868 lines · cpp
1// RUN: %clang_cc1 -verify -fopenmp -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -fexceptions -fcxx-exceptions -o - -fopenmp-version=45 | FileCheck %s2// RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -emit-pch -o %t %s -fopenmp-version=453// RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -debug-info-kind=limited -std=c++11 -include-pch %t -verify %s -emit-llvm -o - -fopenmp-version=45 | FileCheck %s4// RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp -fexceptions -fcxx-exceptions -debug-info-kind=line-tables-only -x c++ -emit-llvm %s -o - -fopenmp-version=45 | FileCheck %s --check-prefix=TERM_DEBUG5// RUN: %clang_cc1 -verify -fopenmp -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -fexceptions -fcxx-exceptions -o - -fopenmp-version=50 -DOMP5 | FileCheck %s --check-prefix=CHECK --check-prefix=OMP5 --check-prefix=OMP5RT6// RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -emit-pch -o %t %s -fopenmp-version=50 -DOMP57// RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -debug-info-kind=limited -std=c++11 -include-pch %t -verify %s -emit-llvm -o - -fopenmp-version=50 -DOMP5 | FileCheck %s --check-prefix=CHECK --check-prefix=OMP5 --check-prefix=OMP5RT8 9// RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -fexceptions -fcxx-exceptions -o - -fopenmp-version=45 | FileCheck %s10// RUN: %clang_cc1 -fopenmp-simd -x c++ -std=c++11 -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -emit-pch -o %t %s -fopenmp-version=4511// RUN: %clang_cc1 -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -debug-info-kind=limited -std=c++11 -include-pch %t -verify %s -emit-llvm -o - -fopenmp-version=45 | FileCheck %s12// RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp-simd -fexceptions -fcxx-exceptions -debug-info-kind=line-tables-only -x c++ -emit-llvm %s -o - -fopenmp-version=45 | FileCheck --check-prefix=TERM_DEBUG %s13// RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -fexceptions -fcxx-exceptions -o - -fopenmp-version=50 -DOMP5 | FileCheck %s --check-prefix=CHECK --check-prefix=OMP514// RUN: %clang_cc1 -fopenmp-simd -x c++ -std=c++11 -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -emit-pch -o %t %s -fopenmp-version=50 -DOMP515// RUN: %clang_cc1 -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -debug-info-kind=limited -std=c++11 -include-pch %t -verify %s -emit-llvm -o - -fopenmp-version=50 -DOMP5 | FileCheck %s --check-prefix=CHECK --check-prefix=OMP516// expected-no-diagnostics17 #ifndef HEADER18 #define HEADER19 20#ifdef OMP521#define CONDITIONAL conditional :22#else23#define CONDITIONAL24#endif //OMP525// CHECK: [[SS_TY:%.+]] = type { i32 }26// OMP5-DAG: [[LAST_IV:@.+]] = {{.*}}common global i64 027// OMP5-DAG: [[LAST_A:@.+]] = {{.*}}common global i32 028 29long long get_val() { extern void mayThrow(); mayThrow(); return 0; }30double *g_ptr;31 32struct S {33 int a, b;34};35 36// CHECK-LABEL: define {{.*void}} @{{.*}}simple{{.*}}(ptr noundef {{.+}}, ptr noundef {{.+}}, ptr noundef {{.+}}, ptr noundef {{.+}})37void simple(float *a, float *b, float *c, float *d) {38 S s, *p;39#ifdef OMP540 #pragma omp simd if (simd: true) nontemporal(a, b, c, d, s)41#else42 #pragma omp simd43#endif44 // CHECK: store i32 0, ptr [[OMP_IV:%[^,]+]]45 46// CHECK: [[IV:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group47// CHECK-NEXT: [[CMP:%.+]] = icmp slt i32 [[IV]], 648// CHECK-NEXT: br i1 [[CMP]], label %[[SIMPLE_LOOP1_BODY:.+]], label %[[SIMPLE_LOOP1_END:[^,]+]]49 for (int i = 3; i < 32; i += 5) {50// CHECK: [[SIMPLE_LOOP1_BODY]]:51// Start of body: calculate i from IV:52// CHECK: [[IV1_1:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group53// CHECK: [[CALC_I_1:%.+]] = mul nsw i32 [[IV1_1]], 554// CHECK-NEXT: [[CALC_I_2:%.+]] = add nsw i32 3, [[CALC_I_1]]55// CHECK-NEXT: store i32 [[CALC_I_2]], ptr [[LC_I:.+]]{{.*}}!llvm.access.group56// ... loop body ...57// End of body: store into a[i]:58// OMP45-NOT: load ptr,{{.*}}!nontemporal59// CHECK-NOT: load float,{{.*}}!nontemporal60// OMP5: load ptr,{{.*}}!nontemporal61// OMP5: load ptr,{{.*}}!nontemporal62// OMP5: load ptr,{{.*}}!nontemporal63// OMP5: load i32,{{.*}}!nontemporal64// OMP5-NOT: load i32,{{.*}}!nontemporal65// OMP5: load ptr,{{.*}}!nontemporal66// CHECK-NOT: load float,{{.*}}!nontemporal67// CHECK: store float [[RESULT:%.+]], ptr {{%.+}}{{.*}}!llvm.access.group68 a[i] = b[i] * c[i] * d[i] + s.a + p->a;69// CHECK: [[IV1_2:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group70// CHECK-NEXT: [[ADD1_2:%.+]] = add nsw i32 [[IV1_2]], 171// CHECK-NEXT: store i32 [[ADD1_2]], ptr [[OMP_IV]]{{.*}}!llvm.access.group72// br label %{{.+}}, !llvm.loop !{{.+}}73 }74// CHECK: [[SIMPLE_LOOP1_END]]:75 76 long long k = get_val();77 78 #pragma omp simd linear(k : 3)79// CHECK: [[K0:%.+]] = call {{.*}}i64 @{{.*}}get_val80// CHECK-NEXT: store i64 [[K0]], ptr [[K_VAR:%[^,]+]]81// CHECK: store i32 0, ptr [[OMP_IV2:%[^,]+]]82// CHECK: [[K0LOAD:%.