1; RUN: opt < %s -aa-pipeline=tbaa,basic-aa -passes=loop-vectorize,dce,instcombine,simplifycfg -force-vector-interleave=1 -force-vector-width=4 -simplifycfg-require-and-preserve-domtree=1 -S | FileCheck %s 2; RUN: opt < %s -aa-pipeline=basic-aa -passes=loop-vectorize,dce,instcombine,simplifycfg -force-vector-interleave=1 -force-vector-width=4 -simplifycfg-require-and-preserve-domtree=1 -S | FileCheck %s --check-prefix=CHECK-NOTBAA 3target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" 4 5; TBAA partitions the accesses in this loop, so it can be vectorized without 6; runtime checks. 7define i32 @test1(ptr nocapture %a, ptr nocapture readonly %b) { 8; CHECK-LABEL: @test1 9; CHECK: entry: 10; CHECK-NEXT: br label %vector.body 11; CHECK: vector.body: 12 13; CHECK: load <4 x float>, ptr %{{.*}}, align 4, !tbaa 14; CHECK: store <4 x i32> %{{.*}}, ptr %{{.*}}, align 4, !tbaa 15 16; CHECK: ret i32 0 17 18; CHECK-NOTBAA-LABEL: @test1 19; CHECK-NOTBAA: entry: 20; CHECK-NOTBAA: icmp ult i64 21; CHECK-NOTBAA-NOT: icmp 22; CHECK-NOTBAA: br i1 {{.+}}, label %for.body, label %vector.body 23 24; CHECK-NOTBAA: load <4 x float>, ptr %{{.*}}, align 4, !tbaa 25; CHECK-NOTBAA: store <4 x i32> %{{.*}}, ptr %{{.*}}, align 4, !tbaa 26 27; CHECK-NOTBAA: ret i32 0 28 29entry: 30 br label %for.body 31 32for.body: ; preds = %for.body, %entry 33 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 34 %arrayidx = getelementptr inbounds float, ptr %b, i64 %indvars.iv 35 %0 = load float, ptr %arrayidx, align 4, !tbaa !0 36 %conv = fptosi float %0 to i32 37 %arrayidx2 = getelementptr inbounds i32, ptr %a, i64 %indvars.iv 38 store i32 %conv, ptr %arrayidx2, align 4, !tbaa !4 39 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 40 %exitcond = icmp eq i64 %indvars.iv.next, 1600 41 br i1 %exitcond, label %for.end, label %for.body 42 43for.end: ; preds = %for.body 44 ret i32 0 45} 46 47; This test is like the first, except here there is still one runtime check 48; required. Without TBAA, however, two checks are required. 49define i32 @test2(ptr nocapture readonly %a, ptr nocapture readonly %b, ptr nocapture %c) { 50; CHECK-LABEL: @test2 51; CHECK: entry: 52; CHECK: icmp ult i64 53; CHECK-NOT: icmp 54; CHECK: br i1 {{.+}}, label %for.body, label %vector.body 55 56; CHECK: load <4 x float>, ptr %{{.*}}, align 4, !tbaa 57; CHECK: store <4 x float> %{{.*}}, ptr %{{.*}}, align 4, !tbaa 58 59; CHECK: ret i32 0 60 61; CHECK-NOTBAA-LABEL: @test2 62; CHECK-NOTBAA: entry: 63; CHECK-NOTBAA: icmp ult i64 64; CHECK-NOTBAA: icmp ult i64 65; CHECK-NOTBAA-NOT: icmp 66; CHECK-NOTBAA: br i1 {{.+}}, label %for.body, label %vector.body 67 68; CHECK-NOTBAA: load <4 x float>, ptr %{{.*}}, align 4, !tbaa 69; CHECK-NOTBAA: store <4 x float> %{{.*}}, ptr %{{.*}}, align 4, !tbaa 70 71; CHECK-NOTBAA: ret i32 0 72 73entry: 74 br label %for.body 75 76for.body: ; preds = %for.body, %entry 77 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 78 %arrayidx = getelementptr inbounds float, ptr %b, i64 %indvars.iv 79 %0 = load float, ptr %arrayidx, align 4, !tbaa !0 80 %arrayidx2 = getelementptr inbounds i32, ptr %a, i64 %indvars.iv 81 %1 = load i32, ptr %arrayidx2, align 4, !tbaa !4 82 %conv = sitofp i32 %1 to float 83 %mul = fmul float %0, %conv 84 %arrayidx4 = getelementptr inbounds float, ptr %c, i64 %indvars.iv 85 store float %mul, ptr %arrayidx4, align 4, !tbaa !0 86 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 87 %exitcond = icmp eq i64 %indvars.iv.next, 1600 88 br i1 %exitcond, label %for.end, label %for.body 89 90for.end: ; preds = %for.body 91 ret i32 0 92} 93 94!0 = !{!1, !1, i64 0} 95!1 = !{!"float", !2, i64 0} 96!2 = !{!"omnipotent char", !3, i64 0} 97!3 = !{!"Simple C/C++ TBAA"} 98!4 = !{!5, !5, i64 0} 99!5 = !{!"int", !2, i64 0} 100