1 // Test host codegen. 2 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-64 3 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s 4 // RUN: %clang_cc1 -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-64 5 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 6 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s 7 // RUN: %clang_cc1 -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 8 9 // Test target codegen - host bc file has to be created first. 10 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm-bc %s -o %t-ppc-host.bc 11 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - | FileCheck %s --check-prefix TCHECK --check-prefix TCHECK-64 12 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o %t %s 13 // RUN: %clang_cc1 -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix TCHECK --check-prefix TCHECK-64 14 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm-bc %s -o %t-x86-host.bc 15 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o - | FileCheck %s --check-prefix TCHECK --check-prefix TCHECK-32 16 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -o %t %s 17 // RUN: %clang_cc1 -fopenmp -x c++ -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -fopenmp-is-device -fopenmp-host-ir-file-path %t-x86-host.bc -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix TCHECK --check-prefix TCHECK-32 18 19 // expected-no-diagnostics 20 #ifndef HEADER 21 #define HEADER 22 23 template<typename tx, typename ty> 24 struct TT{ 25 tx X; 26 ty Y; 27 }; 28 29 // CHECK: [[TT:%.+]] = type { i64, i8 } 30 // CHECK: [[S1:%.+]] = type { double } 31 32 // TCHECK: [[TT:%.+]] = type { i64, i8 } 33 // TCHECK: [[S1:%.+]] = type { double } 34 35 // CHECK-DAG: [[SIZET:@.+]] = private unnamed_addr constant [1 x i{{32|64}}] [i[[SZ:32|64]] 4] 36 // CHECK: [[MAPT:@.+]] = private unnamed_addr constant [1 x i32] [i32 288] 37 // CHECK-DAG: [[MAPT2:@.+]] = private unnamed_addr constant [9 x i32] [i32 288, i32 161, i32 288, i32 161, i32 161, i32 288, i32 288, i32 161, i32 161] 38 // CHECK-DAG: [[SIZET3:@.+]] = private unnamed_addr constant [1 x i{{32|64}}] zeroinitializer 39 // CHECK-DAG: [[MAPT3:@.+]] = private unnamed_addr constant [1 x i32] [i32 32] 40 // CHECK-DAG: [[MAPT4:@.+]] = private unnamed_addr constant [5 x i32] [i32 35, i32 288, i32 288, i32 288, i32 161] 41 // CHECK-DAG: [[SIZET5:@.+]] = private unnamed_addr constant [3 x i{{32|64}}] [i[[SZ]] 4, i[[SZ]] 1, i[[SZ]] 40] 42 // CHECK-DAG: [[MAPT5:@.+]] = private unnamed_addr constant [3 x i32] [i32 288, i32 288, i32 161] 43 // CHECK-DAG: [[SIZET6:@.+]] = private unnamed_addr constant [2 x i{{32|64}}] [i[[SZ]] 4, i[[SZ]] 40] 44 // CHECK-DAG: [[MAPT6:@.+]] = private unnamed_addr constant [2 x i32] [i32 288, i32 161] 45 46 47 // CHECK: define {{.*}}[[FOO:@.+]]( 48 int foo(int n, double *ptr) { 49 int a = 0; 50 short aa = 0; 51 float b[10]; 52 float bn[n]; 53 double c[5][10]; 54 double cn[5][n]; 55 TT<long long, char> d; 56 57 #pragma omp target firstprivate(a) 58 { 59 } 60 61 // a is passed by value to tgt_target 62 // CHECK: [[N_ADDR:%.+]] = alloca i{{[0-9]+}}, 63 // CHECK: [[PTR_ADDR:%.+]] = alloca double*, 64 // CHECK: [[A:%.+]] = alloca i{{[0-9]+}}, 65 // CHECK: [[A2:%.+]] = alloca i{{[0-9]+}}, 66 // CHECK: [[B:%.+]] = alloca [10 x float], 67 // CHECK: [[SSTACK:%.+]] = alloca i8*, 68 // CHECK: [[C:%.+]] = alloca [5 x [10 x double]], 69 // CHECK: [[D:%.+]] = alloca [[TT]], 70 // CHECK: [[ACAST:%.+]] = alloca i{{[0-9]+}}, 71 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 72 // CHECK: [[BASE_PTR_ARR:%.+]] = alloca [1 x i8*], 73 // CHECK: [[PTR_ARR:%.+]] = alloca [1 x i8*], 74 // CHECK: [[A2CAST:%.+]] = alloca i{{[0-9]+}}, 75 // CHECK: [[BASE_PTR_ARR2:%.+]] = alloca [9 x i8*], 76 // CHECK: [[PTR_ARR2:%.+]] = alloca [9 x i8*], 77 // CHECK: [[SIZET2:%.+]] = alloca [9 x i{{[0-9]+}}], 78 // CHECK: [[BASE_PTR_ARR3:%.+]] = alloca [1 x i8*], 79 // CHECK: [[PTR_ARR3:%.+]] = alloca [1 x i8*], 80 // CHECK: [[N_ADDR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[N_ADDR]], 81 // CHECK-64: [[N_EXT:%.+]] = zext i{{[0-9]+}} [[N_ADDR_VAL]] to i{{[0-9]+}} 82 // CHECK: [[SSAVE_RET:%.+]] = call i8* @llvm.stacksave() 83 // CHECK: store i8* [[SSAVE_RET]], i8** [[SSTACK]], 84 // CHECK-64: [[BN_VLA:%.+]] = alloca float, i{{[0-9]+}} [[N_EXT]], 85 // CHECK-32: [[BN_VLA:%.