188 lines · cpp
1// Check code generation2// RUN: %clang_cc1 -verify -triple x86_64-pc-linux-gnu -fopenmp -emit-llvm %s -o - | FileCheck %s --check-prefix=IR3 4// Check same results after serialization round-trip5// RUN: %clang_cc1 -verify -triple x86_64-pc-linux-gnu -fopenmp -emit-pch -o %t %s6// RUN: %clang_cc1 -verify -triple x86_64-pc-linux-gnu -fopenmp -include-pch %t -emit-llvm %s -o - | FileCheck %s --check-prefix=IR7// expected-no-diagnostics8 9#ifndef HEADER10#define HEADER11 12// placeholder for loop body code.13extern "C" void body(...) {}14 15 16// IR-LABEL: @func(17// IR-NEXT: [[ENTRY:.*]]:18// IR-NEXT: %[[START_ADDR:.+]] = alloca i32, align 419// IR-NEXT: %[[END_ADDR:.+]] = alloca i32, align 420// IR-NEXT: %[[STEP_ADDR:.+]] = alloca i32, align 421// IR-NEXT: %[[I:.+]] = alloca i32, align 422// IR-NEXT: %[[DOTOMP_IV:.+]] = alloca i32, align 423// IR-NEXT: %[[TMP:.+]] = alloca i32, align 424// IR-NEXT: %[[DOTCAPTURE_EXPR_:.+]] = alloca i32, align 425// IR-NEXT: %[[DOTCAPTURE_EXPR_1:.+]] = alloca i32, align 426// IR-NEXT: %[[DOTCAPTURE_EXPR_2:.+]] = alloca i32, align 427// IR-NEXT: %[[DOTCAPTURE_EXPR_3:.+]] = alloca i32, align 428// IR-NEXT: %[[DOTCAPTURE_EXPR_6:.+]] = alloca i32, align 429// IR-NEXT: %[[DOTCAPTURE_EXPR_8:.+]] = alloca i32, align 430// IR-NEXT: %[[DOTUNROLLED_IV_I:.+]] = alloca i32, align 431// IR-NEXT: %[[DOTOMP_LB:.+]] = alloca i32, align 432// IR-NEXT: %[[DOTOMP_UB:.+]] = alloca i32, align 433// IR-NEXT: %[[DOTOMP_STRIDE:.+]] = alloca i32, align 434// IR-NEXT: %[[DOTOMP_IS_LAST:.+]] = alloca i32, align 435// IR-NEXT: %[[DOTUNROLLED_IV_I12:.+]] = alloca i32, align 436// IR-NEXT: %[[DOTUNROLL_INNER_IV_I:.+]] = alloca i32, align 437// IR-NEXT: %[[TMP0:.+]] = call i32 @__kmpc_global_thread_num(ptr @2)38// IR-NEXT: store i32 %[[START:.+]], ptr %[[START_ADDR]], align 439// IR-NEXT: store i32 %[[END:.+]], ptr %[[END_ADDR]], align 440// IR-NEXT: store i32 %[[STEP:.+]], ptr %[[STEP_ADDR]], align 441// IR-NEXT: %[[TMP1:.+]] = load i32, ptr %[[START_ADDR]], align 442// IR-NEXT: store i32 %[[TMP1]], ptr %[[DOTCAPTURE_EXPR_]], align 443// IR-NEXT: %[[TMP2:.+]] = load i32, ptr %[[END_ADDR]], align 444// IR-NEXT: store i32 %[[TMP2]], ptr %[[DOTCAPTURE_EXPR_1]], align 445// IR-NEXT: %[[TMP3:.+]] = load i32, ptr %[[STEP_ADDR]], align 446// IR-NEXT: store i32 %[[TMP3]], ptr %[[DOTCAPTURE_EXPR_2]], align 447// IR-NEXT: %[[TMP4:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_1]], align 448// IR-NEXT: %[[TMP5:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_]], align 449// IR-NEXT: %[[SUB:.