+]] = load i64, ptr [[K_VAR]]83// CHECK-NEXT: store i64 [[K0LOAD]], ptr [[LIN0:%[^,]+]]84 85// CHECK: [[IV2:%.+]] = load i32, ptr [[OMP_IV2]]{{.*}}!llvm.access.group86// CHECK-NEXT: [[CMP2:%.+]] = icmp slt i32 [[IV2]], 987// CHECK-NEXT: br i1 [[CMP2]], label %[[SIMPLE_LOOP2_BODY:.+]], label %[[SIMPLE_LOOP2_END:[^,]+]]88 for (int i = 10; i > 1; i--) {89// CHECK: [[SIMPLE_LOOP2_BODY]]:90// Start of body: calculate i from IV:91// CHECK: [[IV2_0:%.+]] = load i32, ptr [[OMP_IV2]]{{.*}}!llvm.access.group92// FIXME: It is interesting, why the following "mul 1" was not constant folded?93// CHECK-NEXT: [[IV2_1:%.+]] = mul nsw i32 [[IV2_0]], 194// CHECK-NEXT: [[LC_I_1:%.+]] = sub nsw i32 10, [[IV2_1]]95// CHECK-NEXT: store i32 [[LC_I_1]], ptr {{.+}}, !llvm.access.group96//97// CHECK-NEXT: [[LIN0_1:%.+]] = load i64, ptr [[LIN0]]{{.*}}!llvm.access.group98// CHECK-NEXT: [[IV2_2:%.+]] = load i32, ptr [[OMP_IV2]]{{.*}}!llvm.access.group99// CHECK-NEXT: [[LIN_MUL1:%.+]] = mul nsw i32 [[IV2_2]], 3100// CHECK-NEXT: [[LIN_EXT1:%.+]] = sext i32 [[LIN_MUL1]] to i64101// CHECK-NEXT: [[LIN_ADD1:%.+]] = add nsw i64 [[LIN0_1]], [[LIN_EXT1]]102// Update of the privatized version of linear variable!103// CHECK-NEXT: store i64 [[LIN_ADD1]], ptr [[K_PRIVATIZED:%[^,]+]]104 a[k]++;105 k = k + 3;106// CHECK: [[IV2_2:%.+]] = load i32, ptr [[OMP_IV2]]{{.*}}!llvm.access.group107// CHECK-NEXT: [[ADD2_2:%.+]] = add nsw i32 [[IV2_2]], 1108// CHECK-NEXT: store i32 [[ADD2_2]], ptr [[OMP_IV2]]{{.*}}!llvm.access.group109// br label {{.+}}, !llvm.loop ![[SIMPLE_LOOP2_ID]]110 }111// CHECK: [[SIMPLE_LOOP2_END]]:112//113// Update linear vars after loop, as the loop was operating on a private version.114// CHECK: [[LIN0_2:%.+]] = load i64, ptr [[K_PRIVATIZED]],115// CHECK-NEXT: store i64 [[LIN0_2]], ptr [[K_VAR]]116//117 118 int lin = 12;119 #pragma omp simd linear(lin : get_val()), linear(g_ptr)120 121// Init linear private var.122// CHECK: store i32 12, ptr [[LIN_VAR:%[^,]+]]123// CHECK: store i64 0, ptr [[OMP_IV3:%[^,]+]]124 125// CHECK: [[LIN_LOAD:%.+]] = load i32, ptr [[LIN_VAR]]126// CHECK-NEXT: store i32 [[LIN_LOAD]], ptr [[LIN_START:%[^,]+]]127// Remember linear step.128// CHECK: [[CALL_VAL:%.+]] = invoke129// CHECK: store i64 [[CALL_VAL]], ptr [[LIN_STEP:%[^,]+]]130 131// CHECK: [[GLIN_LOAD:%.+]] = load ptr, ptr [[GLIN_VAR:@[^,]+]]132// CHECK-NEXT: store ptr [[GLIN_LOAD]], ptr [[GLIN_START:%[^,]+]]133 134// CHECK: [[IV3:%.+]] = load i64, ptr [[OMP_IV3]]{{.*}}!llvm.access.group135// CHECK-NEXT: [[CMP3:%.+]] = icmp ult i64 [[IV3]], 4136// CHECK-NEXT: br i1 [[CMP3]], label %[[SIMPLE_LOOP3_BODY:.+]], label %[[SIMPLE_LOOP3_END:[^,]+]]137 for (unsigned long long it = 2000; it >= 600; it-=400) {138// CHECK: [[SIMPLE_LOOP3_BODY]]:139// Start of body: calculate it from IV:140// CHECK: [[IV3_0:%.+]] = load i64, ptr [[OMP_IV3]]{{.*}}!llvm.access.group141// CHECK-NEXT: [[LC_IT_1:%.+]] = mul i64 [[IV3_0]], 400142// CHECK-NEXT: [[LC_IT_2:%.+]] = sub i64 2000, [[LC_IT_1]]143// CHECK-NEXT: store i64 [[LC_IT_2]], ptr {{.+}}, !llvm.access.group144//145// Linear start and step are used to calculate current value of the linear variable.146// CHECK: [[LINSTART:.+]] = load i32, ptr [[LIN_START]]{{.*}}!llvm.access.group147// CHECK: [[LINSTEP:.+]] = load i64, ptr [[LIN_STEP]]{{.*}}!llvm.access.group148// CHECK-NOT: store i32 {{.+}}, ptr [[LIN_VAR]],{{.*}}!llvm.access.group149// CHECK: store i32 {{.+}}, ptr [[LIN_PRIV:%.+]],{{.*}}!llvm.access.group150// CHECK: [[GLINSTART:.+]] = load ptr, ptr [[GLIN_START]]{{.*}}!llvm.access.group151// CHECK-NEXT: [[IV3_1:%.+]] = load i64, ptr [[OMP_IV3]]{{.*}}!llvm.access.group152// CHECK-NEXT: [[MUL:%.+]] = mul i64 [[IV3_1]], 1153// CHECK: [[GEP:%.+]] = getelementptr{{.*}}[[GLINSTART]]154// CHECK-NEXT: store ptr [[GEP]], ptr [[G_PTR_CUR:%[^,]+]]{{.*}}!llvm.access.group155 *g_ptr++ = 0.0;156// CHECK: [[GEP_VAL:%.+]] = load ptr, ptr [[G_PTR_CUR]]{{.*}}!llvm.access.group157// CHECK: store double{{.*}}[[GEP_VAL]]{{.*}}!llvm.access.group158 a[it + lin]++;159// CHECK: [[FLT_INC:%.+]] = fadd float160// CHECK-NEXT: store float [[FLT_INC]],{{.*}}!llvm.access.group161// CHECK: [[IV3_2:%.+]] = load i64, ptr [[OMP_IV3]]{{.*}}!llvm.access.group162// CHECK-NEXT: [[ADD3_2:%.+]] = add i64 [[IV3_2]], 1163// CHECK-NEXT: store i64 [[ADD3_2]], ptr [[OMP_IV3]]{{.*}}!llvm.access.group164 }165// CHECK: [[SIMPLE_LOOP3_END]]:166//167// Linear start and step are used to calculate final value of the linear variables.168// CHECK: [[LIN:%.+]] = load i32, ptr [[LIN_PRIV]]169// CHECK-NEXT: store i32 [[LIN]], ptr [[LIN_VAR]],170// CHECK: [[GLIN:%.+]] = load ptr, ptr [[G_PTR_CUR]]171// CHECK-NEXT: store ptr [[GLIN]], ptr [[GLIN_VAR]]172 173 #pragma omp simd174// CHECK: store i32 0, ptr [[OMP_IV4:%[^,]+]]175 176// CHECK: [[IV4:%.+]] = load i32, ptr [[OMP_IV4]]{{.*}}!llvm.access.group177// CHECK-NEXT: [[CMP4:%.+]] = icmp slt i32 [[IV4]], 4178// CHECK-NEXT: br i1 [[CMP4]], label %[[SIMPLE_LOOP4_BODY:.