+]] = alloca float, i{{[0-9]+}} [[N_ADDR_VAL]], 86 // CHECK: [[N_ADDR_VAL2:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[N_ADDR]], 87 // CHECK-64: [[N_EXT2:%.+]] = zext i{{[0-9]+}} [[N_ADDR_VAL2]] to i{{[0-9]+}} 88 // CHECK-64: [[CN_SIZE:%.+]] = mul{{.+}} i{{[0-9]+}} 5, [[N_EXT2]] 89 // CHECK-32: [[CN_SIZE:%.+]] = mul{{.+}} i{{[0-9]+}} 5, [[N_ADDR_VAL2]] 90 // CHECK: [[CN_VLA:%.+]] = alloca double, i{{[0-9]+}} [[CN_SIZE]], 91 // CHECK: [[AVAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[A]], 92 // CHECK-64: [[CONV:%.+]] = bitcast i{{[0-9]+}}* [[ACAST]] to i{{[0-9]+}}* 93 // CHECK-64: store i{{[0-9]+}} [[AVAL]], i{{[0-9]+}}* [[CONV]], 94 // CHECK-32: store i{{[0-9]+}} [[AVAL]], i{{[0-9]+}}* [[ACAST]], 95 // CHECK: [[ACAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[ACAST]], 96 // CHECK: [[ACAST_TOPTR:%.+]] = inttoptr i{{[0-9]+}} [[ACAST_VAL]] to i8* 97 // CHECK: [[BASE_PTR_GEP:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BASE_PTR_ARR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 98 // CHECK: store i8* [[ACAST_TOPTR]], i8** [[BASE_PTR_GEP]], 99 // CHECK: [[ACAST_TOPTR2:%.+]] = inttoptr i{{[0-9]+}} [[ACAST_VAL]] to i8* 100 // CHECK: [[PTR_GEP:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PTR_ARR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 101 // CHECK: store i8* [[ACAST_TOPTR2]], i8** [[PTR_GEP]], 102 // CHECK: [[BASE_PTR_GEP_ARG:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BASE_PTR_ARR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 103 // CHECK: [[PTR_GEP_ARG:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PTR_ARR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 104 // CHECK: {{.+}} = call i32 @__tgt_target(i32 -1, {{.+}}, i32 1, i8** [[BASE_PTR_GEP_ARG]], i8** [[PTR_GEP_ARG]], i[[SZ]]* getelementptr inbounds ([1 x i[[SZ]]], [1 x i[[SZ]]]* [[SIZET]], i32 0, i32 0), i32* getelementptr inbounds ([1 x i32], [1 x i32]* [[MAPT]], i32 0, i32 0)) 105 106 // TCHECK: define void @__omp_offloading_{{.+}}(i{{[0-9]+}} [[A_IN:%.+]]) 107 // TCHECK: [[A_ADDR:%.+]] = alloca i{{[0-9]+}}, 108 // TCHECK-NOT: alloca i{{[0-9]+}}, 109 // TCHECK: store i{{[0-9]+}} [[A_IN]], i{{[0-9]+}}* [[A_ADDR]], 110 // TCHECK-NOT: store i{{[0-9]+}} % 111 // TCHECK: ret void 112 113 #pragma omp target firstprivate(aa,b,bn,c,cn,d) 114 { 115 aa += 1; 116 b[2] = 1.0; 117 bn[3] = 1.0; 118 c[1][2] = 1.0; 119 cn[1][3] = 1.0; 120 d.X = 1; 121 d.Y = 1; 122 } 123 124 // CHECK: [[A2VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[A2]], 125 // CHECK: [[A2CASTCONV:%.+]] = bitcast i{{[0-9]+}}* [[A2CAST]] to i{{[0-9]+}}* 126 // CHECK: store i{{[0-9]+}} [[A2VAL]], i{{[0-9]+}}* [[A2CASTCONV]], 127 // CHECK: [[A2CAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[A2CAST]], 128 // CHECK-64: [[BN_SIZE:%.+]] = mul{{.+}} i{{[0-9]+}} [[N_EXT]], 4 129 // CHECK-32: [[BN_SIZE:%.+]] = mul{{.+}} i{{[0-9]+}} [[N_ADDR_VAL]], 4 130 // CHECK-64: [[CN_SIZE_1:%.+]] = mul{{.+}} i{{[0-9]+}} 5, [[N_EXT2]] 131 // CHECK-32: [[CN_SIZE_1:%.+]] = mul{{.+}} i{{[0-9]+}} 5, [[N_ADDR_VAL2]] 132 // CHECK: [[CN_SIZE_2:%.+]] = mul{{.+}} i{{[0-9]+}} [[CN_SIZE_1]], 8 133 134 // firstprivate(aa) --> base_ptr = aa, ptr = aa, size = 2 (short) 135 // CHECK: [[A2CAST_TO_INT:%.+]] = inttoptr i{{[0-9]+}} [[A2CAST_VAL]] to i8* 136 // CHECK: [[BASE_PTR_GEP2_0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 137 // CHECK: store i8* [[A2CAST_TO_INT]], i8** [[BASE_PTR_GEP2_0]], 138 // CHECK: [[A2CAST_TO_INT_2:%.+]] = inttoptr i{{[0-9]+}} [[A2CAST_VAL]] to i8* 139 // CHECK: [[PTR_GEP2_0:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 140 // CHECK: store i8* [[A2CAST_TO_INT_2]], i8** [[PTR_GEP2_0]], 141 // CHECK: [[SIZE_GEPA2:%.+]] = getelementptr inbounds [9 x i{{[0-9]+}}], [9 x i{{[0-9]+}}]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 142 // CHECK: store i{{[0-9]+}} 2, i{{[0-9]+}}* [[SIZE_GEPA2]], 143 144 // firstprivate(b): base_ptr = &b[0], ptr = &b[0], size = 40 (sizeof(float)*10) 145 // CHECK: [[BCAST:%.+]] = bitcast [10 x float]* [[B]] to i8* 146 // CHECK: [[BASE_PTR_GEP2_1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 147 // CHECK: store i8* [[BCAST]], i8** [[BASE_PTR_GEP2_1]], 148 // CHECK: [[BCAST2:%.+]] = bitcast [10 x float]* [[B]] to i8* 149 // CHECK: [[PTR_GEP2_1:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 150 // CHECK: store i8* [[BCAST2]], i8** [[PTR_GEP2_1]], 151 // CHECK: [[SIZE_GEPB:%.