i8* 3539 // CHECK3-NEXT: [[TMP44:%.*]] = bitcast [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]] to i8* 3540 // CHECK3-NEXT: call void @"_omp$reduction$reduction_func4"(i8* [[TMP43]], i8* [[TMP44]]) #[[ATTR1]] 3541 // CHECK3-NEXT: br label [[IFCONT:%.*]] 3542 // CHECK3: else: 3543 // CHECK3-NEXT: br label [[IFCONT]] 3544 // CHECK3: ifcont: 3545 // CHECK3-NEXT: [[TMP45:%.*]] = icmp eq i16 [[TMP8]], 1 3546 // CHECK3-NEXT: [[TMP46:%.*]] = icmp uge i16 [[TMP6]], [[TMP7]] 3547 // CHECK3-NEXT: [[TMP47:%.*]] = and i1 [[TMP45]], [[TMP46]] 3548 // CHECK3-NEXT: br i1 [[TMP47]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 3549 // CHECK3: then4: 3550 // CHECK3-NEXT: [[TMP48:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[DOTOMP_REDUCTION_REMOTE_REDUCE_LIST]], i64 0, i64 0 3551 // CHECK3-NEXT: [[TMP49:%.*]] = load i8*, i8** [[TMP48]], align 8 3552 // CHECK3-NEXT: [[TMP50:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP5]], i64 0, i64 0 3553 // CHECK3-NEXT: [[TMP51:%.*]] = load i8*, i8** [[TMP50]], align 8 3554 // CHECK3-NEXT: [[TMP52:%.*]] = bitcast i8* [[TMP49]] to %"class.std::complex.0"* 3555 // CHECK3-NEXT: [[TMP53:%.*]] = bitcast i8* [[TMP51]] to %"class.std::complex.0"* 3556 // CHECK3-NEXT: [[TMP54:%.*]] = bitcast %"class.std::complex.0"* [[TMP53]] to i8* 3557 // CHECK3-NEXT: [[TMP55:%.*]] = bitcast %"class.std::complex.0"* [[TMP52]] to i8* 3558 // CHECK3-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 [[TMP54]], i8* align 8 [[TMP55]], i64 16, i1 false), !tbaa.struct !27 3559 // CHECK3-NEXT: br label [[IFCONT6:%.*]] 3560 // CHECK3: else5: 3561 // CHECK3-NEXT: br label [[IFCONT6]] 3562 // CHECK3: ifcont6: 3563 // CHECK3-NEXT: ret void 3564 // 3565 // 3566 // CHECK3-LABEL: define {{[^@]+}}@_omp_reduction_inter_warp_copy_func6 3567 // CHECK3-SAME: (i8* [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 3568 // CHECK3-NEXT: entry: 3569 // CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i8*, align 8 3570 // CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 3571 // CHECK3-NEXT: [[DOTCNT_ADDR:%.*]] = alloca i32, align 4 3572 // CHECK3-NEXT: [[TMP2:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1]]) 3573 // CHECK3-NEXT: store i8* [[TMP0]], i8** [[DOTADDR]], align 8, !tbaa [[TBAA12]] 3574 // CHECK3-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 3575 // CHECK3-NEXT: [[NVPTX_TID:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 3576 // CHECK3-NEXT: [[NVPTX_TID2:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 3577 // CHECK3-NEXT: [[NVPTX_LANE_ID:%.*]] = and i32 [[NVPTX_TID2]], 31 3578 // CHECK3-NEXT: [[NVPTX_TID3:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() 3579 // CHECK3-NEXT: [[NVPTX_WARP_ID:%.*]] = ashr i32 [[NVPTX_TID3]], 5 3580 // CHECK3-NEXT: [[TMP3:%.*]] = load i8*, i8** [[DOTADDR]], align 8 3581 // CHECK3-NEXT: [[TMP4:%.*]] = bitcast i8* [[TMP3]] to [1 x i8*]* 3582 // CHECK3-NEXT: store i32 0, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 3583 // CHECK3-NEXT: br label [[PRECOND:%.*]] 3584 // CHECK3: precond: 3585 // CHECK3-NEXT: [[TMP5:%.*]] = load i32, i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 3586 // CHECK3-NEXT: [[TMP6:%.*]] = icmp ult i32 [[TMP5]], 4 3587 // CHECK3-NEXT: br i1 [[TMP6]], label [[BODY:%.*]], label [[EXIT:%.*]] 3588 // CHECK3: body: 3589 // CHECK3-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4]], i32 [[TMP2]]) 3590 // CHECK3-NEXT: [[WARP_MASTER:%.