+]], label %[[SIMPLE_LOOP4_END:[^,]+]]179 for (short it = 6; it <= 20; it-=-4) {180// CHECK: [[SIMPLE_LOOP4_BODY]]:181// Start of body: calculate it from IV:182// CHECK: [[IV4_0:%.+]] = load i32, ptr [[OMP_IV4]]{{.*}}!llvm.access.group183// CHECK-NEXT: [[LC_IT_1:%.+]] = mul nsw i32 [[IV4_0]], 4184// CHECK-NEXT: [[LC_IT_2:%.+]] = add nsw i32 6, [[LC_IT_1]]185// CHECK-NEXT: [[LC_IT_3:%.+]] = trunc i32 [[LC_IT_2]] to i16186// CHECK-NEXT: store i16 [[LC_IT_3]], ptr {{.+}}, !llvm.access.group187 188// CHECK: [[IV4_2:%.+]] = load i32, ptr [[OMP_IV4]]{{.*}}!llvm.access.group189// CHECK-NEXT: [[ADD4_2:%.+]] = add nsw i32 [[IV4_2]], 1190// CHECK-NEXT: store i32 [[ADD4_2]], ptr [[OMP_IV4]]{{.*}}!llvm.access.group191 }192// CHECK: [[SIMPLE_LOOP4_END]]:193 194 #pragma omp simd195// CHECK: store i32 0, ptr [[OMP_IV5:%[^,]+]]196 197// CHECK: [[IV5:%.+]] = load i32, ptr [[OMP_IV5]]{{.*}}!llvm.access.group198// CHECK-NEXT: [[CMP5:%.+]] = icmp slt i32 [[IV5]], 26199// CHECK-NEXT: br i1 [[CMP5]], label %[[SIMPLE_LOOP5_BODY:.+]], label %[[SIMPLE_LOOP5_END:[^,]+]]200 for (unsigned char it = 'z'; it >= 'a'; it+=-1) {201// CHECK: [[SIMPLE_LOOP5_BODY]]:202// Start of body: calculate it from IV:203// CHECK: [[IV5_0:%.+]] = load i32, ptr [[OMP_IV5]]{{.*}}!llvm.access.group204// CHECK-NEXT: [[IV5_1:%.+]] = mul nsw i32 [[IV5_0]], 1205// CHECK-NEXT: [[LC_IT_1:%.+]] = sub nsw i32 122, [[IV5_1]]206// CHECK-NEXT: [[LC_IT_2:%.+]] = trunc i32 [[LC_IT_1]] to i8207// CHECK-NEXT: store i8 [[LC_IT_2]], ptr {{.+}}, !llvm.access.group208 209// CHECK: [[IV5_2:%.+]] = load i32, ptr [[OMP_IV5]]{{.*}}!llvm.access.group210// CHECK-NEXT: [[ADD5_2:%.+]] = add nsw i32 [[IV5_2]], 1211// CHECK-NEXT: store i32 [[ADD5_2]], ptr [[OMP_IV5]]{{.*}}!llvm.access.group212 }213// CHECK: [[SIMPLE_LOOP5_END]]:214 215// CHECK-NOT: mul i32 %{{.+}}, 10216 #pragma omp simd217 for (unsigned i=100; i<10; i+=10) {218 }219 220 int A;221 // CHECK: store i32 -1, ptr [[A:%.+]],222 A = -1;223 #pragma omp simd lastprivate(CONDITIONAL A)224// CHECK: store i64 0, ptr [[OMP_IV7:%[^,]+]]225// CHECK: br label %[[SIMD_LOOP7_COND:[^,]+]]226// CHECK: [[SIMD_LOOP7_COND]]:227// CHECK-NEXT: [[IV7:%.+]] = load i64, ptr [[OMP_IV7]]{{.*}}!llvm.access.group228// CHECK-NEXT: [[CMP7:%.+]] = icmp slt i64 [[IV7]], 7229// CHECK-NEXT: br i1 [[CMP7]], label %[[SIMPLE_LOOP7_BODY:.+]], label %[[SIMPLE_LOOP7_END:[^,]+]]230 for (long long i = -10; i < 10; i += 3) {231// CHECK: [[SIMPLE_LOOP7_BODY]]:232// Start of body: calculate i from IV:233// CHECK: [[IV7_0:%.+]] = load i64, ptr [[OMP_IV7]]{{.*}}!llvm.access.group234// CHECK-NEXT: [[LC_IT_1:%.+]] = mul nsw i64 [[IV7_0]], 3235// CHECK-NEXT: [[LC_IT_2:%.+]] = add nsw i64 -10, [[LC_IT_1]]236// CHECK-NEXT: store i64 [[LC_IT_2]], ptr [[LC:%[^,]+]],{{.+}}!llvm.access.group237// CHECK-NEXT: [[LC_VAL:%.+]] = load i64, ptr [[LC]]{{.+}}!llvm.access.group238// CHECK-NEXT: [[A_VAL:%.+]] = load i32, ptr [[A_PRIV:%[^,]+]],{{.+}}!llvm.access.group239// CHECK-NEXT: [[CAST:%.+]] = sext i32 [[A_VAL]] to i64240// CHECK-NEXT: [[ADD:%.+]] = add nsw i64 [[CAST]], [[LC_VAL]]241// CHECK-NEXT: [[CONV:%.+]] = trunc i64 [[ADD]] to i32242// CHECK-NEXT: store i32 [[CONV]], ptr [[A_PRIV]],{{.+}}!llvm.access.group243// OMP5-NEXT: [[IV:%.+]] = load i64, ptr [[OMP_IV7]],{{.+}}!llvm.access.group244// OMP5RT: call void @__kmpc_critical(ptr {{.+}}, i32 [[GTID:%.+]], ptr [[A_REGION:@.+]]),{{.+}}!llvm.access.group245// OMP5-NEXT: [[LAST_IV_VAL:%.+]] = load i64, ptr [[LAST_IV]],{{.+}}!llvm.access.group246// OMP5-NEXT: [[CMP:%.+]] = icmp sle i64 [[LAST_IV_VAL]], [[IV]]247// OMP5-NEXT: br i1 [[CMP]], label %[[LP_THEN:.+]], label %[[LP_DONE:[^,]+]]248// OMP5: [[LP_THEN]]:249// OMP5-NEXT: store i64 [[IV]], ptr [[LAST_IV]],{{.+}}!llvm.access.group250// OMP5-NEXT: [[A_VAL:%.+]] = load i32, ptr [[A_PRIV]],{{.+}}!llvm.access.group251// OMP5-NEXT: store i32 [[A_VAL]], ptr [[LAST_A]],{{.+}}!llvm.access.group252// OMP5-NEXT: br label %[[LP_DONE]]253// OMP5: [[LP_DONE]]:254// OMP5RT-NEXT: call void @__kmpc_end_critical(ptr {{.+}}, i32 [[GTID]], ptr [[A_REGION]]),{{.+}}!llvm.access.group255 A += i;256// CHECK: [[IV7_2:%.+]] = load i64, ptr [[OMP_IV7]]{{.*}}!llvm.access.group257// CHECK-NEXT: [[ADD7_2:%.+]] = add nsw i64 [[IV7_2]], 1258// CHECK-NEXT: store i64 [[ADD7_2]], ptr [[OMP_IV7]]{{.*}}!llvm.access.group259 }260// CHECK: [[SIMPLE_LOOP7_END]]:261// CHECK-NEXT: store i64 11, ptr262// OMP5-NEXT: [[LAST_A_VAL:%.+]] = load i32, ptr [[LAST_A]],263// OMP5-NEXT: store i32 [[LAST_A_VAL]], ptr [[A_PRIV]],264// CHECK-NEXT: [[A_PRIV_VAL:%.+]] = load i32, ptr [[A_PRIV]],265// CHECK-NEXT: store i32 [[A_PRIV_VAL]], ptr [[A]],266 int R;267 // CHECK: store i32 -1, ptr [[R:%[^,]+]],268 R = -1;269// CHECK: store i64 0, ptr [[OMP_IV8:%[^,]+]],270// CHECK: store i32 1, ptr [[R_PRIV:%[^,]+]],271#ifdef OMP5272 #pragma omp simd reduction(*:R) if(A)273#else274 #pragma omp simd reduction(*:R)275#endif276// OMP5: [[A_VAL:%.