+]] = getelementptr inbounds [9 x i{{[0-9]+}}], [9 x i{{[0-9]+}}]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 152 // CHECK: store i{{[0-9]+}} 40, i{{[0-9]+}}* [[SIZE_GEPB]], 153 154 // firstprivate(bn), 2 entries, n and bn: (1) base_ptr = n, ptr = n, size = 8 ; (2) base_ptr = &c[0], ptr = &c[0], size = n*sizeof(float) 155 // CHECK-64: [[N_EXT3_1:%.+]] = inttoptr i{{[0-9]+}} [[N_EXT]] to i8* 156 // CHECK-32: [[N_EXT3_1:%.+]] = inttoptr i{{[0-9]+}} [[N_ADDR_VAL]] to i8* 157 // CHECK: [[BASE_PTR_GEP2_2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 158 // CHECK: store i8* [[N_EXT3_1]], i8** [[BASE_PTR_GEP2_2]], 159 // CHECK-64: [[N_EXT3_2:%.+]] = inttoptr i{{[0-9]+}} [[N_EXT]] to i8* 160 // CHECK-32: [[N_EXT3_2:%.+]] = inttoptr i{{[0-9]+}} [[N_ADDR_VAL]] to i8* 161 // CHECK: [[PTR_GEP2_2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 162 // CHECK: store i8* [[N_EXT3_2]], i8** [[PTR_GEP2_2]], 163 // CHECK: [[SIZE_GEPBN_1:%.+]] = getelementptr inbounds [9 x i{{[0-9]+}}], [9 x i{{[0-9]+}}]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 164 // CHECK: store i{{[0-9]+}} {{[0-9]}}, i{{[0-9]+}}* [[SIZE_GEPBN_1]], 165 // CHECK: [[VLABN_BCAST:%.+]] = bitcast float* [[BN_VLA]] to i8* 166 // CHECK: [[BASE_PTR_GEP2_3:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 167 // CHECK: store i8* [[VLABN_BCAST]], i8** [[BASE_PTR_GEP2_3]], 168 // CHECK: [[SIZE_GEPBN_3:%.+]] = getelementptr inbounds [9 x i{{[0-9]+}}], [9 x i{{[0-9]+}}]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 169 // CHECK: store i{{[0-9]+}} [[BN_SIZE]], i{{[0-9]+}}* [[SIZE_GEPBN_3]] 170 171 // firstprivate(c): base_ptr = &c[0], ptr = &c[0], size = 400 (5*10*sizeof(double)) 172 // CHECK: [[C_BCAST:%.+]] = bitcast [5 x [10 x double]]* [[C]] to i8* 173 // CHECK: [[BASE_PTR_GEP2_4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 4 174 // CHECK: store i8* [[C_BCAST]], i8** [[BASE_PTR_GEP2_4]], 175 // CHECK: [[C_BCAST2:%.+]] = bitcast [5 x [10 x double]]* [[C]] to i8* 176 // CHECK: [[PTR_GEP2_4:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 4 177 // CHECK: store i8* [[C_BCAST2]], i8** [[PTR_GEP2_4]], 178 // CHECK: [[SIZE_GEPC_4:%.+]] = getelementptr inbounds [9 x i{{[0-9]+}}], [9 x i{{[0-9]+}}]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 4 179 // CHECK: store i{{[0-9]+}} 400, i{{[0-9]+}}* [[SIZE_GEPC_4]], 180 181 // firstprivate(cn), 3 entries, 5, n, cn: (1) base_ptr = 5, ptr = 5, size = 8; (2) (1) base_ptr = n, ptr = n, size = 8; (3) base_ptr = &cn[0], ptr = &cn[0], size = 5*n*sizeof(double) 182 // CHECK: [[BASE_PTR_GEP2_5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 5 183 // CHECK: store i8* inttoptr (i{{[0-9]+}} 5 to i8*), i8** [[BASE_PTR_GEP2_5]], 184 // CHECK: [[PTR_GEP2_5:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 5 185 // CHECK: store i8* inttoptr (i{{[0-9]+}} 5 to i8*), i8** [[PTR_GEP2_5]], 186 // CHECK: [[SIZE_GEPCN_5:%.+]] = getelementptr inbounds [9 x i{{[0-9]+}}], [9 x i{{[0-9]+}}]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 5 187 // CHECK: store i{{[0-9]+}} {{[0-9]}}, i{{[0-9]+}}* [[SIZE_GEPCN_5]], 188 // CHECK-64: [[CN_SZ_2_1:%.+]] = inttoptr i{{[0-9]+}} [[N_EXT2]] to i8* 189 // CHECK-32: [[CN_SZ_2_1:%.+]] = inttoptr i{{[0-9]+}} [[N_ADDR_VAL2]] to i8* 190 // CHECK: [[BASE_PTR_GEP2_6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 6 191 // CHECK: store i8* [[CN_SZ_2_1]], i8** [[BASE_PTR_GEP2_6]], 192 // CHECK-64: [[CN_SZ_2_2:%.+]] = inttoptr i{{[0-9]+}} [[N_EXT2]] to i8* 193 // CHECK-32: [[CN_SZ_2_2:%.+]] = inttoptr i{{[0-9]+}} [[N_ADDR_VAL2]] to i8* 194 // CHECK: [[PTR_GEP2_6:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 6 195 // CHECK: store i8* [[CN_SZ_2_2]], i8** [[PTR_GEP2_6]], 196 // CHECK: [[SIZE_GEPCN_6:%.+]] = getelementptr inbounds [9 x i{{[0-9]+}}], [9 x i{{[0-9]+}}]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 6 197 // CHECK: store i{{[0-9]+}} {{[0-9]}}, i{{[0-9]+}}* [[SIZE_GEPCN_6]], 198 // CHECK: [[VLA_CN_BCAST:%.+]] = bitcast double* [[CN_VLA]] to i8* 199 // CHECK: [[BASE_PTR_GEP2_7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 7 200 // CHECK: store i8* [[VLA_CN_BCAST]], i8** [[BASE_PTR_GEP2_7]], 201 // CHECK: [[VLA_CN_BCAST2:%.+]] = bitcast double* [[CN_VLA]] to i8* 202 // CHECK: [[PTR_GEP2_7:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 7 203 // CHECK: store i8* [[VLA_CN_BCAST2]], i8** [[PTR_GEP2_7]], 204 // CHECK: [[SIZE_GEPCN_7:%.