+]] = sub i32 %[[TMP4]], %[[TMP5]]50// IR-NEXT: %[[SUB4:.+]] = sub i32 %[[SUB]], 151// IR-NEXT: %[[TMP6:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_2]], align 452// IR-NEXT: %[[ADD:.+]] = add i32 %[[SUB4]], %[[TMP6]]53// IR-NEXT: %[[TMP7:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_2]], align 454// IR-NEXT: %[[DIV:.+]] = udiv i32 %[[ADD]], %[[TMP7]]55// IR-NEXT: %[[SUB5:.+]] = sub i32 %[[DIV]], 156// IR-NEXT: store i32 %[[SUB5]], ptr %[[DOTCAPTURE_EXPR_3]], align 457// IR-NEXT: %[[TMP8:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_3]], align 458// IR-NEXT: %[[ADD7:.+]] = add i32 %[[TMP8]], 159// IR-NEXT: store i32 %[[ADD7]], ptr %[[DOTCAPTURE_EXPR_6]], align 460// IR-NEXT: %[[TMP9:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_6]], align 461// IR-NEXT: %[[SUB9:.+]] = sub i32 %[[TMP9]], -162// IR-NEXT: %[[DIV10:.+]] = udiv i32 %[[SUB9]], 263// IR-NEXT: %[[SUB11:.+]] = sub i32 %[[DIV10]], 164// IR-NEXT: store i32 %[[SUB11]], ptr %[[DOTCAPTURE_EXPR_8]], align 465// IR-NEXT: store i32 0, ptr %[[DOTUNROLLED_IV_I]], align 466// IR-NEXT: %[[TMP10:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_6]], align 467// IR-NEXT: %[[CMP:.+]] = icmp ult i32 0, %[[TMP10]]68// IR-NEXT: br i1 %[[CMP]], label %[[OMP_PRECOND_THEN:.+]], label %[[OMP_PRECOND_END:.+]]69// IR-EMPTY:70// IR-NEXT: [[OMP_PRECOND_THEN]]:71// IR-NEXT: store i32 0, ptr %[[DOTOMP_LB]], align 472// IR-NEXT: %[[TMP11:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_8]], align 473// IR-NEXT: store i32 %[[TMP11]], ptr %[[DOTOMP_UB]], align 474// IR-NEXT: store i32 1, ptr %[[DOTOMP_STRIDE]], align 475// IR-NEXT: store i32 0, ptr %[[DOTOMP_IS_LAST]], align 476// IR-NEXT: call void @__kmpc_for_static_init_4u(ptr @1, i32 %[[TMP0]], i32 34, ptr %[[DOTOMP_IS_LAST]], ptr %[[DOTOMP_LB]], ptr %[[DOTOMP_UB]], ptr %[[DOTOMP_STRIDE]], i32 1, i32 1)77// IR-NEXT: %[[TMP12:.+]] = load i32, ptr %[[DOTOMP_UB]], align 478// IR-NEXT: %[[TMP13:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_8]], align 479// IR-NEXT: %[[CMP13:.+]] = icmp ugt i32 %[[TMP12]], %[[TMP13]]80// IR-NEXT: br i1 %[[CMP13]], label %[[COND_TRUE:.+]], label %[[COND_FALSE:.+]]81// IR-EMPTY:82// IR-NEXT: [[COND_TRUE]]:83// IR-NEXT: %[[TMP14:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_8]], align 484// IR-NEXT: br label %[[COND_END:.