*]] = icmp eq i32 [[NVPTX_LANE_ID]], 0 3591 // CHECK3-NEXT: br i1 [[WARP_MASTER]], label [[THEN:%.*]], label [[ELSE:%.*]] 3592 // CHECK3: then: 3593 // CHECK3-NEXT: [[TMP7:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 3594 // CHECK3-NEXT: [[TMP8:%.*]] = load i8*, i8** [[TMP7]], align 8, !tbaa [[TBAA12]] 3595 // CHECK3-NEXT: [[TMP9:%.*]] = bitcast i8* [[TMP8]] to i32* 3596 // CHECK3-NEXT: [[TMP10:%.*]] = getelementptr i32, i32* [[TMP9]], i32 [[TMP5]] 3597 // CHECK3-NEXT: [[TMP11:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_WARP_ID]] 3598 // CHECK3-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP10]], align 4 3599 // CHECK3-NEXT: store volatile i32 [[TMP12]], i32 addrspace(3)* [[TMP11]], align 4 3600 // CHECK3-NEXT: br label [[IFCONT:%.*]] 3601 // CHECK3: else: 3602 // CHECK3-NEXT: br label [[IFCONT]] 3603 // CHECK3: ifcont: 3604 // CHECK3-NEXT: call void @__kmpc_barrier(%struct.ident_t* @[[GLOB4]], i32 [[TMP2]]) 3605 // CHECK3-NEXT: [[TMP13:%.*]] = load i32, i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 3606 // CHECK3-NEXT: [[IS_ACTIVE_THREAD:%.*]] = icmp ult i32 [[NVPTX_TID]], [[TMP13]] 3607 // CHECK3-NEXT: br i1 [[IS_ACTIVE_THREAD]], label [[THEN4:%.*]], label [[ELSE5:%.*]] 3608 // CHECK3: then4: 3609 // CHECK3-NEXT: [[TMP14:%.*]] = getelementptr inbounds [32 x i32], [32 x i32] addrspace(3)* @__openmp_nvptx_data_transfer_temporary_storage, i64 0, i32 [[NVPTX_TID]] 3610 // CHECK3-NEXT: [[TMP15:%.*]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[TMP4]], i64 0, i64 0 3611 // CHECK3-NEXT: [[TMP16:%.*]] = load i8*, i8** [[TMP15]], align 8, !tbaa [[TBAA12]] 3612 // CHECK3-NEXT: [[TMP17:%.*]] = bitcast i8* [[TMP16]] to i32* 3613 // CHECK3-NEXT: [[TMP18:%.*]] = getelementptr i32, i32* [[TMP17]], i32 [[TMP5]] 3614 // CHECK3-NEXT: [[TMP19:%.*]] = load volatile i32, i32 addrspace(3)* [[TMP14]], align 4, !tbaa [[TBAA8]] 3615 // CHECK3-NEXT: store i32 [[TMP19]], i32* [[TMP18]], align 4, !tbaa [[TBAA8]] 3616 // CHECK3-NEXT: br label [[IFCONT6:%.*]] 3617 // CHECK3: else5: 3618 // CHECK3-NEXT: br label [[IFCONT6]] 3619 // CHECK3: ifcont6: 3620 // CHECK3-NEXT: [[TMP20:%.*]] = add nsw i32 [[TMP5]], 1 3621 // CHECK3-NEXT: store i32 [[TMP20]], i32* [[DOTCNT_ADDR]], align 4, !tbaa [[TBAA8]] 3622 // CHECK3-NEXT: br label [[PRECOND]] 3623 // CHECK3: exit: 3624 // CHECK3-NEXT: ret void 3625 // 3626 // 3627 // CHECK3-LABEL: define {{[^@]+}}@__omp_outlined__3_wrapper 3628 // CHECK3-SAME: (i16 zeroext [[TMP0:%.*]], i32 [[TMP1:%.*]]) #[[ATTR0]] { 3629 // CHECK3-NEXT: entry: 3630 // CHECK3-NEXT: [[DOTADDR:%.*]] = alloca i16, align 2 3631 // CHECK3-NEXT: [[DOTADDR1:%.*]] = alloca i32, align 4 3632 // CHECK3-NEXT: [[DOTZERO_ADDR:%.*]] = alloca i32, align 4 3633 // CHECK3-NEXT: [[GLOBAL_ARGS:%.*]] = alloca i8**, align 8 3634 // CHECK3-NEXT: store i32 0, i32* [[DOTZERO_ADDR]], align 4 3635 // CHECK3-NEXT: store i16 [[TMP0]], i16* [[DOTADDR]], align 2, !tbaa [[TBAA19]] 3636 // CHECK3-NEXT: store i32 [[TMP1]], i32* [[DOTADDR1]], align 4, !tbaa [[TBAA8]] 3637 // CHECK3-NEXT: call void @__kmpc_get_shared_variables(i8*** [[GLOBAL_ARGS]]) 3638 // CHECK3-NEXT: [[TMP2:%.*]] = load i8**, i8*** [[GLOBAL_ARGS]], align 8 3639 // CHECK3-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 0 3640 // CHECK3-NEXT: [[TMP4:%.*]] = bitcast i8** [[TMP3]] to i32** 3641 // CHECK3-NEXT: [[TMP5:%.