+]] = getelementptr inbounds [9 x i{{[0-9]+}}], [9 x i{{[0-9]+}}]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 7 205 // CHECK: store i{{[0-9]+}} [[CN_SIZE_2]], i{{[0-9]+}}* [[SIZE_GEPCN_7]], 206 207 // firstprivate(d): base_ptr = &d, ptr = &d, size = 16 208 // CHECK: [[D_REF:%.+]] = bitcast [[TT]]* [[D]] to i8* 209 // CHECK: [[BASE_PTR_GEP2_8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 8 210 // CHECK: store i8* [[D_REF]], i8** [[BASE_PTR_GEP2_8]], 211 // CHECK: [[D_REF2:%.+]] = bitcast [[TT]]* [[D]] to i8* 212 // CHECK: [[PTR_GEP2_8:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 8 213 // CHECK: store i8* [[D_REF2]], i8** [[PTR_GEP2_8]], 214 // CHECK: [[SIZE_GEPCN_8:%.+]] = getelementptr inbounds [9 x i{{[0-9]+}}], [9 x i{{[0-9]+}}]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 8 215 // CHECK: store i{{[0-9]+}} {{[0-9]+}}, i{{[0-9]+}}* [[SIZE_GEPCN_8]], 216 217 218 // CHECK: [[BASE_PTR_GEP_ARG2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[BASE_PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 219 // CHECK: [[PTR_GEP_ARG2:%.+]] = getelementptr inbounds [9 x i8*], [9 x i8*]* [[PTR_ARR2]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 220 // CHECK: [[SIZES_ARG2:%.+]] = getelementptr inbounds [9 x i[[SZ]]], [9 x i[[SZ]]]* [[SIZET2]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 221 // CHECK: {{.+}} = call i32 @__tgt_target(i32 -1, {{.+}}, i32 9, i8** [[BASE_PTR_GEP_ARG2]], i8** [[PTR_GEP_ARG2]], i[[SZ]]* [[SIZES_ARG2]], i32* getelementptr inbounds ([9 x i32], [9 x i32]* [[MAPT2]], i32 0, i32 0)) 222 223 // make sure that firstprivate variables are generated in all cases and that we use those instances for operations inside the 224 // target region 225 // TCHECK: define void @__omp_offloading_{{.+}}(i{{[0-9]+}} [[A2_IN:%.+]], [10 x float]* {{.+}} [[B_IN:%.+]], i{{[0-9]+}} [[BN_SZ:%.+]], float* {{.+}} [[BN_IN:%.+]], [5 x [10 x double]]* {{.+}} [[C_IN:%.+]], i{{[0-9]+}} [[CN_SZ1:%.+]], i{{[0-9]+}} [[CN_SZ2:%.+]], double* {{.+}} [[CN_IN:%.+]], [[TT]]* {{.+}} [[D_IN:%.+]]) 226 // TCHECK: [[A2_ADDR:%.+]] = alloca i{{[0-9]+}}, 227 // TCHECK: [[B_ADDR:%.+]] = alloca [10 x float]*, 228 // TCHECK: [[VLA_ADDR:%.+]] = alloca i{{[0-9]+}}, 229 // TCHECK: [[BN_ADDR:%.+]] = alloca float*, 230 // TCHECK: [[C_ADDR:%.+]] = alloca [5 x [10 x double]]*, 231 // TCHECK: [[VLA_ADDR2:%.+]] = alloca i{{[0-9]+}}, 232 // TCHECK: [[VLA_ADDR4:%.+]] = alloca i{{[0-9]+}}, 233 // TCHECK: [[CN_ADDR:%.+]] = alloca double*, 234 // TCHECK: [[D_ADDR:%.+]] = alloca [[TT]]*, 235 // TCHECK-NOT: alloca i{{[0-9]+}}, 236 // TCHECK: [[B_PRIV:%.+]] = alloca [10 x float], 237 // TCHECK: [[SSTACK:%.+]] = alloca i8*, 238 // TCHECK: [[C_PRIV:%.+]] = alloca [5 x [10 x double]], 239 // TCHECK: [[D_PRIV:%.+]] = alloca [[TT]], 240 // TCHECK: store i{{[0-9]+}} [[A2_IN]], i{{[0-9]+}}* [[A2_ADDR]], 241 // TCHECK: store [10 x float]* [[B_IN]], [10 x float]** [[B_ADDR]], 242 // TCHECK: store i{{[0-9]+}} [[BN_SZ]], i{{[0-9]+}}* [[VLA_ADDR]], 243 // TCHECK: store float* [[BN_IN]], float** [[BN_ADDR]], 244 // TCHECK: store [5 x [10 x double]]* [[C_IN]], [5 x [10 x double]]** [[C_ADDR]], 245 // TCHECK: store i{{[0-9]+}} [[CN_SZ1]], i{{[0-9]+}}* [[VLA_ADDR2]], 246 // TCHECK: store i{{[0-9]+}} [[CN_SZ2]], i{{[0-9]+}}* [[VLA_ADDR4]], 247 // TCHECK: store double* [[CN_IN]], double** [[CN_ADDR]], 248 // TCHECK: store [[TT]]* [[D_IN]], [[TT]]** [[D_ADDR]], 249 // TCHECK: [[CONV_A2ADDR:%.+]] = bitcast i{{[0-9]+}}* [[A2_ADDR]] to i{{[0-9]+}}* 250 // TCHECK: [[B_ADDR_REF:%.+]] = load [10 x float]*, [10 x float]** [[B_ADDR]], 251 // TCHECK: [[BN_SZ_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[VLA_ADDR]], 252 // TCHECK: [[BN_ADDR_REF:%.+]] = load float*, float** [[BN_ADDR]], 253 // TCHECK: [[C_ADDR_REF:%.+]] = load [5 x [10 x double]]*, [5 x [10 x double]]** [[C_ADDR]], 254 // TCHECK: [[CN_SZ1_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[VLA_ADDR2]], 255 // TCHECK: [[CN_SZ2_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[VLA_ADDR4]], 256 // TCHECK: [[CN_ADDR_REF:%.+]] = load double*, double** [[CN_ADDR]], 257 // TCHECK: [[D_ADDR_REF:%.+]] = load [[TT]]*, [[TT]]** [[D_ADDR]], 258 259 // firstprivate(aa): a_priv = a_in 260 // TCHECK-NOT: store i{{[0-9]+}} % 261 262 // firstprivate(b): memcpy(b_priv,b_in) 263 // TCHECK: [[B_PRIV_BCAST:%.