+]] = load i32, ptr [[A]],277// OMP5-NEXT: [[COND:%.+]] = icmp ne i32 [[A_VAL]], 0278// OMP5-NEXT: br i1 [[COND]], label {{%?}}[[THEN:[^,]+]], label {{%?}}[[ELSE:[^,]+]]279// OMP5: [[THEN]]:280 281// CHECK: br label %[[SIMD_LOOP8_COND:[^,]+]]282// CHECK: [[SIMD_LOOP8_COND]]:283// CHECK-NEXT: [[IV8:%.+]] = load i64, ptr [[OMP_IV8]]{{.*}}!llvm.access.group284// CHECK-NEXT: [[CMP8:%.+]] = icmp slt i64 [[IV8]], 7285// CHECK-NEXT: br i1 [[CMP8]], label %[[SIMPLE_LOOP8_BODY:.+]], label %[[SIMPLE_LOOP8_END:[^,]+]]286 for (long long i = -10; i < 10; i += 3) {287// CHECK: [[SIMPLE_LOOP8_BODY]]:288// Start of body: calculate i from IV:289// CHECK: [[IV8_0:%.+]] = load i64, ptr [[OMP_IV8]]{{.*}}!llvm.access.group290// CHECK-NEXT: [[LC_IT_1:%.+]] = mul nsw i64 [[IV8_0]], 3291// CHECK-NEXT: [[LC_IT_2:%.+]] = add nsw i64 -10, [[LC_IT_1]]292// CHECK-NEXT: store i64 [[LC_IT_2]], ptr [[LC:%[^,]+]],{{.+}}!llvm.access.group293// CHECK-NEXT: [[LC_VAL:%.+]] = load i64, ptr [[LC]]{{.+}}!llvm.access.group294// CHECK: store i32 %{{.+}}, ptr [[R_PRIV]],{{.+}}!llvm.access.group295 R *= i;296// CHECK: [[IV8_2:%.+]] = load i64, ptr [[OMP_IV8]]{{.*}}!llvm.access.group297// CHECK-NEXT: [[ADD8_2:%.+]] = add nsw i64 [[IV8_2]], 1298// CHECK-NEXT: store i64 [[ADD8_2]], ptr [[OMP_IV8]]{{.*}}!llvm.access.group299 }300// CHECK: [[SIMPLE_LOOP8_END]]:301// OMP5: br label {{%?}}[[EXIT:[^,]+]]302// OMP5: br label %[[SIMD_LOOP8_COND:[^,]+]]303// OMP5: [[SIMD_LOOP8_COND]]:304// OMP5-NEXT: [[IV8:%.+]] = load i64, ptr [[OMP_IV8]],{{[^!]*}}305// OMP5-NEXT: [[CMP8:%.+]] = icmp slt i64 [[IV8]], 7306// OMP5-NEXT: br i1 [[CMP8]], label %[[SIMPLE_LOOP8_BODY:.+]], label %[[SIMPLE_LOOP8_END:[^,]+]]307// OMP5: [[SIMPLE_LOOP8_BODY]]:308// Start of body: calculate i from IV:309// OMP5: [[IV8_0:%.+]] = load i64, ptr [[OMP_IV8]],{{[^!]*}}310// OMP5-NEXT: [[LC_IT_1:%.+]] = mul nsw i64 [[IV8_0]], 3311// OMP5-NEXT: [[LC_IT_2:%.+]] = add nsw i64 -10, [[LC_IT_1]]312// OMP5-NEXT: store i64 [[LC_IT_2]], ptr [[LC:%[^,]+]],{{[^!]*}}313// OMP5-NEXT: [[LC_VAL:%.+]] = load i64, ptr [[LC]],{{[^!]*}}314// OMP5: store i32 %{{.+}}, ptr [[R_PRIV]],{{[^!]*}}315// OMP5: [[IV8_2:%.+]] = load i64, ptr [[OMP_IV8]],{{[^!]*}}316// OMP5-NEXT: [[ADD8_2:%.+]] = add nsw i64 [[IV8_2]], 1317// OMP5-NEXT: store i64 [[ADD8_2]], ptr [[OMP_IV8]],{{[^!]*}}318// OMP5: br label {{%?}}[[SIMD_LOOP8_COND]], {{.*}}!llvm.loop ![[DISABLE_VECT:.+]]319// OMP5: [[SIMPLE_LOOP8_END]]:320// OMP5: br label {{%?}}[[EXIT]]321// OMP5: [[EXIT]]:322 323// CHECK-DAG: [[R_VAL:%.+]] = load i32, ptr [[R]],324// CHECK-DAG: [[R_PRIV_VAL:%.+]] = load i32, ptr [[R_PRIV]],325// CHECK: [[RED:%.+]] = mul nsw i32 [[R_VAL]], [[R_PRIV_VAL]]326// CHECK-NEXT: store i32 [[RED]], ptr [[R]],327// CHECK-NEXT: ret void328}329 330template <class T, unsigned K> T tfoo(T a) { return a + K; }331 332template <typename T, unsigned N>333int templ1(T a, T *z) {334 #pragma omp simd collapse(N)335 for (int i = 0; i < N * 2; i++) {336 for (long long j = 0; j < (N + N + N + N); j += 2) {337 z[i + j] = a + tfoo<T, N>(i + j);338 }339 }340 return 0;341}342 343// Instatiation templ1<float,2>344// CHECK-LABEL: define {{.*i32}} @{{.*}}templ1{{.*}}(float noundef {{.+}}, ptr noundef {{.+}})345// CHECK: store i64 0, ptr [[T1_OMP_IV:[^,]+]]346// ...347// CHECK: [[IV:%.+]] = load i64, ptr [[T1_OMP_IV]]{{.*}}!llvm.access.group348// CHECK-NEXT: [[CMP1:%.+]] = icmp slt i64 [[IV]], 16349// CHECK-NEXT: br i1 [[CMP1]], label %[[T1_BODY:.+]], label %[[T1_END:[^,]+]]350// CHECK: [[T1_BODY]]:351// Loop counters i and j updates:352// CHECK: [[IV1:%.+]] = load i64, ptr [[T1_OMP_IV]]{{.*}}!llvm.access.group353// CHECK-NEXT: [[I_1:%.+]] = sdiv i64 [[IV1]], 4354// CHECK-NEXT: [[I_1_MUL1:%.+]] = mul nsw i64 [[I_1]], 1355// CHECK-NEXT: [[I_1_ADD0:%.+]] = add nsw i64 0, [[I_1_MUL1]]356// CHECK-NEXT: [[I_2:%.+]] = trunc i64 [[I_1_ADD0]] to i32357// CHECK-NEXT: store i32 [[I_2]], ptr {{%.+}}{{.*}}!llvm.access.group358// CHECK: [[IV2:%.+]] = load i64, ptr [[T1_OMP_IV]]{{.*}}!llvm.access.group359// CHECK: [[IV2_1:%.+]] = load i64, ptr [[T1_OMP_IV]]{{.*}}!llvm.access.group360// CHECK-NEXT: [[J_1_DIV1:%.+]] = sdiv i64 [[IV2_1]], 4361// CHECK-NEXT: [[J_1_MUL1:%.+]] = mul nsw i64 [[J_1_DIV1]], 4362// CHECK-NEXT: [[J_1_SUB0:%.+]] = sub nsw i64 [[IV2]], [[J_1_MUL1]]363// CHECK-NEXT: [[J_2:%.+]] = mul nsw i64 [[J_1_SUB0]], 2364// CHECK-NEXT: [[J_2_ADD0:%.+]] = add nsw i64 0, [[J_2]]365// CHECK-NEXT: store i64 [[J_2_ADD0]], ptr {{%.+}}{{.*}}!llvm.access.group366// simd.for.inc:367// CHECK: [[IV3:%.+]] = load i64, ptr [[T1_OMP_IV]]{{.*}}!llvm.access.group368// CHECK-NEXT: [[INC:%.+]] = add nsw i64 [[IV3]], 1369// CHECK-NEXT: store i64 [[INC]], ptr [[T1_OMP_IV]]{{.*}}!llvm.access.group370// CHECK-NEXT: br label {{%.