1; RUN: opt -basicaa -loop-accesses -analyze < %s | FileCheck %s 2 3; This loop: 4; 5; int **A; 6; for (i) 7; for (j) { 8; A[i][j] = A[i-1][j] * B[j] 9; B[j+1] = 2 // backward dep between this and the previous 10; } 11; 12; is transformed by Load-PRE to stash away A[i] for the next iteration of the 13; outer loop: 14; 15; Curr = A[0]; // Prev_0 16; for (i: 1..N) { 17; Prev = Curr; // Prev = PHI (Prev_0, Curr) 18; Curr = A[i]; 19; for (j: 0..N) { 20; Curr[j] = Prev[j] * B[j] 21; B[j+1] = 2 // backward dep between this and the previous 22; } 23; } 24; 25; Since A[i] and A[i-1] are likely to be independent, getUnderlyingObjects 26; should not assume that Curr and Prev share the same underlying object. 27; 28; If it did we would try to dependence-analyze Curr and Prev and the analysis 29; would fail with non-constant distance. 30; 31; To illustrate one of the negative consequences of this, if the loop has a 32; backward dependence we won't detect this but instead fully fall back on 33; memchecks (that is what LAA does after encountering a case of non-constant 34; distance). 35 36target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128" 37target triple = "x86_64-apple-macosx10.10.0" 38 39; CHECK: for_j.body: 40; CHECK-NEXT: Report: unsafe dependent memory operations in loop 41; CHECK-NEXT: Dependences: 42; CHECK-NEXT: Backward: 43; CHECK-NEXT: %loadB = load i8, i8* %gepB, align 1 -> 44; CHECK-NEXT: store i8 2, i8* %gepB_plus_one, align 1 45 46define void @f(i8** noalias %A, i8* noalias %B, i64 %N) { 47for_i.preheader: 48 %prev_0 = load i8*, i8** %A, align 8 49 br label %for_i.body 50 51for_i.body: 52 %i = phi i64 [1, %for_i.preheader], [%i.1, %for_j.end] 53 %prev = phi i8* [%prev_0, %for_i.preheader], [%curr, %for_j.end] 54 %gep = getelementptr inbounds i8*, i8** %A, i64 %i 55 %curr = load i8*, i8** %gep, align 8 56 br label %for_j.preheader 57 58for_j.preheader: 59 br label %for_j.body 60 61for_j.body: 62 %j = phi i64 [0, %for_j.preheader], [%j.1, %for_j.body] 63 64 %gepPrev = getelementptr inbounds i8, i8* %prev, i64 %j 65 %gepCurr = getelementptr inbounds i8, i8* %curr, i64 %j 66 %gepB = getelementptr inbounds i8, i8* %B, i64 %j 67 68 %loadPrev = load i8, i8* %gepPrev, align 1 69 %loadB = load i8, i8* %gepB, align 1 70 71 %mul = mul i8 %loadPrev, %loadB 72 73 store i8 %mul, i8* %gepCurr, align 1 74 75 %gepB_plus_one = getelementptr inbounds i8, i8* %gepB, i64 1 76 store i8 2, i8* %gepB_plus_one, align 1 77 78 %j.1 = add nuw i64 %j, 1 79 %exitcondj = icmp eq i64 %j.1, %N 80 br i1 %exitcondj, label %for_j.end, label %for_j.body 81 82for_j.end: 83 84 %i.1 = add nuw i64 %i, 1 85 %exitcond = icmp eq i64 %i.1, %N 86 br i1 %exitcond, label %for_i.end, label %for_i.body 87 88for_i.end: 89 ret void 90} 91