+]] = bitcast [10 x float]* [[B_PRIV]] to i8* 264 // TCHECK: [[B_ADDR_REF_BCAST:%.+]] = bitcast [10 x float]* [[B_ADDR_REF]] to i8* 265 // TCHECK: call void @llvm.memcpy.{{.+}}(i8* [[B_PRIV_BCAST]], i8* [[B_ADDR_REF_BCAST]], {{.+}}) 266 267 // TCHECK: [[RET_STACK:%.+]] = call i8* @llvm.stacksave() 268 // TCHECK: store i8* [[RET_STACK]], i8** [[SSTACK]], 269 270 // firstprivate(bn) 271 // TCHECK: [[BN_PRIV:%.+]] = alloca float, i{{[0-9]+}} [[BN_SZ_VAL]], 272 // TCHECK: [[BN_COPY_SZ:%.+]] = mul{{.+}} i{{[0-9]+}} [[BN_SZ_VAL]], 4 273 // TCHECK: [[BN_PRIV__BCAST:%.+]] = bitcast float* [[BN_PRIV]] to i8* 274 // TCHECK: [[BN_REF_IN_BCAST:%.+]] = bitcast float* [[BN_ADDR_REF]] to i8* 275 // TCHECK: call void @llvm.memcpy.{{.+}}(i8* [[BN_PRIV__BCAST]], i8* [[BN_REF_IN_BCAST]], i{{[0-9]+}} [[BN_COPY_SZ]],{{.+}}) 276 277 // firstprivate(c) 278 // TCHECK: [[C_PRIV_BCAST:%.+]] = bitcast [5 x [10 x double]]* [[C_PRIV]] to i8* 279 // TCHECK: [[C_IN_BCAST:%.+]] = bitcast [5 x [10 x double]]* [[C_ADDR_REF]] to i8* 280 // TCHECK: call void @llvm.memcpy.{{.+}}(i8* [[C_PRIV_BCAST]], i8* [[C_IN_BCAST]],{{.+}}) 281 282 // firstprivate(cn) 283 // TCHECK: [[CN_SZ:%.+]] = mul{{.+}} i{{[0-9]+}} [[CN_SZ1_VAL]], [[CN_SZ2_VAL]] 284 // TCHECK: [[CN_PRIV:%.+]] = alloca double, i{{[0-9]+}} [[CN_SZ]], 285 // TCHECK: [[CN_SZ2:%.+]] = mul{{.+}} i{{[0-9]+}} [[CN_SZ1_VAL]], [[CN_SZ2_VAL]] 286 // TCHECK: [[CN_SZ2_CPY:%.+]] = mul{{.+}} i{{[0-9]+}} [[CN_SZ2]], 8 287 // TCHECK: [[CN_PRIV_BCAST:%.+]] = bitcast double* [[CN_PRIV]] to i8* 288 // TCHECK: [[CN_IN_BCAST:%.+]] = bitcast double* [[CN_ADDR_REF]] to i8* 289 // TCHECK: call void @llvm.memcpy.{{.+}}(i8* [[CN_PRIV_BCAST]], i8* [[CN_IN_BCAST]], i{{[0-9]+}} [[CN_SZ2_CPY]],{{.+}}) 290 291 // firstprivate(d) 292 // TCHECK: [[D_PRIV_BCAST:%.+]] = bitcast [[TT]]* [[D_PRIV]] to i8* 293 // TCHECK: [[D_IN_BCAST:%.+]] = bitcast [[TT]]* [[D_ADDR_REF]] to i8* 294 // TCHECK: call void @llvm.memcpy.{{.+}}(i8* [[D_PRIV_BCAST]], i8* [[D_IN_BCAST]],{{.+}}) 295 296 297 #pragma omp target firstprivate(ptr) 298 { 299 ptr[0]++; 300 } 301 // CHECK: [[PTR_ADDR_REF:%.+]] = load double*, double** [[PTR_ADDR]], 302 // CHECK: [[PTR_ADDR_BCAST:%.+]] = bitcast double* [[PTR_ADDR_REF]] to i8* 303 304 // CHECK: [[BASE_PTR_GEP3_0:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BASE_PTR_ARR3]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 305 // CHECK: store i8* [[PTR_ADDR_BCAST]], i8** [[BASE_PTR_GEP3_0]], 306 // CHECK: [[PTR_ADDR_BCAST2:%.+]] = bitcast double* [[PTR_ADDR_REF]] to i8* 307 // CHECK: [[PTR_GEP3_0:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PTR_ARR3]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 308 // CHECK: store i8* [[PTR_ADDR_BCAST2]], i8** [[PTR_GEP3_0]], 309 310 // CHECK: [[BASE_PTR_GEP_ARG3:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[BASE_PTR_ARR3]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 311 // CHECK: [[PTR_GEP_ARG3:%.+]] = getelementptr inbounds [1 x i8*], [1 x i8*]* [[PTR_ARR3]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 312 // CHECK: {{.+}} = call i32 @__tgt_target(i32 -1, {{.+}}, i32 1, i8** [[BASE_PTR_GEP_ARG3]], i8** [[PTR_GEP_ARG3]], i[[SZ]]* getelementptr inbounds ([1 x i[[SZ]]], [1 x i[[SZ]]]* [[SIZET3]], i32 0, i32 0), i32* getelementptr inbounds ([1 x i32], [1 x i32]* [[MAPT3]], i32 0, i32 0)) 313 314 // TCHECK: define void @__omp_offloading_{{.+}}(double* [[PTR_IN:%.+]]) 315 // TCHECK: [[PTR_ADDR:%.+]] = alloca double*, 316 // TCHECK-NOT: alloca double*, 317 // TCHECK: store double* [[PTR_IN]], double** [[PTR_ADDR]], 318 // TCHECK-NOT: store double* % 319 320 return a; 321 } 322 323 324 template<typename tx> 325 tx ftemplate(int n) { 326 tx a = 0; 327 tx b[10]; 328 329 #pragma omp target firstprivate(a,b) 330 { 331 a += 1; 332 b[2] += 1; 333 } 334 335 return a; 336 } 337 338 static 339 int fstatic(int n) { 340 int a = 0; 341 char aaa = 0; 342 int b[10]; 343 344 #pragma omp target firstprivate(a,aaa,b) 345 { 346 a += 1; 347 aaa += 1; 348 b[2] += 1; 349 } 350 351 return a; 352 } 353 354 // TCHECK: define void @__omp_offloading_{{.+}}(i{{[0-9]+}} [[A_IN:%.+]], i{{[0-9]+}} [[A3_IN:%.+]], [10 x i{{[0-9]+}}]*{{.+}} [[B_IN:%.+]]) 355 // TCHECK: [[A_ADDR:%.+]] = alloca i{{[0-9]+}}, 356 // TCHECK: [[A3_ADDR:%.+]] = alloca i{{[0-9]+}}, 357 // TCHECK: [[B_ADDR:%.+]] = alloca [10 x i{{[0-9]+}}]*, 358 // TCHECK-NOT: alloca i{{[0-9]+}}, 359 // TCHECK: [[B_PRIV:%.