+}}371// CHECK: [[T1_END]]:372// CHECK: ret i32 0373//374void inst_templ1() {375 float a;376 float z[100];377 templ1<float,2> (a, z);378}379 380 381typedef int MyIdx;382 383class IterDouble {384 double *Ptr;385public:386 IterDouble operator++ () const {387 IterDouble n;388 n.Ptr = Ptr + 1;389 return n;390 }391 bool operator < (const IterDouble &that) const {392 return Ptr < that.Ptr;393 }394 double & operator *() const {395 return *Ptr;396 }397 MyIdx operator - (const IterDouble &that) const {398 return (MyIdx) (Ptr - that.Ptr);399 }400 IterDouble operator + (int Delta) {401 IterDouble re;402 re.Ptr = Ptr + Delta;403 return re;404 }405 406 ///~IterDouble() {}407};408 409// CHECK-LABEL: define {{.*void}} @{{.*}}iter_simple{{.*}}410void iter_simple(IterDouble ia, IterDouble ib, IterDouble ic) {411//412// Calculate number of iterations before the loop body.413// CHECK: [[DIFF1:%.+]] = invoke {{.*}}i32 @{{.*}}IterDouble{{.*}}414// CHECK: [[DIFF2:%.+]] = sub nsw i32 [[DIFF1]], 1415// CHECK-NEXT: [[DIFF3:%.+]] = add nsw i32 [[DIFF2]], 1416// CHECK-NEXT: [[DIFF4:%.+]] = sdiv i32 [[DIFF3]], 1417// CHECK-NEXT: [[DIFF5:%.+]] = sub nsw i32 [[DIFF4]], 1418// CHECK-NEXT: store i32 [[DIFF5]], ptr [[OMP_LAST_IT:%[^,]+]]{{.+}}419// CHECK: store i32 0, ptr [[IT_OMP_IV:%[^,]+]]420 #pragma omp simd421 422// CHECK: [[IV:%.+]] = load i32, ptr [[IT_OMP_IV]]{{.+}} !llvm.access.group423// CHECK-NEXT: [[LAST_IT:%.+]] = load i32, ptr [[OMP_LAST_IT]]{{.+}}!llvm.access.group424// CHECK-NEXT: [[NUM_IT:%.+]] = add nsw i32 [[LAST_IT]], 1425// CHECK-NEXT: [[CMP:%.+]] = icmp slt i32 [[IV]], [[NUM_IT]]426// CHECK-NEXT: br i1 [[CMP]], label %[[IT_BODY:[^,]+]], label %[[IT_END:[^,]+]]427 for (IterDouble i = ia; i < ib; ++i) {428// CHECK: [[IT_BODY]]:429// Start of body: calculate i from index:430// CHECK: [[IV1:%.+]] = load i32, ptr [[IT_OMP_IV]]{{.+}}!llvm.access.group431// Call of operator+ (i, IV).432// CHECK: {{%.+}} = invoke {{.+}} @{{.*}}IterDouble{{.*}}433// ... loop body ...434 *i = *ic * 0.5;435// Float multiply and save result.436// CHECK: [[MULR:%.+]] = fmul double {{%.+}}, 5.000000e-01437// CHECK-NEXT: invoke {{.+}} @{{.*}}IterDouble{{.*}}438// CHECK: store double [[MULR:%.+]], ptr [[RESULT_ADDR:%.+]], !llvm.access.group439 ++ic;440//441// CHECK: [[IV2:%.+]] = load i32, ptr [[IT_OMP_IV]]{{.+}}!llvm.access.group442// CHECK-NEXT: [[ADD2:%.+]] = add nsw i32 [[IV2]], 1443// CHECK-NEXT: store i32 [[ADD2]], ptr [[IT_OMP_IV]]{{.+}}!llvm.access.group444// br label %{{.*}}, !llvm.loop ![[ITER_LOOP_ID]]445 }446// CHECK: [[IT_END]]:447// CHECK: ret void448}449 450 451// CHECK-LABEL: define {{.*void}} @{{.*}}collapsed{{.*}}452void collapsed(float *a, float *b, float *c, float *d) {453 int i; // outer loop counter454 unsigned j; // middle loop couter, leads to unsigned icmp in loop header.455 // k declared in the loop init below456 short l; // inner loop counter457// CHECK: store i32 0, ptr [[OMP_IV:[^,]+]]458//459 #pragma omp simd collapse(4)460 461// CHECK: [[IV:%.+]] = load i32, ptr [[OMP_IV]]{{.+}}!llvm.access.group462// CHECK-NEXT: [[CMP:%.+]] = icmp ult i32 [[IV]], 120463// CHECK-NEXT: br i1 [[CMP]], label %[[COLL1_BODY:[^,]+]], label %[[COLL1_END:[^,]+]]464 for (i = 1; i < 3; i++) // 2 iterations465 for (j = 2u; j < 5u; j++) //3 iterations466 for (int k = 3; k <= 6; k++) // 4 iterations467 for (l = 4; l < 9; ++l) // 5 iterations468 {469// CHECK: [[COLL1_BODY]]:470// Start of body: calculate i from index:471// CHECK: [[IV1:%.+]] = load i32, ptr [[OMP_IV]]{{.+}}!llvm.access.group472// Calculation of the loop counters values.473// CHECK: [[CALC_I_1:%.+]] = udiv i32 [[IV1]], 60474// CHECK-NEXT: [[CALC_I_1_MUL1:%.+]] = mul i32 [[CALC_I_1]], 1475// CHECK-NEXT: [[CALC_I_2:%.+]] = add i32 1, [[CALC_I_1_MUL1]]476// CHECK-NEXT: store i32 [[CALC_I_2]], ptr [[LC_I:.+]]477 478// CHECK: [[IV1_2:%.+]] = load i32, ptr [[OMP_IV]]{{.+}}!llvm.access.group479// CHECK: [[IV1_2_1:%.+]] = load i32, ptr [[OMP_IV]]{{.+}}!llvm.access.group480// CHECK-NEXT: [[CALC_J_1:%.+]] = udiv i32 [[IV1_2_1]], 60481// CHECK-NEXT: [[MUL_1:%.+]] = mul i32 [[CALC_J_1]], 60482// CHECK-NEXT: [[SUB_3:%.+]] = sub i32 [[IV1_2]], [[MUL_1]]483// CHECK-NEXT: [[CALC_J_2:%.+]] = udiv i32 [[SUB_3]], 20484// CHECK-NEXT: [[CALC_J_2_MUL1:%.+]] = mul i32 [[CALC_J_2]], 1485// CHECK-NEXT: [[CALC_J_3:%.+]] = add i32 2, [[CALC_J_2_MUL1]]486// CHECK-NEXT: store i32 [[CALC_J_3]], ptr [[LC_J:.+]]487 488// CHECK: [[IV1_3:%.+]] = load i32, ptr [[OMP_IV]]{{.+}}!llvm.access.group489// CHECK: [[IV1_3_1:%.+]] = load i32, ptr [[OMP_IV]]{{.+}}!llvm.access.group490// CHECK-NEXT: [[DIV_1:%.+]] = udiv i32 [[IV1_3_1]], 60491// CHECK-NEXT: [[MUL_2:%.+]] = mul i32 [[DIV_1]], 60492// CHECK-NEXT: [[ADD_3:%.+]] = sub i32 [[IV1_3]], [[MUL_2]]493 494// CHECK: [[IV1_4:%.+]] = load i32, ptr [[OMP_IV]]495// CHECK: [[IV1_4_1:%.+]] = load i32, ptr [[OMP_IV]]496// CHECK-NEXT: [[DIV_2:%.+]] = udiv i32 [[IV1_4_1]], 60497// CHECK-NEXT: [[MUL_3:%.+]] = mul i32 [[DIV_2]], 60498// CHECK-NEXT: [[SUB_6:%.+]] = sub i32 [[IV1_4]], [[MUL_3]]499// CHECK-NEXT: [[DIV_3:%.+]] = udiv i32 [[SUB_6]], 20500// CHECK-NEXT: [[MUL_4:%.+]] = mul i32 [[DIV_3]], 20501// CHECK-NEXT: [[ADD_5:%.