| 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 |
;; In conclusion, we must not interchange the loops. |
| 19 |
|
| 20 |
; CHECK: --- !Missed |
| 21 |
; CHECK-NEXT: Pass: loop-interchange |
| 22 |
; CHECK-NEXT: Name: Dependence |
| 23 |
; CHECK-NEXT: Function: f |
| 24 |
; CHECK-NEXT: Args: |
| 25 |
; CHECK-NEXT: - String: All loops have dependencies in all directions. |
| 26 |
; CHECK-NEXT: ... |
| 27 |
|
| 28 |
|
| 29 |
@a = dso_local global [16 x [16 x [16 x i32]]] zeroinitializer, align 4 |
| 30 |
|
| 31 |
define dso_local void @f() { |
| 32 |
entry: |
| 33 |
br label %for.cond1.preheader |
| 34 |
|
| 35 |
for.cond1.preheader: |
| 36 |
%i.039 = phi i32 [ 0, %entry ], [ %inc26, %for.cond.cleanup3 ] |
| 37 |
%sub = add nuw nsw i32 %i.039, 3 |
| 38 |
%idxprom = zext nneg i32 %sub to i64 |
| 39 |
%mul = shl nuw nsw i32 %i.039, 1 |
| 40 |
%idxprom13 = zext nneg i32 %mul to i64 |
| 41 |
br label %for.cond5.preheader |
| 42 |
|
| 43 |
for.cond.cleanup: |
| 44 |
ret void |
| 45 |
|
| 46 |
for.cond5.preheader: |
| 47 |
%j.038 = phi i32 [ 1, %for.cond1.preheader ], [ %inc23, %for.cond.cleanup7 ] |
| 48 |
%idxprom11 = zext nneg i32 %j.038 to i64 |
| 49 |
%sub18 = add nsw i32 %j.038, -1 |
| 50 |
%idxprom19 = sext i32 %sub18 to i64 |
| 51 |
br label %for.body8 |
| 52 |
|
| 53 |
for.cond.cleanup3: |
| 54 |
%inc26 = add nuw nsw i32 %i.039, 1 |
| 55 |
%cmp = icmp samesign ult i32 %i.039, 3 |
| 56 |
br i1 %cmp, label %for.cond1.preheader, label %for.cond.cleanup |
| 57 |
|
| 58 |
for.cond.cleanup7: |
| 59 |
%inc23 = add nuw nsw i32 %j.038, 1 |
| 60 |
%cmp2 = icmp samesign ult i32 %j.038, 7 |
| 61 |
br i1 %cmp2, label %for.cond5.preheader, label %for.cond.cleanup3 |
| 62 |
|
| 63 |
for.body8: |
| 64 |
%k.037 = phi i32 [ 1, %for.cond5.preheader ], [ %add15, %for.body8 ] |
| 65 |
%idxprom9 = zext nneg i32 %k.037 to i64 |
| 66 |
%arrayidx12 = getelementptr inbounds nuw [16 x [16 x [16 x i32]]], ptr @a, i64 0, i64 %idxprom, i64 %idxprom9, i64 %idxprom11 |
| 67 |
%0 = load i32, ptr %arrayidx12, align 4 |
| 68 |
%add15 = add nuw nsw i32 %k.037, 1 |
| 69 |
%idxprom16 = zext nneg i32 %add15 to i64 |
| 70 |
%arrayidx20 = getelementptr inbounds [16 x [16 x [16 x i32]]], ptr @a, i64 0, i64 %idxprom13, i64 %idxprom16, i64 %idxprom19 |
| 71 |
%1 = load i32, ptr %arrayidx20, align 4 |
| 72 |
%sub21 = sub nsw i32 %1, %0 |
| 73 |
store i32 %sub21, ptr %arrayidx20, align 4 |
| 74 |
%cmp6 = icmp samesign ult i32 %k.037, 7 |
| 75 |
br i1 %cmp6, label %for.body8, label %for.cond.cleanup7 |
| 76 |
} |