+]] = alloca [10 x i{{[0-9]+}}], 360 // TCHECK: store i{{[0-9]+}} [[A_IN]], i{{[0-9]+}}* [[A_ADDR]], 361 // TCHECK: store i{{[0-9]+}} [[A3_IN]], i{{[0-9]+}}* [[A3_ADDR]], 362 // TCHECK: store [10 x i{{[0-9]+}}]* [[B_IN]], [10 x i{{[0-9]+}}]** [[B_ADDR]], 363 // TCHECK-64: [[A_CONV:%.+]] = bitcast i{{[0-9]+}}* [[A_ADDR]] to i{{[0-9]+}}* 364 // TCHECK: [[A3_CONV:%.+]] = bitcast i{{[0-9]+}}* [[A3_ADDR]] to i8* 365 // TCHECK: [[B_ADDR_REF:%.+]] = load [10 x i{{[0-9]+}}]*, [10 x i{{[0-9]+}}]** [[B_ADDR]], 366 367 // firstprivate(a): a_priv = a_in 368 369 // firstprivate(aaa) 370 // TCHECK-NOT: store i{{[0-9]+}} % 371 372 // firstprivate(b) 373 // TCHECK: [[B_PRIV_BCAST:%.+]] = bitcast [10 x i{{[0-9]+}}]* [[B_PRIV]] to i8* 374 // TCHECK: [[B_IN_BCAST:%.+]] = bitcast [10 x i{{[0-9]+}}]* [[B_ADDR_REF]] to i8* 375 // TCHECK: call void @llvm.memcpy.{{.+}}(i8* [[B_PRIV_BCAST]], i8* [[B_IN_BCAST]],{{.+}}) 376 377 // TCHECK: ret void 378 379 struct S1 { 380 double a; 381 382 int r1(int n){ 383 int b = n+1; 384 short int c[2][n]; 385 386 #pragma omp target firstprivate(b,c) 387 { 388 this->a = (double)b + 1.5; 389 c[1][1] = ++a; 390 } 391 392 return c[1][1] + (int)b; 393 } 394 395 // on the host side, we first generate r1, then the static function and the template above 396 // CHECK: define{{.+}} i32 {{.+}}([[S1]]* {{.+}}, i{{[0-9]+}} {{.+}}) 397 // CHECK: [[BASE_PTRS4:%.+]] = alloca [5 x i8*], 398 // CHECK: [[PTRS4:%.+]] = alloca [5 x i8*], 399 // CHECK: [[SIZET4:%.+]] = alloca [5 x i{{[0-9]+}}], 400 401 // map(this: this ptr is implicitly captured (not firstprivate matter) 402 // CHECK: {{.+}} = getelementptr inbounds [5 x i8*], [5 x i8*]* [[BASE_PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 403 // CHECK: store {{.+}}, {{.+}}, 404 // CHECK: {{.+}} = getelementptr inbounds [5 x i8*], [5 x i8*]* [[PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 405 // CHECK: store {{.+}}, {{.+}}, 406 // CHECK: {{.+}} getelementptr inbounds [5 x i{{[0-9]+}}], [5 x i{{[0-9]+}}]* [[SIZET4]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 407 // CHECK: store {{.+}}, {{.+}} 408 409 // firstprivate(b): base_ptr = b, ptr = b, size = 4 (pass by-value) 410 // CHECK: [[B_CAST_PTR:%.+]] = inttoptr i{{[0-9]+}} [[B_CAST:%.+]] to i8* 411 // CHECK: [[BASE_PTRS_GEP4_1:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[BASE_PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 412 // CHECK: store i8* [[B_CAST_PTR]], i8** [[BASE_PTRS_GEP4_1]], 413 // CHECK: [[B_CAST_PTR2:%.+]] = inttoptr i{{[0-9]+}} [[B_CAST:%.+]] to i8* 414 // CHECK: [[PTRS_GEP4_1:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 415 // CHECK: store i8* [[B_CAST_PTR2]], i8** [[PTRS_GEP4_1]], 416 // CHECK: [[SIZES_GEP4_1:%.+]] = getelementptr inbounds [5 x i{{[0-9]+}}], [5 x i{{[0-9]+}}]* [[SIZET4]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 417 // CHECK: store i{{[0-9]+}} 4, i{{[0-9]+}}* [[SIZES_GEP4_1]], 418 419 // firstprivate(c), 3 entries: 2, n, c 420 // CHECK: [[BASE_PTRS_GEP4_2:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[BASE_PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 421 // CHECK: store i8* inttoptr (i{{[0-9]+}} 2 to i8*), i8** [[BASE_PTRS_GEP4_2]], 422 // CHECK: [[PTRS_GEP4_2:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 423 // CHECK: store i8* inttoptr (i{{[0-9]+}} 2 to i8*), i8** [[PTRS_GEP4_2]], 424 // CHECK: [[SIZES_GEP4_2:%.+]] = getelementptr inbounds [5 x i{{[0-9]+}}], [5 x i{{[0-9]+}}]* [[SIZET4]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 425 // CHECK-64: store i{{[0-9]+}} 8, i{{[0-9]+}}* [[SIZES_GEP4_2]], 426 // CHECK-32: store i{{[0-9]+}} 4, i{{[0-9]+}}* [[SIZES_GEP4_2]], 427 // CHECK: [[N_PTR:%.+]] = inttoptr i{{[0-9]+}} [[N:%.+]] to i8* 428 // CHECK: [[BASE_PTRS_GEP4_3:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[BASE_PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 429 // CHECK: store i8* [[N_PTR]], i8** [[BASE_PTRS_GEP4_3]], 430 // CHECK: [[N_PTR2:%.+]] = inttoptr i{{[0-9]+}} [[N:%.+]] to i8* 431 // CHECK: [[PTRS_GEP4_3:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 432 // CHECK: store i8* [[N_PTR2]], i8** [[PTRS_GEP4_3]], 433 // CHECK: [[SIZES_GEP4_3:%.+]] = getelementptr inbounds [5 x i{{[0-9]+}}], [5 x i{{[0-9]+}}]* [[SIZET4]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 434 // CHECK-64: store i{{[0-9]+}} 8, i{{[0-9]+}}* [[SIZES_GEP4_3]], 435 // CHECK-32: store i{{[0-9]+}} 4, i{{[0-9]+}}* [[SIZES_GEP4_3]], 436 // CHECK: [[B_BCAST:%.