+]] = sub i32 [[ADD_3]], [[MUL_4]]502// CHECK-NEXT: [[DIV_4:%.+]] = udiv i32 [[ADD_5]], 5503// CHECK-NEXT: [[MUL_5:%.+]] = mul i32 [[DIV_4]], 1504// CHECK-NEXT: [[ADD_6:%.+]] = add i32 3, [[MUL_5]]505// CHECK-NEXT: store i32 [[ADD_6]], ptr [[LC_K:.+]]506 507// CHECK: [[IV1_5:%.+]] = load i32, ptr [[OMP_IV]]{{.+}}!llvm.access.group508// CHECK: [[IV1_5_1:%.+]] = load i32, ptr [[OMP_IV]]{{.+}}!llvm.access.group509// CHECK-NEXT: [[DIV_5:%.+]] = udiv i32 [[IV1_5_1]], 60510// CHECK-NEXT: [[MUL_6:%.+]] = mul i32 [[DIV_5]], 60511// CHECK-NEXT: [[SUB_7:%.+]] = sub i32 [[IV1_5]], [[MUL_6]]512 513// CHECK: [[IV1_6:%.+]] = load i32, ptr [[OMP_IV]]514// CHECK: [[IV1_6_1:%.+]] = load i32, ptr [[OMP_IV]]515// CHECK-NEXT: [[DIV_6:%.+]] = udiv i32 [[IV1_6_1]], 60516// CHECK-NEXT: [[MUL_7:%.+]] = mul i32 [[DIV_6]], 60517// CHECK-NEXT: [[SUB_10:%.+]] = sub i32 [[IV1_6]], [[MUL_7]]518// CHECK-NEXT: [[DIV_7:%.+]] = udiv i32 [[SUB_10]], 20519// CHECK-NEXT: [[MUL_8:%.+]] = mul i32 [[DIV_7]], 20520// CHECK-NEXT: [[SUB_11:%.+]] = sub i32 [[SUB_7]], [[MUL_8]]521 522// CHECK: [[IV1_7:%.+]] = load i32, ptr [[OMP_IV]]523// CHECK: [[IV1_7_1:%.+]] = load i32, ptr [[OMP_IV]]524// CHECK-NEXT: [[DIV_8:%.+]] = udiv i32 [[IV1_7_1]], 60525// CHECK-NEXT: [[MUL_9:%.+]] = mul i32 [[DIV_8]], 60526// CHECK-NEXT: [[SUB_12:%.+]] = sub i32 [[IV1_7]], [[MUL_9]]527 528// CHECK: [[IV1_8:%.+]] = load i32, ptr [[OMP_IV]]529// CHECK: [[IV1_8_1:%.+]] = load i32, ptr [[OMP_IV]]530// CHECK-NEXT: [[DIV_3:%.+]] = udiv i32 [[IV1_8_1]], 60531// CHECK-NEXT: [[MUL_4:%.+]] = mul i32 [[DIV_3]], 60532// CHECK-NEXT: [[SUB_7:%.+]] = sub i32 [[IV1_8]], [[MUL_4]]533// CHECK-NEXT: [[DIV_4:%.+]] = udiv i32 [[SUB_7]], 20534// CHECK-NEXT: [[MUL_5:%.+]] = mul i32 [[DIV_4]], 20535// CHECK-NEXT: [[SUB_8:%.+]] = sub i32 [[SUB_12]], [[MUL_5]]536// CHECK-NEXT: [[DIV_5:%.+]] = udiv i32 [[SUB_8]], 5537// CHECK-NEXT: [[MUL_6:%.+]] = mul i32 [[DIV_5]], 5538// CHECK-NEXT: [[SUB_9:%.+]] = sub i32 [[SUB_11]], [[MUL_6]]539// CHECK-NEXT: [[MUL_6:%.+]] = mul i32 [[SUB_9]], 1540// CHECK-NEXT: [[CALC_L_2:%.+]] = add i32 4, [[MUL_6]]541// CHECK-NEXT: [[CALC_L_3:%.+]] = trunc i32 [[CALC_L_2]] to i16542// CHECK-NEXT: store i16 [[CALC_L_3]], ptr [[LC_L:.+]]543// ... loop body ...544// End of body: store into a[i]:545// CHECK: store float [[RESULT:%.+]], ptr [[RESULT_ADDR:%.+]]{{.+}}!llvm.access.group546 float res = b[j] * c[k];547 a[i] = res * d[l];548// CHECK: [[IV2:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group549// CHECK-NEXT: [[ADD2:%.+]] = add i32 [[IV2]], 1550// CHECK-NEXT: store i32 [[ADD2]], ptr [[OMP_IV]]{{.*}}!llvm.access.group551// br label %{{[^,]+}}, !llvm.loop ![[COLL1_LOOP_ID]]552// CHECK: [[COLL1_END]]:553 }554// i,j,l are updated; k is not updated.555// CHECK: store i32 3, ptr556// CHECK-NEXT: store i32 5, ptr557// CHECK-NEXT: store i32 7, ptr558// CHECK-NEXT: store i16 9, ptr559// CHECK: ret void560}561 562extern char foo();563extern double globalfloat;564 565// CHECK-LABEL: define {{.*void}} @{{.*}}widened{{.*}}566void widened(float *a, float *b, float *c, float *d) {567 int i; // outer loop counter568 short j; // inner loop counter569 globalfloat = 1.0;570 int localint = 1;571// CHECK: store double {{.+}}, ptr [[GLOBALFLOAT:@.+]]572// Counter is widened to 64 bits.573// CHECK: store i64 0, ptr [[OMP_IV:[^,]+]]574//575 #pragma omp simd collapse(2) private(globalfloat, localint)576 577// CHECK: [[IV:%.+]] = load i64, ptr [[OMP_IV]]{{.+}}!llvm.access.group578// CHECK-NEXT: [[LI:%.+]] = load i64, ptr [[OMP_LI:%[^,]+]]{{.+}}!llvm.access.group579// CHECK-NEXT: [[NUMIT:%.+]] = add nsw i64 [[LI]], 1580// CHECK-NEXT: [[CMP:%.+]] = icmp slt i64 [[IV]], [[NUMIT]]581// CHECK-NEXT: br i1 [[CMP]], label %[[WIDE1_BODY:[^,]+]], label %[[WIDE1_END:[^,]+]]582 for (i = 1; i < 3; i++) // 2 iterations583 for (j = 0; j < foo(); j++) // foo() iterations584 {585// CHECK: [[WIDE1_BODY]]:586// Start of body: calculate i from index:587// CHECK: [[IV1:%.+]] = load i64, ptr [[OMP_IV]]{{.+}}!llvm.access.group588// Calculation of the loop counters values...589// CHECK: store i32 {{[^,]+}}, ptr [[LC_I:.+]]590// CHECK: [[IV1_2:%.+]] = load i64, ptr [[OMP_IV]]{{.+}}!llvm.access.group591// CHECK: store i16 {{[^,]+}}, ptr [[LC_J:.+]]592// ... loop body ...593//594// Here we expect store into private double var, not global595// CHECK-NOT: store double {{.+}}, ptr [[GLOBALFLOAT]]596 globalfloat = (float)j/i;597 float res = b[j] * c[j];598// Store into a[i]:599// CHECK: store float [[RESULT:%.+]], ptr [[RESULT_ADDR:%.+]]{{.+}}!llvm.access.group600 a[i] = res * d[i];601// Then there's a store into private var localint:602// CHECK: store i32 {{.+}}, ptr [[LOCALINT:%[^,]+]]{{.+}}!llvm.access.group603 localint = (int)j;604// CHECK: [[IV2:%.+]] = load i64, ptr [[OMP_IV]]{{.*}}!llvm.access.group605// CHECK-NEXT: [[ADD2:%.+]] = add nsw i64 [[IV2]], 1606// CHECK-NEXT: store i64 [[ADD2]], ptr [[OMP_IV]]{{.