Skip to content

Commit b1d69e9

Browse files
kasuga-fjkasuga-fj
committed
Revert "[LoopUtils] Fix metadata generated by makeFollowupLoopID"
This reverts commit 889f40c.
1 parent 889f40c commit b1d69e9

File tree

2 files changed

+20
-111
lines changed

2 files changed

+20
-111
lines changed

‎llvm/lib/Transforms/Utils/LoopUtils.cpp

Lines changed: 0 additions & 29 deletions
Original file line numberDiff line numberDiff line change
@@ -317,35 +317,6 @@ std::optional<MDNode *> llvm::makeFollowupLoopID(
317317

318318
HasAnyFollowup = true;
319319
for (const MDOperand &Option : drop_begin(FollowupNode->operands())) {
320-
// The followup metadata typically forms as follows:
321-
//
322-
// !0 = distinct !{!0, !1, !2}
323-
// !1 = !{!"llvm.loop.distribute.enable", i1 true}
324-
// !2 = !{!"llvm.loop.distribute.followup_all", !3}
325-
// !3 = distinct !{!3, !4}
326-
// !4 = !{!"llvm.loop.vectorize.enable", i1 true}
327-
//
328-
// If we push Option (!3 in this case) in MDs, the new metadata looks
329-
// something like:
330-
//
331-
// !5 = distinct !{!5, !3}
332-
//
333-
// This doesn't contain !4, so the vectorization pass doesn't recognize
334-
// this loop as vectorization enabled. To make the new metadata contain !4
335-
// instead of !3, traverse all of Option's operands and push them into
336-
// MDs if Option seems to be a LoopID.
337-
if (auto *MDN = dyn_cast<MDNode>(Option)) {
338-
// TODO: Is there a proper way to detect LoopID?
339-
if (MDN->getNumOperands() > 1 && MDN->getOperand(0) == MDN) {
340-
for (const MDOperand &NestedOption : drop_begin(MDN->operands())) {
341-
MDs.push_back(NestedOption.get());
342-
Changed = true;
343-
}
344-
continue;
345-
}
346-
}
347-
348-
// If Option does't seem to be a LoopID, push it as it is.
349320
MDs.push_back(Option.get());
350321
Changed = true;
351322
}