+]] = bitcast i{{[0-9]+}}* [[B:%.+]] to i8* 437 // CHECK: [[BASE_PTRS_GEP4_4:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[BASE_PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 4 438 // CHECK: store i8* [[B_BCAST]], i8** [[BASE_PTRS_GEP4_4]], 439 // CHECK: [[B_BCAST2:%.+]] = bitcast i{{[0-9]+}}* [[B:%.+]] to i8* 440 // CHECK: [[PTRS_GEP4_4:%.+]] = getelementptr inbounds [5 x i8*], [5 x i8*]* [[PTRS4]], i{{[0-9]+}} 0, i{{[0-9]+}} 4 441 // CHECK: store i8* [[B_BCAST2]], i8** [[PTRS_GEP4_4]], 442 // CHECK: [[SIZES_GEP4_4:%.+]] = getelementptr inbounds [5 x i{{[0-9]+}}], [5 x i{{[0-9]+}}]* [[SIZET4]], i{{[0-9]+}} 0, i{{[0-9]+}} 4 443 // CHECK: store i{{[0-9]+}} [[B_SIZE:%.+]], i{{[0-9]+}}* [[SIZES_GEP4_4]], 444 445 // only check that we use the map types stored in the global variable 446 // CHECK: call i32 @__tgt_target(i32 -1, {{.+}}, i32 5, i8** {{.+}}, i8** {{.+}}, i{{[0-9]+}}* {{.+}}, i32* getelementptr inbounds ([5 x i32], [5 x i32]* [[MAPT4]], i32 0, i32 0)) 447 448 // TCHECK: define void @__omp_offloading_{{.+}}([[S1]]* [[TH:%.+]], i{{[0-9]+}} [[B_IN:%.+]], i{{[0-9]+}} [[VLA:%.+]], i{{[0-9]+}} [[VLA1:%.+]], i{{[0-9]+}}{{.+}} [[C_IN:%.+]]) 449 // TCHECK: [[TH_ADDR:%.+]] = alloca [[S1]]*, 450 // TCHECK: [[B_ADDR:%.+]] = alloca i{{[0-9]+}}, 451 // TCHECK: [[VLA_ADDR:%.+]] = alloca i{{[0-9]+}}, 452 // TCHECK: [[VLA_ADDR2:%.+]] = alloca i{{[0-9]+}}, 453 // TCHECK: [[C_ADDR:%.+]] = alloca i{{[0-9]+}}*, 454 // TCHECK-NOT: alloca i{{[0-9]+}}, 455 // TCHECK: [[SSTACK:%.+]] = alloca i8*, 456 457 // TCHECK: store [[S1]]* [[TH]], [[S1]]** [[TH_ADDR]], 458 // TCHECK: store i{{[0-9]+}} [[B_IN]], i{{[0-9]+}}* [[B_ADDR]], 459 // TCHECK: store i{{[0-9]+}} [[VLA]], i{{[0-9]+}}* [[VLA_ADDR]], 460 // TCHECK: store i{{[0-9]+}} [[VLA1]], i{{[0-9]+}}* [[VLA_ADDR2]], 461 // TCHECK: store i{{[0-9]+}}* [[C_IN]], i{{[0-9]+}}** [[C_ADDR]], 462 // TCHECK: [[TH_ADDR_REF:%.+]] = load [[S1]]*, [[S1]]** [[TH_ADDR]], 463 // TCHECK-64: [[B_ADDR_CONV:%.+]] = bitcast i{{[0-9]+}}* [[B_ADDR]] to i{{[0-9]+}}* 464 // TCHECK: [[VLA_ADDR_REF:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[VLA_ADDR]], 465 // TCHECK: [[VLA_ADDR_REF2:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[VLA_ADDR2]], 466 // TCHECK: [[C_ADDR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[C_ADDR]], 467 468 // firstprivate(b) 469 // TCHECK-NOT: store i{{[0-9]+}} % 470 471 // TCHECK: [[RET_STACK:%.+]] = call i8* @llvm.stacksave() 472 // TCHECK: store i8* [[RET_STACK:%.+]], i8** [[SSTACK]], 473 474 // firstprivate(c) 475 // TCHECK: [[C_SZ:%.+]] = mul{{.+}} i{{[0-9]+}} [[VLA_ADDR_REF]], [[VLA_ADDR_REF2]] 476 // TCHECK: [[C_PRIV:%.+]] = alloca i{{[0-9]+}}, i{{[0-9]+}} [[C_SZ]], 477 // TCHECK: [[C_SZ2:%.+]] = mul{{.+}} i{{[0-9]+}} [[VLA_ADDR_REF]], [[VLA_ADDR_REF2]] 478 // TCHECK: [[C_SZ_CPY:%.+]] = mul{{.+}} i{{[0-9]+}} [[C_SZ2]], 2 479 // TCHECK: [[C_PRIV_BCAST:%.+]] = bitcast i{{[0-9]+}}* [[C_PRIV]] to i8* 480 // TCHECK: [[C_IN_BCAST:%.+]] = bitcast i{{[0-9]+}}* [[C_ADDR_REF]] to i8* 481 // TCHECK: call void @llvm.memcpy.{{.+}}(i8* [[C_PRIV_BCAST]], i8* [[C_IN_BCAST]],{{.+}}) 482 483 // finish 484 // TCHECK: [[RELOAD_SSTACK:%.+]] = load i8*, i8** [[SSTACK]], 485 // TCHECK: call void @llvm.stackrestore(i8* [[RELOAD_SSTACK]]) 486 // TCHECK: ret void 487 488 489 // static host function 490 // CHECK: define{{.+}} i32 {{.+}}(i{{[0-9]+}} {{.+}}) 491 // CHECK: [[BASE_PTRS5:%.+]] = alloca [3 x i8*], 492 // CHECK: [[PTRS5:%.+]] = alloca [3 x i8*], 493 494 // firstprivate(a): by value 495 // CHECK: [[A_CAST_PTR:%.+]] = inttoptr i{{[0-9]+}} [[A_CAST:%.+]] to i8* 496 // CHECK: [[BASE_PTRS_GEP5_0:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[BASE_PTRS5]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 497 // CHECK: store i8* [[A_CAST_PTR]], i8** [[BASE_PTRS_GEP5_0]], 498 // CHECK: [[A_CAST_PTR2:%.+]] = inttoptr i{{[0-9]+}} [[A_CAST:%.+]] to i8* 499 // CHECK: [[PTRS_GEP5_0:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[PTRS5]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 500 // CHECK: store i8* [[A_CAST_PTR2]], i8** [[PTRS_GEP5_0]], 501 502 // firstprivate(aaa): by value 503 // CHECK: [[A3_CAST_PTR:%.+]] = inttoptr i{{[0-9]+}} [[A3_CAST:%.