|
| 1 | +; RUN: opt < %s -passes=loop-interchange -pass-remarks-missed='loop-interchange' -pass-remarks-output=%t \ |
| 2 | +; RUN: -verify-dom-info -verify-loop-info -verify-loop-lcssa |
| 3 | +; RUN: FileCheck --input-file=%t %s |
| 4 | + |
| 5 | +;; The original code: |
| 6 | +;; |
| 7 | +;; #define N 4 |
| 8 | +;; int a[N*N][N*N][N*N]; |
| 9 | +;; void f() { |
| 10 | +;; for (int i = 0; i < N; i++) |
| 11 | +;; for (int j = 1; j < 2*N; j++) |
| 12 | +;; for (int k = 1; k < 2*N; k++) |
| 13 | +;; a[2*i][k+1][j-1] -= a[i+N-1][k][j]; |
| 14 | +;; } |
| 15 | +;; |
| 16 | +;; The entry of the direction vector for the outermost loop is `DVEntry::LE`. |
| 17 | +;; We need to treat this as `*`, not `<`. See issue #123920 for details. |
| 18 | + |
| 19 | +; CHECK: --- !Missed |
| 20 | +; CHECK-NEXT: Pass: loop-interchange |
| 21 | +; CHECK-NEXT: Name: Dependence |
| 22 | +; CHECK-NEXT: Function: f |
| 23 | +; CHECK: --- !Missed |
| 24 | +; CHECK-NEXT: Pass: loop-interchange |
| 25 | +; CHECK-NEXT: Name: Dependence |
| 26 | +; CHECK-NEXT: Function: f |
| 27 | + |
| 28 | +@a = dso_local global [16 x [16 x [16 x i32]]] zeroinitializer, align 4 |
| 29 | + |
| 30 | +define dso_local void @f() { |
| 31 | +entry: |
| 32 | + br label %for.cond1.preheader |
| 33 | + |
| 34 | +for.cond1.preheader: |
| 35 | + %i.039 = phi i32 [ 0, %entry ], [ %inc26, %for.cond.cleanup3 ] |
| 36 | + %sub = add nuw nsw i32 %i.039, 3 |
| 37 | + %idxprom = zext nneg i32 %sub to i64 |
| 38 | + %mul = shl nuw nsw i32 %i.039, 1 |
| 39 | + %idxprom13 = zext nneg i32 %mul to i64 |
| 40 | + br label %for.cond5.preheader |
| 41 | + |
| 42 | +for.cond.cleanup: |
| 43 | + ret void |
| 44 | + |
| 45 | +for.cond5.preheader: |
| 46 | + %j.038 = phi i32 [ 1, %for.cond1.preheader ], [ %inc23, %for.cond.cleanup7 ] |
| 47 | + %idxprom11 = zext nneg i32 %j.038 to i64 |
| 48 | + %sub18 = add nsw i32 %j.038, -1 |
| 49 | + %idxprom19 = sext i32 %sub18 to i64 |
| 50 | + br label %for.body8 |
| 51 | + |
| 52 | +for.cond.cleanup3: |
| 53 | + %inc26 = add nuw nsw i32 %i.039, 1 |
| 54 | + %cmp = icmp samesign ult i32 %i.039, 3 |
| 55 | + br i1 %cmp, label %for.cond1.preheader, label %for.cond.cleanup |
| 56 | + |
| 57 | +for.cond.cleanup7: |
| 58 | + %inc23 = add nuw nsw i32 %j.038, 1 |
| 59 | + %cmp2 = icmp samesign ult i32 %j.038, 7 |
| 60 | + br i1 %cmp2, label %for.cond5.preheader, label %for.cond.cleanup3 |
| 61 | + |
| 62 | +for.body8: |
| 63 | + %k.037 = phi i32 [ 1, %for.cond5.preheader ], [ %add15, %for.body8 ] |
| 64 | + %idxprom9 = zext nneg i32 %k.037 to i64 |
| 65 | + %arrayidx12 = getelementptr inbounds nuw [16 x [16 x [16 x i32]]], ptr @a, i64 0, i64 %idxprom, i64 %idxprom9, i64 %idxprom11 |
| 66 | + %0 = load i32, ptr %arrayidx12, align 4 |
| 67 | + %add15 = add nuw nsw i32 %k.037, 1 |
| 68 | + %idxprom16 = zext nneg i32 %add15 to i64 |
| 69 | + %arrayidx20 = getelementptr inbounds [16 x [16 x [16 x i32]]], ptr @a, i64 0, i64 %idxprom13, i64 %idxprom16, i64 %idxprom19 |
| 70 | + %1 = load i32, ptr %arrayidx20, align 4 |
| 71 | + %sub21 = sub nsw i32 %1, %0 |
| 72 | + store i32 %sub21, ptr %arrayidx20, align 4 |
| 73 | + %cmp6 = icmp samesign ult i32 %k.037, 7 |
| 74 | + br i1 %cmp6, label %for.body8, label %for.cond.cleanup7 |
| 75 | +} |
0 commit comments