1; RUN: opt -basicaa -loop-distribute -enable-loop-distribute -loop-simplify -scoped-noalias \ 2; RUN: -loop-versioning -S < %s | FileCheck %s 3 4; Test the metadata generated when versioning an already versioned loop. Here 5; we invoke loop distribution to perform the first round of versioning. It 6; adds memchecks for accesses that can alias across the distribution boundary. 7; Then we further version the distributed loops to fully disambiguate accesses 8; within each. 9; 10; So as an example, we add noalias between C and A during the versioning 11; within loop distribution and then add noalias between C and D during the 12; second explicit versioning step: 13; 14; for (i = 0; i < n; i++) { 15; A[i + 1] = A[i] * B[i]; 16; ------------------------------- 17; C[i] = D[i] * E[i]; 18; } 19 20; To see it easier what's going on, I expanded every noalias/scope metadata 21; reference below in a comment. For a scope I use the format scope(domain), 22; e.g. scope 17 in domain 15 is written as 17(15). 23 24target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128" 25 26@B = common global i32* null, align 8 27@A = common global i32* null, align 8 28@C = common global i32* null, align 8 29@D = common global i32* null, align 8 30@E = common global i32* null, align 8 31 32define void @f() { 33entry: 34 %a = load i32*, i32** @A, align 8 35 %b = load i32*, i32** @B, align 8 36 %c = load i32*, i32** @C, align 8 37 %d = load i32*, i32** @D, align 8 38 %e = load i32*, i32** @E, align 8 39 br label %for.body 40 41for.body: ; preds = %for.body, %entry 42 %ind = phi i64 [ 0, %entry ], [ %add, %for.body ] 43 44 %arrayidxA = getelementptr inbounds i32, i32* %a, i64 %ind 45 46; CHECK: %loadA.ldist1 = {{.*}} !noalias !25 47; A noalias C: !25 -> { 17(15), 18(15), 19(15), 26(24) } 48; ^^^^^^ 49 %loadA = load i32, i32* %arrayidxA, align 4 50 51 %arrayidxB = getelementptr inbounds i32, i32* %b, i64 %ind 52 %loadB = load i32, i32* %arrayidxB, align 4 53 54 %mulA = mul i32 %loadB, %loadA 55 56 %add = add nuw nsw i64 %ind, 1 57 %arrayidxA_plus_4 = getelementptr inbounds i32, i32* %a, i64 %add 58 store i32 %mulA, i32* %arrayidxA_plus_4, align 4 59 60; CHECK: for.body: 61 62 %arrayidxD = getelementptr inbounds i32, i32* %d, i64 %ind 63 64; CHECK: %loadD = {{.*}} !alias.scope !31 65; D's scope: !31 -> { 18(15), 32(33) } 66; ^^^^^^ 67 %loadD = load i32, i32* %arrayidxD, align 4 68 69 %arrayidxE = getelementptr inbounds i32, i32* %e, i64 %ind 70 71; CHECK: %loadE = {{.*}} !alias.scope !34 72; E's scope: !34 -> { 19(15), 35(33) } 73; ^^^^^^ 74 %loadE = load i32, i32* %arrayidxE, align 4 75 76 %mulC = mul i32 %loadD, %loadE 77 78 %arrayidxC = getelementptr inbounds i32, i32* %c, i64 %ind 79 80; CHECK: store i32 %mulC, {{.*}} !alias.scope !36, !noalias !38 81; C's scope: !36 -> { 17(15), 37(33) } 82; ^^^^^^ 83; C noalias D and E: !38 -> { 21(15), 32(33), 35(33) } 84; ^^^^^^ ^^^^^^ 85 store i32 %mulC, i32* %arrayidxC, align 4 86 87 %exitcond = icmp eq i64 %add, 20 88 br i1 %exitcond, label %for.end, label %for.body 89 90for.end: ; preds = %for.body 91 ret void 92} 93 94; Domain for the second loop versioning for the top loop after 95; distribution. 96; CHECK: !15 = distinct !{!15, !"LVerDomain"} 97; CHECK: !17 = distinct !{!17, !15} 98; CHECK: !25 = !{!17, !18, !19, !26} 99; CHECK: !31 = !{!18, !32} 100; CHECK: !32 = distinct !{!32, !33} 101; Domain for the second loop versioning for the bottom loop after 102; distribution. 103; CHECK: !33 = distinct !{!33, !"LVerDomain"} 104; CHECK: !34 = !{!19, !35} 105; CHECK: !35 = distinct !{!35, !33} 106; CHECK: !36 = !{!17, !37} 107; CHECK: !38 = !{!21, !32, !35} 108