+]] to i8* 504 // CHECK: [[BASE_PTRS_GEP5_1:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[BASE_PTRS5]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 505 // CHECK: store i8* [[A3_CAST_PTR]], i8** [[BASE_PTRS_GEP5_1]], 506 // CHECK: [[A3_CAST_PTR2:%.+]] = inttoptr i{{[0-9]+}} [[A3_CAST:%.+]] to i8* 507 // CHECK: [[PTRS_GEP5_1:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[PTRS5]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 508 // CHECK: store i8* [[A3_CAST_PTR2]], i8** [[PTRS_GEP5_1]], 509 510 // firstprivate(b): base_ptr = &b[0], ptr= &b[0] 511 // CHECK: [[B_BCAST:%.+]] = bitcast [10 x i{{[0-9]+}}]* [[B:%.+]] to i8* 512 // CHECK: [[BASE_PTRS_GEP5_2:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[BASE_PTRS5]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 513 // CHECK: store i8* [[B_BCAST]], i8** [[BASE_PTRS_GEP5_2]], 514 // CHECK: [[B_BCAST2:%.+]] = bitcast [10 x i{{[0-9]+}}]* [[B:%.+]] to i8* 515 // CHECK: [[PTRS_GEP5_2:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[PTRS5]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 516 // CHECK: store i8* [[B_BCAST2]], i8** [[PTRS_GEP5_2]], 517 518 // only check that the right sizes and map types are used 519 // CHECK: call i32 @__tgt_target(i32 -1, {{.+}}, i32 3, i8** {{.+}}, i8** {{.+}}, i[[SZ]]* getelementptr inbounds ([3 x i[[SZ]]], [3 x i[[SZ]]]* [[SIZET5]], i32 0, i32 0), i32* getelementptr inbounds ([3 x i32], [3 x i32]* [[MAPT5]], i32 0, i32 0)) 520 }; 521 522 523 524 int bar(int n, double *ptr){ 525 int a = 0; 526 a += foo(n, ptr); 527 S1 S; 528 a += S.r1(n); 529 a += fstatic(n); 530 a += ftemplate<int>(n); 531 532 return a; 533 } 534 535 // template host and device 536 537 // CHECK: define{{.+}} i32 {{.+}}(i{{[0-9]+}} {{.+}}) 538 // CHECK: [[BASE_PTRS6:%.+]] = alloca [2 x i8*], 539 // CHECK: [[PTRS6:%.+]] = alloca [2 x i8*], 540 541 // firstprivate(a): by value 542 // CHECK: [[AT_CAST_PTR:%.+]] = inttoptr i{{[0-9]+}} [[AT_CAST:%.+]] to i8* 543 // CHECK: [[BASE_PTRS_GEP6_0:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BASE_PTRS6]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 544 // CHECK: store i8* [[AT_CAST_PTR]], i8** [[BASE_PTRS_GEP6_0]], 545 // CHECK: [[AT_CAST_PTR2:%.+]] = inttoptr i{{[0-9]+}} [[AT_CAST:%.+]] to i8* 546 // CHECK: [[PTRS_GEP6_0:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[PTRS6]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 547 // CHECK: store i8* [[AT_CAST_PTR2]], i8** [[PTRS_GEP6_0]], 548 549 // firstprivate(b): pointer 550 // CHECK: [[B_BCAST:%.+]] = bitcast [10 x i{{[0-9]+}}]* [[B:%.+]] to i8* 551 // CHECK: [[BASE_PTRS_GEP6_1:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[BASE_PTRS6]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 552 // CHECK: store i8* [[B_BCAST]], i8** [[BASE_PTRS_GEP6_1]], 553 // CHECK: [[B_BCAST2:%.+]] = bitcast [10 x i{{[0-9]+}}]* [[B:%.+]] to i8* 554 // CHECK: [[PTRS_GEP6_1:%.+]] = getelementptr inbounds [2 x i8*], [2 x i8*]* [[PTRS6]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 555 // CHECK: store i8* [[B_BCAST2]], i8** [[PTRS_GEP6_1]], 556 557 // CHECK: call i32 @__tgt_target(i32 -1, {{.+}}, i32 2, i8** {{.+}}, i8** {{.+}}, i[[SZ]]* getelementptr inbounds ([2 x i[[SZ]]], [2 x i[[SZ]]]* [[SIZET6]], i32 0, i32 0), i32* getelementptr inbounds ([2 x i32], [2 x i32]* [[MAPT6]], i32 0, i32 0)) 558 559 560 // TCHECK: define void @__omp_offloading_{{.+}}(i{{[0-9]+}} [[A_IN:%.+]], [10 x i{{[0-9]+}}]*{{.+}} [[B_IN:%.+]]) 561 // TCHECK: [[A_ADDR:%.+]] = alloca i{{[0-9]+}}, 562 // TCHECK: [[B_ADDR:%.+]] = alloca [10 x i{{[0-9]+}}]*, 563 // TCHECK-NOT: alloca i{{[0-9]+}}, 564 // TCHECK: [[B_PRIV:%.+]] = alloca [10 x i{{[0-9]+}}], 565 // TCHECK: store i{{[0-9]+}} [[A_IN]], i{{[0-9]+}}* [[A_ADDR]], 566 // TCHECK: store [10 x i{{[0-9]+}}]* [[B_IN]], [10 x i{{[0-9]+}}]** [[B_ADDR]], 567 // TCHECK-64: [[A_ADDR_CONV:%.+]] = bitcast i{{[0-9]+}}* [[A_ADDR]] to i{{[0-9]+}}* 568 // TCHECK: [[B_ADDR_REF:%.+]] = load [10 x i{{[0-9]+}}]*, [10 x i{{[0-9]+}}]** [[B_ADDR]], 569 570 // firstprivate(a) 571 // TCHECK-NOT: store i{{[0-9]+}} % 572 573 // firstprivate(b) 574 // TCHECK: [[B_PRIV_BCAST:%.+]] = bitcast [10 x i{{[0-9]+}}]* [[B_PRIV]] to i8* 575 // TCHECK: [[B_IN_BCAST:%.+]] = bitcast [10 x i{{[0-9]+}}]* [[B_ADDR_REF]] to i8* 576 // TCHECK: call void @llvm.memcpy.{{.+}}(i8* [[B_PRIV_BCAST]], i8* [[B_IN_BCAST]],{{.+}}) 577 578 // TCHECK: ret void 579 580 #endif 581