+]]85// IR-EMPTY:86// IR-NEXT: [[COND_FALSE]]:87// IR-NEXT: %[[TMP15:.+]] = load i32, ptr %[[DOTOMP_UB]], align 488// IR-NEXT: br label %[[COND_END]]89// IR-EMPTY:90// IR-NEXT: [[COND_END]]:91// IR-NEXT: %[[COND:.+]] = phi i32 [ %[[TMP14]], %[[COND_TRUE]] ], [ %[[TMP15]], %[[COND_FALSE]] ]92// IR-NEXT: store i32 %[[COND]], ptr %[[DOTOMP_UB]], align 493// IR-NEXT: %[[TMP16:.+]] = load i32, ptr %[[DOTOMP_LB]], align 494// IR-NEXT: store i32 %[[TMP16]], ptr %[[DOTOMP_IV]], align 495// IR-NEXT: br label %[[OMP_INNER_FOR_COND:.+]]96// IR-EMPTY:97// IR-NEXT: [[OMP_INNER_FOR_COND]]:98// IR-NEXT: %[[TMP17:.+]] = load i32, ptr %[[DOTOMP_IV]], align 499// IR-NEXT: %[[TMP18:.+]] = load i32, ptr %[[DOTOMP_UB]], align 4100// IR-NEXT: %[[ADD14:.+]] = add i32 %[[TMP18]], 1101// IR-NEXT: %[[CMP15:.+]] = icmp ult i32 %[[TMP17]], %[[ADD14]]102// IR-NEXT: br i1 %[[CMP15]], label %[[OMP_INNER_FOR_BODY:.+]], label %[[OMP_INNER_FOR_END:.+]]103// IR-EMPTY:104// IR-NEXT: [[OMP_INNER_FOR_BODY]]:105// IR-NEXT: %[[TMP19:.+]] = load i32, ptr %[[DOTOMP_IV]], align 4106// IR-NEXT: %[[MUL:.+]] = mul i32 %[[TMP19]], 2107// IR-NEXT: %[[ADD16:.+]] = add i32 0, %[[MUL]]108// IR-NEXT: store i32 %[[ADD16]], ptr %[[DOTUNROLLED_IV_I12]], align 4109// IR-NEXT: %[[TMP20:.+]] = load i32, ptr %[[DOTUNROLLED_IV_I12]], align 4110// IR-NEXT: store i32 %[[TMP20]], ptr %[[DOTUNROLL_INNER_IV_I]], align 4111// IR-NEXT: br label %[[FOR_COND:.+]]112// IR-EMPTY:113// IR-NEXT: [[FOR_COND]]:114// IR-NEXT: %[[TMP21:.+]] = load i32, ptr %[[DOTUNROLL_INNER_IV_I]], align 4115// IR-NEXT: %[[TMP22:.+]] = load i32, ptr %[[DOTUNROLLED_IV_I12]], align 4116// IR-NEXT: %[[ADD17:.+]] = add i32 %[[TMP22]], 2117// IR-NEXT: %[[CMP18:.+]] = icmp ult i32 %[[TMP21]], %[[ADD17]]118// IR-NEXT: br i1 %[[CMP18]], label %[[LAND_RHS:.+]], label %[[LAND_END:.+]]119// IR-EMPTY:120// IR-NEXT: [[LAND_RHS]]:121// IR-NEXT: %[[TMP23:.+]] = load i32, ptr %[[DOTUNROLL_INNER_IV_I]], align 4122// IR-NEXT: %[[TMP24:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_3]], align 4123// IR-NEXT: %[[ADD19:.+]] = add i32 %[[TMP24]], 1124// IR-NEXT: %[[CMP20:.+]] = icmp ult i32 %[[TMP23]], %[[ADD19]]125// IR-NEXT: br label %[[LAND_END]]126// IR-EMPTY:127// IR-NEXT: [[LAND_END]]:128// IR-NEXT: %[[TMP25:.+]] = phi i1 [ false, %[[FOR_COND]] ], [ %[[CMP20]], %[[LAND_RHS]] ]129// IR-NEXT: br i1 %[[TMP25]], label %[[FOR_BODY:.+]], label %[[FOR_END:.