*]] = load i32*, i32** [[TMP4]], align 8, !tbaa [[TBAA12]] 3642 // CHECK3-NEXT: [[TMP6:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 1 3643 // CHECK3-NEXT: [[TMP7:%.*]] = bitcast i8** [[TMP6]] to i32** 3644 // CHECK3-NEXT: [[TMP8:%.*]] = load i32*, i32** [[TMP7]], align 8, !tbaa [[TBAA12]] 3645 // CHECK3-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8*, i8** [[TMP2]], i64 2 3646 // CHECK3-NEXT: [[TMP10:%.*]] = bitcast i8** [[TMP9]] to %"class.std::complex.0"** 3647 // CHECK3-NEXT: [[TMP11:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[TMP10]], align 8, !tbaa [[TBAA12]] 3648 // CHECK3-NEXT: call void @__omp_outlined__3(i32* [[DOTADDR1]], i32* [[DOTZERO_ADDR]], i32* [[TMP5]], i32* [[TMP8]], %"class.std::complex.0"* [[TMP11]]) #[[ATTR1]] 3649 // CHECK3-NEXT: ret void 3650 // 3651 // 3652 // CHECK3-LABEL: define {{[^@]+}}@_ZNSt7complexIfEC2ERKfS2_ 3653 // CHECK3-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]], float* nonnull align 4 dereferenceable(4) [[__RE:%.*]], float* nonnull align 4 dereferenceable(4) [[__IM:%.*]]) unnamed_addr #[[ATTR3]] comdat align 2 { 3654 // CHECK3-NEXT: entry: 3655 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 3656 // CHECK3-NEXT: [[__RE_ADDR:%.*]] = alloca float*, align 8 3657 // CHECK3-NEXT: [[__IM_ADDR:%.*]] = alloca float*, align 8 3658 // CHECK3-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 3659 // CHECK3-NEXT: store float* [[__RE]], float** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 3660 // CHECK3-NEXT: store float* [[__IM]], float** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 3661 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 3662 // CHECK3-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 0 3663 // CHECK3-NEXT: [[TMP0:%.*]] = load float*, float** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 3664 // CHECK3-NEXT: [[TMP1:%.*]] = load float, float* [[TMP0]], align 4, !tbaa [[TBAA14]] 3665 // CHECK3-NEXT: store float [[TMP1]], float* [[__RE_]], align 4, !tbaa [[TBAA16]] 3666 // CHECK3-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 1 3667 // CHECK3-NEXT: [[TMP2:%.*]] = load float*, float** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 3668 // CHECK3-NEXT: [[TMP3:%.*]] = load float, float* [[TMP2]], align 4, !tbaa [[TBAA14]] 3669 // CHECK3-NEXT: store float [[TMP3]], float* [[__IM_]], align 4, !tbaa [[TBAA18]] 3670 // CHECK3-NEXT: ret void 3671 // 3672 // 3673 // CHECK3-LABEL: define {{[^@]+}}@_ZNKSt7complexIfE4realEv 3674 // CHECK3-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 3675 // CHECK3-NEXT: entry: 3676 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 3677 // CHECK3-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 3678 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 3679 // CHECK3-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 0 3680 // CHECK3-NEXT: [[TMP0:%.*]] = load float, float* [[__RE_]], align 4, !tbaa [[TBAA16]] 3681 // CHECK3-NEXT: ret float [[TMP0]] 3682 // 3683 // 3684 // CHECK3-LABEL: define {{[^@]+}}@_ZNKSt7complexIfE4imagEv 3685 // CHECK3-SAME: (%"class.std::complex"* nonnull align 4 dereferenceable(8) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 3686 // CHECK3-NEXT: entry: 3687 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex"*, align 8 3688 // CHECK3-NEXT: store %"class.std::complex"* [[THIS]], %"class.std::complex"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 3689 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex"*, %"class.std::complex"** [[THIS_ADDR]], align 8 3690 // CHECK3-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex", %"class.std::complex"* [[THIS1]], i32 0, i32 1 3691 // CHECK3-NEXT: [[TMP0:%.