*}}!llvm.access.group607//608// br label %{{[^,]+}}, !llvm.loop ![[WIDE1_LOOP_ID]]609// CHECK: [[WIDE1_END]]:610 }611// i,j are updated.612// CHECK: store i32 3, ptr [[I:%[^,]+]]613// CHECK: store i16614//615// Here we expect store into original localint, not its privatized version.616// CHECK-NOT: store i32 {{.+}}, ptr [[LOCALINT]]617 localint = (int)j;618// CHECK: ret void619}620 621// CHECK-LABEL: define {{.*void}} @{{.*}}linear{{.*}}(ptr noundef {{.+}})622void linear(float *a) {623 // CHECK: [[VAL_ADDR:%.+]] = alloca i64,624 // CHECK: [[K_ADDR:%.+]] = alloca ptr,625 long long val = 0;626 long long &k = val;627 628 #pragma omp simd linear(k : 3)629// CHECK: store ptr [[VAL_ADDR]], ptr [[K_ADDR]],630// CHECK: [[VAL_REF:%.+]] = load ptr, ptr [[K_ADDR]],631// CHECK: store ptr [[VAL_REF]], ptr [[K_ADDR_REF:%.+]],632// CHECK: store i32 0, ptr [[OMP_IV:%[^,]+]]633// CHECK: [[K_REF:%.+]] = load ptr, ptr [[K_ADDR_REF]],634// CHECK: [[K0LOAD:%.+]] = load i64, ptr [[K_REF]]635// CHECK-NEXT: store i64 [[K0LOAD]], ptr [[LIN0:%[^,]+]]636 637// CHECK: [[IV:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group638// CHECK-NEXT: [[CMP2:%.+]] = icmp slt i32 [[IV]], 9639// CHECK-NEXT: br i1 [[CMP2]], label %[[SIMPLE_LOOP_BODY:.+]], label %[[SIMPLE_LOOP_END:[^,]+]]640 for (int i = 10; i > 1; i--) {641// CHECK: [[SIMPLE_LOOP_BODY]]:642// Start of body: calculate i from IV:643// CHECK: [[IV_0:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group644// FIXME: It is interesting, why the following "mul 1" was not constant folded?645// CHECK-NEXT: [[IV_1:%.+]] = mul nsw i32 [[IV_0]], 1646// CHECK-NEXT: [[LC_I_1:%.+]] = sub nsw i32 10, [[IV_1]]647// CHECK-NEXT: store i32 [[LC_I_1]], ptr {{.+}}, !llvm.access.group648//649// CHECK-NEXT: [[LIN0_1:%.+]] = load i64, ptr [[LIN0]]{{.*}}!llvm.access.group650// CHECK-NEXT: [[IV_2:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group651// CHECK-NEXT: [[LIN_MUL1:%.+]] = mul nsw i32 [[IV_2]], 3652// CHECK-NEXT: [[LIN_EXT1:%.+]] = sext i32 [[LIN_MUL1]] to i64653// CHECK-NEXT: [[LIN_ADD1:%.+]] = add nsw i64 [[LIN0_1]], [[LIN_EXT1]]654// Update of the privatized version of linear variable!655// CHECK-NEXT: store i64 [[LIN_ADD1]], ptr [[K_PRIVATIZED:%[^,]+]]656 a[k]++;657 k = k + 3;658// CHECK: [[IV_2:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group659// CHECK-NEXT: [[ADD2_2:%.+]] = add nsw i32 [[IV_2]], 1660// CHECK-NEXT: store i32 [[ADD2_2]], ptr [[OMP_IV]]{{.*}}!llvm.access.group661// br label {{.+}}, !llvm.loop ![[SIMPLE_LOOP_ID]]662 }663// CHECK: [[SIMPLE_LOOP_END]]:664//665// Update linear vars after loop, as the loop was operating on a private version.666// CHECK: [[K_REF:%.+]] = load ptr, ptr [[K_ADDR_REF]],667// CHECK: store ptr [[K_REF]], ptr [[K_PRIV_REF:%.+]],668// CHECK: [[LIN0_2:%.+]] = load i64, ptr [[K_PRIVATIZED]]669// CHECK-NEXT: [[K_REF:%.+]] = load ptr, ptr [[K_PRIV_REF]],670// CHECK-NEXT: store i64 [[LIN0_2]], ptr [[K_REF]]671//672 673 #pragma omp simd linear(val(k) : 3)674// CHECK: [[VAL_REF:%.+]] = load ptr, ptr [[K_ADDR]],675// CHECK: store ptr [[VAL_REF]], ptr [[K_ADDR_REF:%.+]],676// CHECK: store i32 0, ptr [[OMP_IV:%[^,]+]]677// CHECK: [[K_REF:%.+]] = load ptr, ptr [[K_ADDR_REF]],678// CHECK: [[K0LOAD:%.+]] = load i64, ptr [[K_REF]]679// CHECK-NEXT: store i64 [[K0LOAD]], ptr [[LIN0:%[^,]+]]680 681// CHECK: [[IV:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group682// CHECK-NEXT: [[CMP2:%.+]] = icmp slt i32 [[IV]], 9683// CHECK-NEXT: br i1 [[CMP2]], label %[[SIMPLE_LOOP_BODY:.+]], label %[[SIMPLE_LOOP_END:[^,]+]]684 for (int i = 10; i > 1; i--) {685// CHECK: [[SIMPLE_LOOP_BODY]]:686// Start of body: calculate i from IV:687// CHECK: [[IV_0:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group688// FIXME: It is interesting, why the following "mul 1" was not constant folded?689// CHECK-NEXT: [[IV_1:%.+]] = mul nsw i32 [[IV_0]], 1690// CHECK-NEXT: [[LC_I_1:%.+]] = sub nsw i32 10, [[IV_1]]691// CHECK-NEXT: store i32 [[LC_I_1]], ptr {{.+}}, !llvm.access.group692//693// CHECK-NEXT: [[LIN0_1:%.+]] = load i64, ptr [[LIN0]]{{.*}}!llvm.access.group694// CHECK-NEXT: [[IV_2:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group695// CHECK-NEXT: [[LIN_MUL1:%.+]] = mul nsw i32 [[IV_2]], 3696// CHECK-NEXT: [[LIN_EXT1:%.+]] = sext i32 [[LIN_MUL1]] to i64697// CHECK-NEXT: [[LIN_ADD1:%.+]] = add nsw i64 [[LIN0_1]], [[LIN_EXT1]]698// Update of the privatized version of linear variable!699// CHECK-NEXT: store i64 [[LIN_ADD1]], ptr [[K_PRIVATIZED:%[^,]+]]700 a[k]++;701 k = k + 3;702// CHECK: [[IV_2:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group703// CHECK-NEXT: [[ADD2_2:%.+]] = add nsw i32 [[IV_2]], 1704// CHECK-NEXT: store i32 [[ADD2_2]], ptr [[OMP_IV]]{{.*}}!llvm.access.group705// br label {{.+}}, !llvm.loop ![[SIMPLE_LOOP_ID]]706 }707// CHECK: [[SIMPLE_LOOP_END]]:708//709// Update linear vars after loop, as the loop was operating on a private version.710// CHECK: [[K_REF:%.