108 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 62; RUN: opt < %s -O3 -S | FileCheck %s3 4target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"5target triple = "x86_64-apple-macosx11.0.0"6 7; We expect the loop to be vectorized even if it8; is not unrolled.9; Verify that fdiv is not sunk back into the loop10; after getting hoisted - based on this C source:11;12; void vdiv(ptr a, float b) {13; for (int i = 0; i != 1024; ++i)14; a[i] /= b;15; }16 17define void @vdiv(ptr %a, float %b) {18; CHECK-LABEL: define void @vdiv(19; CHECK-SAME: ptr captures(none) [[A:%.*]], float [[B:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {20; CHECK-NEXT: [[ENTRY:.*]]:21; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x float> poison, float [[B]], i64 022; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x float> [[BROADCAST_SPLATINSERT]], <4 x float> poison, <4 x i32> zeroinitializer23; CHECK-NEXT: [[TMP0:%.*]] = fdiv fast <4 x float> splat (float 1.000000e+00), [[BROADCAST_SPLAT]]24; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]25; CHECK: [[VECTOR_BODY]]:26; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]27; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX]]28; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP1]], align 4, !tbaa [[FLOAT_TBAA3:![0-9]+]]29; CHECK-NEXT: [[TMP3:%.*]] = fmul fast <4 x float> [[WIDE_LOAD]], [[TMP0]]30; CHECK-NEXT: store <4 x float> [[TMP3]], ptr [[TMP1]], align 4, !tbaa [[FLOAT_TBAA3]]31; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 432; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], 102433; CHECK-NEXT: br i1 [[TMP5]], label %[[FOR_COND_CLEANUP:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP7:![0-9]+]]34; CHECK: [[FOR_COND_CLEANUP]]:35; CHECK-NEXT: ret void36;37entry:38 %a.addr = alloca ptr, align 839 %b.addr = alloca float, align 440 %i = alloca i32, align 441 store ptr %a, ptr %a.addr, align 8, !tbaa !342 store float %b, ptr %b.addr, align 4, !tbaa !743 call void @llvm.lifetime.start.p0(ptr %i)44 store i32 0, ptr %i, align 4, !tbaa !945 br label %for.cond46 47for.cond: ; preds = %for.inc, %entry48 %0 = load i32, ptr %i, align 4, !tbaa !949 %cmp = icmp ne i32 %0, 102450 br i1 %cmp, label %for.body, label %for.cond.cleanup51 52for.cond.cleanup: ; preds = %for.cond53 call void @llvm.lifetime.end.p0(ptr %i)54 br label %for.end55 56for.body: ; preds = %for.cond57 %1 = load float, ptr %b.addr, align 4, !tbaa !758 %2 = load ptr, ptr %a.addr, align 8, !tbaa !359 %3 = load i32, ptr %i, align 4, !tbaa !960 %idxprom = sext i32 %3 to i6461 %arrayidx = getelementptr inbounds float, ptr %2, i64 %idxprom62 %4 = load float, ptr %arrayidx, align 4, !tbaa !763 %div = fdiv fast float %4, %164 store float %div, ptr %arrayidx, align 4, !tbaa !765 br label %for.inc66 67for.inc: ; preds = %for.body68 %5 = load i32, ptr %i, align 4, !tbaa !969 %inc = add nsw i32 %5, 170 store i32 %inc, ptr %i, align 4, !tbaa !971 br label %for.cond, !llvm.loop !1172 73for.end: ; preds = %for.cond.cleanup74 ret void75}76 77declare void @llvm.lifetime.start.p0(ptr nocapture)78declare void @llvm.lifetime.end.p0(ptr nocapture)79 80!llvm.module.flags = !{!0, !1}81!llvm.ident = !{!2}82 83!0 = !{i32 1, !"wchar_size", i32 4}84!1 = !{i32 7, !"PIC Level", i32 2}85!2 = !{!"clang version 13.0.0 (https://github.com/llvm/llvm-project.git 437fb428178716db014f75701f3cdea6dea63f59)"}86!3 = !{!4, !4, i64 0}87!4 = !{!"any pointer", !5, i64 0}88!5 = !{!"omnipotent char", !6, i64 0}89!6 = !{!"Simple C/C++ TBAA"}90!7 = !{!8, !8, i64 0}91!8 = !{!"float", !5, i64 0}92!9 = !{!10, !10, i64 0}93!10 = !{!"int", !5, i64 0}94!11 = distinct !{!11, !12, !13}95!12 = !{!"llvm.loop.mustprogress"}96!13 = !{!"llvm.loop.unroll.disable"}97;.98; CHECK: [[FLOAT_TBAA3]] = !{[[META4:![0-9]+]], [[META4]], i64 0}99; CHECK: [[META4]] = !{!"float", [[META5:![0-9]+]], i64 0}100; CHECK: [[META5]] = !{!"omnipotent char", [[META6:![0-9]+]], i64 0}101; CHECK: [[META6]] = !{!"Simple C/C++ TBAA"}102; CHECK: [[LOOP7]] = distinct !{[[LOOP7]], [[META8:![0-9]+]], [[META9:![0-9]+]], [[META10:![0-9]+]], [[META11:![0-9]+]]}103; CHECK: [[META8]] = !{!"llvm.loop.mustprogress"}104; CHECK: [[META9]] = !{!"llvm.loop.unroll.disable"}105; CHECK: [[META10]] = !{!"llvm.loop.isvectorized", i32 1}106; CHECK: [[META11]] = !{!"llvm.loop.unroll.runtime.disable"}107;.108