*]] = load float, float* [[__IM_]], align 4, !tbaa [[TBAA18]] 3692 // CHECK3-NEXT: ret float [[TMP0]] 3693 // 3694 // 3695 // CHECK3-LABEL: define {{[^@]+}}@_ZNSt7complexIdEC2ERKdS2_ 3696 // CHECK3-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]], double* nonnull align 8 dereferenceable(8) [[__RE:%.*]], double* nonnull align 8 dereferenceable(8) [[__IM:%.*]]) unnamed_addr #[[ATTR3]] comdat align 2 { 3697 // CHECK3-NEXT: entry: 3698 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 3699 // CHECK3-NEXT: [[__RE_ADDR:%.*]] = alloca double*, align 8 3700 // CHECK3-NEXT: [[__IM_ADDR:%.*]] = alloca double*, align 8 3701 // CHECK3-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 3702 // CHECK3-NEXT: store double* [[__RE]], double** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 3703 // CHECK3-NEXT: store double* [[__IM]], double** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 3704 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 3705 // CHECK3-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 0 3706 // CHECK3-NEXT: [[TMP0:%.*]] = load double*, double** [[__RE_ADDR]], align 8, !tbaa [[TBAA12]] 3707 // CHECK3-NEXT: [[TMP1:%.*]] = load double, double* [[TMP0]], align 8, !tbaa [[TBAA22]] 3708 // CHECK3-NEXT: store double [[TMP1]], double* [[__RE_]], align 8, !tbaa [[TBAA24]] 3709 // CHECK3-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 1 3710 // CHECK3-NEXT: [[TMP2:%.*]] = load double*, double** [[__IM_ADDR]], align 8, !tbaa [[TBAA12]] 3711 // CHECK3-NEXT: [[TMP3:%.*]] = load double, double* [[TMP2]], align 8, !tbaa [[TBAA22]] 3712 // CHECK3-NEXT: store double [[TMP3]], double* [[__IM_]], align 8, !tbaa [[TBAA26]] 3713 // CHECK3-NEXT: ret void 3714 // 3715 // 3716 // CHECK3-LABEL: define {{[^@]+}}@_ZNKSt7complexIdE4realEv 3717 // CHECK3-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 3718 // CHECK3-NEXT: entry: 3719 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 3720 // CHECK3-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 3721 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 3722 // CHECK3-NEXT: [[__RE_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 0 3723 // CHECK3-NEXT: [[TMP0:%.*]] = load double, double* [[__RE_]], align 8, !tbaa [[TBAA24]] 3724 // CHECK3-NEXT: ret double [[TMP0]] 3725 // 3726 // 3727 // CHECK3-LABEL: define {{[^@]+}}@_ZNKSt7complexIdE4imagEv 3728 // CHECK3-SAME: (%"class.std::complex.0"* nonnull align 8 dereferenceable(16) [[THIS:%.*]]) #[[ATTR4]] comdat align 2 { 3729 // CHECK3-NEXT: entry: 3730 // CHECK3-NEXT: [[THIS_ADDR:%.*]] = alloca %"class.std::complex.0"*, align 8 3731 // CHECK3-NEXT: store %"class.std::complex.0"* [[THIS]], %"class.std::complex.0"** [[THIS_ADDR]], align 8, !tbaa [[TBAA12]] 3732 // CHECK3-NEXT: [[THIS1:%.*]] = load %"class.std::complex.0"*, %"class.std::complex.0"** [[THIS_ADDR]], align 8 3733 // CHECK3-NEXT: [[__IM_:%.*]] = getelementptr inbounds %"class.std::complex.0", %"class.std::complex.0"* [[THIS1]], i32 0, i32 1 3734 // CHECK3-NEXT: [[TMP0:%.*]] = load double, double* [[__IM_]], align 8, !tbaa [[TBAA26]] 3735 // CHECK3-NEXT: ret double [[TMP0]] 3736 // 3737