| 1 |
; REQUIRES: asserts |
| 2 |
|
| 3 |
; RUN: opt -passes=loop-interchange -debug -disable-output %s 2>&1 | FileCheck %s |
| 4 |
|
| 5 |
@A = global [16 x [16 x i32]] zeroinitializer |
| 6 |
|
| 7 |
; Check that the CacheCost is calculated only when required. In this case, it |
| 8 |
; is computed after passing the legality check. |
| 9 |
; |
| 10 |
; for (i = 0; i < 16; i++) |
| 11 |
; for (j = 0; j < 16; j++) |
| 12 |
; A[j][i] += 1; |
| 13 |
|
| 14 |
; CHECK: Loops 'for.i.header' and 'for.j' are legal to interchange |
| 15 |
; CHECK: Compute CacheCost |
| 16 |
define void @legal_to_interchange() { |
| 17 |
entry: |
| 18 |
br label %for.i.header |
| 19 |
|
| 20 |
for.i.header: |
| 21 |
%i = phi i32 [ 0, %entry ], [ %i.next, %for.i.latch ] |
| 22 |
br label %for.j |
| 23 |
|
| 24 |
for.j: |
| 25 |
%j = phi i32 [ 0, %for.i.header ], [ %j.next, %for.j ] |
| 26 |
%idx = getelementptr inbounds [16 x [16 x i32]], ptr @A, i32 0, i32 %j, i32 %i |
| 27 |
%val = load i32, ptr %idx |
| 28 |
%inc = add i32 %val, 1 |
| 29 |
store i32 %inc, ptr %idx |
| 30 |
%j.next = add i32 %j, 1 |
| 31 |
%j.exit = icmp eq i32 %j.next, 16 |
| 32 |
br i1 %j.exit, label %for.i.latch, label %for.j |
| 33 |
|
| 34 |
for.i.latch: |
| 35 |
%i.next = add i32 %i, 1 |
| 36 |
%i.exit = icmp eq i32 %i.next, 16 |
| 37 |
br i1 %i.exit, label %exit, label %for.i.header |
| 38 |
|
| 39 |
exit: |
| 40 |
ret void |
| 41 |
} |
| 42 |
|
| 43 |
; Check that the CacheCost is not calculated when not required. In this case, |
| 44 |
; the legality check always fails so that we do not need to compute the |
| 45 |
; CacheCost. |
| 46 |
; |
| 47 |
; for (i = 0; i < 16; i++) |
| 48 |
; for (j = 0; j < 16; j++) |
| 49 |
; A[j][i] = A[i][j]; |
| 50 |
|
| 51 |
; CHECK-NOT: Compute CacheCost |
| 52 |
define void @illegal_to_interchange() { |
| 53 |
entry: |
| 54 |
br label %for.i.header |
| 55 |
|
| 56 |
for.i.header: |
| 57 |
%i = phi i32 [ 0, %entry ], [ %i.next, %for.i.latch ] |
| 58 |
br label %for.j |
| 59 |
|
| 60 |
for.j: |
| 61 |
%j = phi i32 [ 0, %for.i.header ], [ %j.next, %for.j ] |
| 62 |
%idx.load = getelementptr inbounds [16 x [16 x i32]], ptr @A, i32 0, i32 %i, i32 %j |
| 63 |
%idx.store = getelementptr inbounds [16 x [16 x i32]], ptr @A, i32 0, i32 %j, i32 %i |
| 64 |
%val = load i32, ptr %idx.load |
| 65 |
store i32 %val, ptr %idx.store |
| 66 |
%j.next = add i32 %j, 1 |
| 67 |
%j.exit = icmp eq i32 %j.next, 16 |
| 68 |
br i1 %j.exit, label %for.i.latch, label %for.j |
| 69 |
|
| 70 |
for.i.latch: |
| 71 |
%i.next = add i32 %i, 1 |
| 72 |
%i.exit = icmp eq i32 %i.next, 16 |
| 73 |
br i1 %i.exit, label %exit, label %for.i.header |
| 74 |
|
| 75 |
exit: |
| 76 |
ret void |
| 77 |
} |