80 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 52 3; RUN: opt < %s -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -S | FileCheck %s4 5define void @select_with_fastmath_flags(ptr noalias %a, ptr noalias %b, ptr noalias %c, i64 %N) {6; CHECK-LABEL: define void @select_with_fastmath_flags(7; CHECK-SAME: ptr noalias [[A:%.*]], ptr noalias [[B:%.*]], ptr noalias [[C:%.*]], i64 [[N:%.*]]) {8; CHECK-NEXT: [[ENTRY:.*]]:9; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 410; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]11; CHECK: [[VECTOR_PH]]:12; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], 413; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]14; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]15; CHECK: [[VECTOR_BODY]]:16; CHECK-NEXT: [[TMP0:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]17; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw float, ptr [[B]], i64 [[TMP0]]18; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP1]], align 419; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds nuw float, ptr [[C]], i64 [[TMP0]]20; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <4 x float>, ptr [[TMP3]], align 421; CHECK-NEXT: [[TMP5:%.*]] = fcmp fast ogt <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD1]]22; CHECK-NEXT: [[TMP6:%.*]] = fadd fast <4 x float> [[WIDE_LOAD]], splat (float 1.000000e+01)23; CHECK-NEXT: [[TMP7:%.*]] = select fast <4 x i1> [[TMP5]], <4 x float> [[TMP6]], <4 x float> [[WIDE_LOAD1]]24; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[TMP0]]25; CHECK-NEXT: store <4 x float> [[TMP7]], ptr [[TMP8]], align 426; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[TMP0]], 427; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]28; CHECK-NEXT: br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]29; CHECK: [[MIDDLE_BLOCK]]:30; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]31; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]32; CHECK: [[SCALAR_PH]]:33; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]34; CHECK-NEXT: br label %[[FOR_BODY:.*]]35; CHECK: [[FOR_BODY]]:36; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], %[[FOR_BODY]] ], [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ]37; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds nuw float, ptr [[B]], i64 [[IV]]38; CHECK-NEXT: [[TMP11:%.*]] = load float, ptr [[GEP]], align 439; CHECK-NEXT: [[GEP3:%.*]] = getelementptr inbounds nuw float, ptr [[C]], i64 [[IV]]40; CHECK-NEXT: [[TMP12:%.*]] = load float, ptr [[GEP3]], align 441; CHECK-NEXT: [[CMP4:%.*]] = fcmp fast ogt float [[TMP11]], [[TMP12]]42; CHECK-NEXT: [[ADD:%.*]] = fadd fast float [[TMP11]], 1.000000e+0143; CHECK-NEXT: [[COND:%.*]] = select fast i1 [[CMP4]], float [[ADD]], float [[TMP12]]44; CHECK-NEXT: [[GEP11:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV]]45; CHECK-NEXT: store float [[COND]], ptr [[GEP11]], align 446; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 147; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]48; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label %[[EXIT]], label %[[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]49; CHECK: [[EXIT]]:50; CHECK-NEXT: ret void51;52entry:53 br label %for.body54 55for.body:56 %iv = phi i64 [ %iv.next, %for.body ], [ 0, %entry ]57 %gep = getelementptr inbounds nuw float, ptr %b, i64 %iv58 %0 = load float, ptr %gep, align 459 %gep3 = getelementptr inbounds nuw float, ptr %c, i64 %iv60 %1 = load float, ptr %gep3, align 461 %cmp4 = fcmp fast ogt float %0, %162 %add = fadd fast float %0, 1.000000e+0163 %cond = select fast i1 %cmp4, float %add, float %164 %gep11 = getelementptr inbounds nuw float, ptr %a, i64 %iv65 store float %cond, ptr %gep11, align 466 %iv.next = add nuw nsw i64 %iv, 167 %exitcond.not = icmp eq i64 %iv.next, %N68 br i1 %exitcond.not, label %exit, label %for.body69 70exit:71 ret void72}73 74;.75; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}76; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}77; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}78; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}79;.80