+]] = load ptr, ptr [[K_ADDR_REF]],711// CHECK: store ptr [[K_REF]], ptr [[K_PRIV_REF:%.+]],712// CHECK: [[LIN0_2:%.+]] = load i64, ptr [[K_PRIVATIZED]]713// CHECK-NEXT: [[K_REF:%.+]] = load ptr, ptr [[K_PRIV_REF]],714// CHECK-NEXT: store i64 [[LIN0_2]], ptr [[K_REF]]715//716 #pragma omp simd linear(uval(k) : 3)717// CHECK: store i32 0, ptr [[OMP_IV:%[^,]+]]718// CHECK: [[K0LOAD:%.+]] = load i64, ptr [[VAL_ADDR]]719// CHECK-NEXT: store i64 [[K0LOAD]], ptr [[LIN0:%[^,]+]]720 721// CHECK: [[IV:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group722// CHECK-NEXT: [[CMP2:%.+]] = icmp slt i32 [[IV]], 9723// CHECK-NEXT: br i1 [[CMP2]], label %[[SIMPLE_LOOP_BODY:.+]], label %[[SIMPLE_LOOP_END:[^,]+]]724 for (int i = 10; i > 1; i--) {725// CHECK: [[SIMPLE_LOOP_BODY]]:726// Start of body: calculate i from IV:727// CHECK: [[IV_0:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group728// FIXME: It is interesting, why the following "mul 1" was not constant folded?729// CHECK-NEXT: [[IV_1:%.+]] = mul nsw i32 [[IV_0]], 1730// CHECK-NEXT: [[LC_I_1:%.+]] = sub nsw i32 10, [[IV_1]]731// CHECK-NEXT: store i32 [[LC_I_1]], ptr {{.+}}, !llvm.access.group732//733// CHECK-NEXT: [[LIN0_1:%.+]] = load i64, ptr [[LIN0]]{{.*}}!llvm.access.group734// CHECK-NEXT: [[IV_2:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group735// CHECK-NEXT: [[LIN_MUL1:%.+]] = mul nsw i32 [[IV_2]], 3736// CHECK-NEXT: [[LIN_EXT1:%.+]] = sext i32 [[LIN_MUL1]] to i64737// CHECK-NEXT: [[LIN_ADD1:%.+]] = add nsw i64 [[LIN0_1]], [[LIN_EXT1]]738// Update of the privatized version of linear variable!739// CHECK-NEXT: store i64 [[LIN_ADD1]], ptr [[K_PRIVATIZED:%[^,]+]]740 a[k]++;741 k = k + 3;742// CHECK: [[IV_2:%.+]] = load i32, ptr [[OMP_IV]]{{.*}}!llvm.access.group743// CHECK-NEXT: [[ADD2_2:%.+]] = add nsw i32 [[IV_2]], 1744// CHECK-NEXT: store i32 [[ADD2_2]], ptr [[OMP_IV]]{{.*}}!llvm.access.group745// br label {{.+}}, !llvm.loop ![[SIMPLE_LOOP_ID]]746 }747// CHECK: [[SIMPLE_LOOP_END]]:748//749// Update linear vars after loop, as the loop was operating on a private version.750// CHECK: [[LIN0_2:%.+]] = load i64, ptr [[K_PRIVATIZED]]751// CHECK-NEXT: store i64 [[LIN0_2]], ptr [[VAL_ADDR]]752//753}754 755#ifdef OMP5756// OMP5-LABEL: inner_simd757void inner_simd() {758 double a, b;759#pragma omp simd nontemporal(a)760 for (int i = 0; i < 10; ++i) {761#pragma omp simd nontemporal(b)762 for (int k = 0; k < 10; ++k) {763 // OMP5: load double,{{.*}}!nontemporal764 // OMP5: store double{{.*}}!nontemporal765 a = b;766 }767 // OMP5-NOT: load double,{{.*}}!nontemporal768 // OMP5: load double,769 // OMP5: store double{{.*}}!nontemporal770 a = b;771 }772}773 774extern struct T t;775struct Base {776 float a;777};778struct T : public Base {779 void foo() {780#pragma omp simd nontemporal(Base::a)781 for (int i = 0; i < 10; ++i) {782 // OMP5: store float{{.*}}!nontemporal783 // OMP5-NOT: nontemporal784 // OMP5-NEXT: store float785 Base::a = 0;786 t.a = 0;787 }788 }789} t;790 791void bartfoo() {792 t.foo();793}794 795#endif // OMP5796// TERM_DEBUG-LABEL: bar797int bar() { extern void mayThrow(); mayThrow(); return 0; };798 799// TERM_DEBUG-LABEL: parallel_simd800void parallel_simd(float *a) {801#pragma omp parallel802#pragma omp simd803 // TERM_DEBUG-NOT: __kmpc_global_thread_num804 // TERM_DEBUG: invoke noundef i32 {{.*}}bar{{.*}}()805 // TERM_DEBUG: unwind label %[[TERM_LPAD:[^,]+]],806 // TERM_DEBUG-NOT: __kmpc_global_thread_num807 // TERM_DEBUG: [[TERM_LPAD]]808 // TERM_DEBUG: call void @__clang_call_terminate809 // TERM_DEBUG: unreachable810 for (unsigned i = 131071; i <= 2147483647; i += 127)811 a[i] += bar();812}813// TERM_DEBUG: !{{[0-9]+}} = !DILocation(line: [[@LINE-11]],814 815// CHECK-LABEL: S8816// CHECK-DAG: call void @llvm.assume(i1817// CHECK-DAG: call void @llvm.assume(i1818// CHECK-DAG: call void @llvm.assume(i1819// CHECK-DAG: call void @llvm.assume(i1820struct SS {821 SS(): a(0) {}822 SS(int v) : a(v) {}823 int a;824 typedef int type;825};826 827template <typename T>828class S7 : public T {829protected:830 T *a;831 T b[2];832 S7() : a(0) {}833 834public:835 S7(typename T::type &v) : a((T*)&v) {836#pragma omp simd aligned(a)837 for (int k = 0; k < a->a; ++k)838 ++this->a->a;839#pragma omp simd aligned(this->b : 8)840 for (int k = 0; k < a->a; ++k)841 ++a->a;842 }843};844 845class S8 : private IterDouble, public S7<SS> {846 S8() {}847 848public:849 S8(int v) : S7<SS>(v){850#pragma omp parallel private(a)851#pragma omp simd aligned(S7<SS>::a)852 for (int k = 0; k < a->a; ++k)853 ++this->a->a;854#pragma omp parallel shared(b)855#pragma omp simd aligned(this->b: 4)856 for (int k = 0; k < a->a; ++k)857 ++a->a;858 }859};860S8 s8(0);861 862// TERM_DEBUG-NOT: line: 0,863// TERM_DEBUG: distinct !DISubprogram(linkageName: "_GLOBAL__sub_I_simd_codegen.cpp",864// OMP5-DAG: ![[NOVECT:.+]] = !{!"llvm.loop.vectorize.enable", i1 false}865// OMP5-DAG: ![[DISABLE_VECT]] = distinct !{{.*}}![[NOVECT]]{{[,}]}}866#endif // HEADER867 868