+]]130// IR-EMPTY:131// IR-NEXT: [[FOR_BODY]]:132// IR-NEXT: %[[TMP26:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_]], align 4133// IR-NEXT: %[[TMP27:.+]] = load i32, ptr %[[DOTUNROLL_INNER_IV_I]], align 4134// IR-NEXT: %[[TMP28:.+]] = load i32, ptr %[[DOTCAPTURE_EXPR_2]], align 4135// IR-NEXT: %[[MUL21:.+]] = mul i32 %[[TMP27]], %[[TMP28]]136// IR-NEXT: %[[ADD22:.+]] = add i32 %[[TMP26]], %[[MUL21]]137// IR-NEXT: store i32 %[[ADD22]], ptr %[[I]], align 4138// IR-NEXT: %[[TMP29:.+]] = load i32, ptr %[[START_ADDR]], align 4139// IR-NEXT: %[[TMP30:.+]] = load i32, ptr %[[END_ADDR]], align 4140// IR-NEXT: %[[TMP31:.+]] = load i32, ptr %[[STEP_ADDR]], align 4141// IR-NEXT: %[[TMP32:.+]] = load i32, ptr %[[I]], align 4142// IR-NEXT: call void (...) @body(i32 noundef %[[TMP29]], i32 noundef %[[TMP30]], i32 noundef %[[TMP31]], i32 noundef %[[TMP32]])143// IR-NEXT: br label %[[FOR_INC:.+]]144// IR-EMPTY:145// IR-NEXT: [[FOR_INC]]:146// IR-NEXT: %[[TMP33:.+]] = load i32, ptr %[[DOTUNROLL_INNER_IV_I]], align 4147// IR-NEXT: %[[INC:.+]] = add i32 %[[TMP33]], 1148// IR-NEXT: store i32 %[[INC]], ptr %[[DOTUNROLL_INNER_IV_I]], align 4149// IR-NEXT: br label %[[FOR_COND]], !llvm.loop ![[LOOP2:[0-9]+]]150// IR-EMPTY:151// IR-NEXT: [[FOR_END]]:152// IR-NEXT: br label %[[OMP_BODY_CONTINUE:.+]]153// IR-EMPTY:154// IR-NEXT: [[OMP_BODY_CONTINUE]]:155// IR-NEXT: br label %[[OMP_INNER_FOR_INC:.+]]156// IR-EMPTY:157// IR-NEXT: [[OMP_INNER_FOR_INC]]:158// IR-NEXT: %[[TMP34:.+]] = load i32, ptr %[[DOTOMP_IV]], align 4159// IR-NEXT: %[[ADD23:.+]] = add i32 %[[TMP34]], 1160// IR-NEXT: store i32 %[[ADD23]], ptr %[[DOTOMP_IV]], align 4161// IR-NEXT: br label %[[OMP_INNER_FOR_COND]]162// IR-EMPTY:163// IR-NEXT: [[OMP_INNER_FOR_END]]:164// IR-NEXT: br label %[[OMP_LOOP_EXIT:.+]]165// IR-EMPTY:166// IR-NEXT: [[OMP_LOOP_EXIT]]:167// IR-NEXT: call void @__kmpc_for_static_fini(ptr @1, i32 %[[TMP0]])168// IR-NEXT: br label %[[OMP_PRECOND_END]]169// IR-EMPTY:170// IR-NEXT: [[OMP_PRECOND_END]]:171// IR-NEXT: call void @__kmpc_barrier(ptr @3, i32 %[[TMP0]])172// IR-NEXT: ret void173// IR-NEXT: }174extern "C" void func(int start, int end, int step) {175 int i;176 #pragma omp for177 #pragma omp unroll partial178 for (i = start; i < end; i+=step)179 body(start, end, step, i);180}181 182#endif /* HEADER */183 184 185// IR: ![[LOOP2]] = distinct !{![[LOOP2]], ![[LOOPPROP3:[0-9]+]], ![[LOOPPROP4:[0-9]+]]}186// IR: ![[LOOPPROP3]] = !{!"llvm.loop.mustprogress"}187// IR: ![[LOOPPROP4]] = !{!"llvm.loop.unroll.count", i32 2}188