‎llvm/test/Transforms/LoopVectorize/make-followup-loop-id.ll

Lines changed: 20 additions & 82 deletions
Original file line numberDiff line numberDiff line change
@@ -11,6 +11,10 @@
1111
; a[i] *= x;
1212
; }
1313
; }
14+
;
15+
; FIXME: Currently unrolling is not applied. This is because the new Loop ID
16+
; created after vectorization does not directly contain unroll metadata.
17+
; Unexpected nests have been created.
1418
define void @f(ptr noundef captures(none) %a, float noundef %x) {
1519
; CHECK-LABEL: define void @f(
1620
; CHECK-SAME: ptr noundef captures(none) [[A:%.*]], float noundef [[X:%.*]]) {
@@ -21,47 +25,14 @@ define void @f(ptr noundef captures(none) %a, float noundef %x) {
2125
; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x float> [[BROADCAST_SPLATINSERT]], <4 x float> poison, <4 x i32> zeroinitializer
2226
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
2327
; CHECK: [[VECTOR_BODY]]:
24-
; CHECK-NEXT: [[INDEX_NEXT_6:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
28+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
29+
; CHECK-NEXT: [[INDEX_NEXT_6:%.*]] = add i64 [[INDEX]], 0
2530
; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_6]]
26-
; CHECK-NEXT: [[WIDE_LOAD_7:%.*]] = load <4 x float>, ptr [[TMP14]], align 4
31+
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw float, ptr [[TMP14]], i32 0
32+
; CHECK-NEXT: [[WIDE_LOAD_7:%.*]] = load <4 x float>, ptr [[TMP2]], align 4
2733
; CHECK-NEXT: [[TMP15:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_7]]
28-
; CHECK-NEXT: store <4 x float> [[TMP15]], ptr [[TMP14]], align 4
29-
; CHECK-NEXT: [[INDEX_NEXT1:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 4
30-
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT1]]
31-
; CHECK-NEXT: [[WIDE_LOAD_1:%.*]] = load <4 x float>, ptr [[TMP2]], align 4
32-
; CHECK-NEXT: [[TMP3:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_1]]
33-
; CHECK-NEXT: store <4 x float> [[TMP3]], ptr [[TMP2]], align 4
34-
; CHECK-NEXT: [[INDEX_NEXT_1:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 8
35-
; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_1]]
36-
; CHECK-NEXT: [[WIDE_LOAD_2:%.*]] = load <4 x float>, ptr [[TMP16]], align 4
37-
; CHECK-NEXT: [[TMP5:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_2]]
38-
; CHECK-NEXT: store <4 x float> [[TMP5]], ptr [[TMP16]], align 4
39-
; CHECK-NEXT: [[INDEX_NEXT_2:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 12
40-
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_2]]
41-
; CHECK-NEXT: [[WIDE_LOAD_3:%.*]] = load <4 x float>, ptr [[TMP6]], align 4
42-
; CHECK-NEXT: [[TMP7:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_3]]
43-
; CHECK-NEXT: store <4 x float> [[TMP7]], ptr [[TMP6]], align 4
44-
; CHECK-NEXT: [[INDEX_NEXT_3:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 16
45-
; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_3]]
46-
; CHECK-NEXT: [[WIDE_LOAD_4:%.*]] = load <4 x float>, ptr [[TMP8]], align 4
47-
; CHECK-NEXT: [[TMP9:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_4]]
48-
; CHECK-NEXT: store <4 x float> [[TMP9]], ptr [[TMP8]], align 4
49-
; CHECK-NEXT: [[INDEX_NEXT_4:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 20
50-
; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_4]]
51-
; CHECK-NEXT: [[WIDE_LOAD_5:%.*]] = load <4 x float>, ptr [[TMP10]], align 4
52-
; CHECK-NEXT: [[TMP11:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_5]]
53-
; CHECK-NEXT: store <4 x float> [[TMP11]], ptr [[TMP10]], align 4
54-
; CHECK-NEXT: [[INDEX_NEXT_5:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 24
55-
; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_5]]
56-
; CHECK-NEXT: [[WIDE_LOAD_6:%.*]] = load <4 x float>, ptr [[TMP12]], align 4
57-
; CHECK-NEXT: [[TMP13:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_6]]
58-
; CHECK-NEXT: store <4 x float> [[TMP13]], ptr [[TMP12]], align 4
59-
; CHECK-NEXT: [[INDEX_NEXT_7:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 28
60-
; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_7]]
61-
; CHECK-NEXT: [[WIDE_LOAD_8:%.*]] = load <4 x float>, ptr [[TMP17]], align 4
62-
; CHECK-NEXT: [[TMP18:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_8]]
63-
; CHECK-NEXT: store <4 x float> [[TMP18]], ptr [[TMP17]], align 4
64-
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw nsw i64 [[INDEX_NEXT_6]], 32
34+
; CHECK-NEXT: store <4 x float> [[TMP15]], ptr [[TMP2]], align 4
35+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
6536
; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
6637
; CHECK-NEXT: br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
6738
; CHECK: [[MIDDLE_BLOCK]]:
@@ -70,49 +41,14 @@ define void @f(ptr noundef captures(none) %a, float noundef %x) {
7041
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ 1024, %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
7142
; CHECK-NEXT: br label %[[FOR_BODY:.*]]
7243
; CHECK: [[FOR_BODY]]:
73-
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT_7:%.*]], %[[FOR_BODY]] ]
44+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[FOR_BODY]] ]
7445
; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV]]
7546
; CHECK-NEXT: [[LOAD:%.*]] = load float, ptr [[ARRAYIDX]], align 4
7647
; CHECK-NEXT: [[MUL:%.*]] = fmul float [[X]], [[LOAD]]
7748
; CHECK-NEXT: store float [[MUL]], ptr [[ARRAYIDX]], align 4
78-
; CHECK-NEXT: [[IV_NEXT:%.*]] = add nuw nsw i64 [[IV]], 1
79-
; CHECK-NEXT: [[ARRAYIDX_1:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV_NEXT]]
80-
; CHECK-NEXT: [[LOAD_1:%.*]] = load float, ptr [[ARRAYIDX_1]], align 4
81-
; CHECK-NEXT: [[MUL_1:%.*]] = fmul float [[X]], [[LOAD_1]]
82-
; CHECK-NEXT: store float [[MUL_1]], ptr [[ARRAYIDX_1]], align 4
83-
; CHECK-NEXT: [[IV_NEXT_1:%.*]] = add nuw nsw i64 [[IV]], 2
84-
; CHECK-NEXT: [[ARRAYIDX_2:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV_NEXT_1]]
85-
; CHECK-NEXT: [[LOAD_2:%.*]] = load float, ptr [[ARRAYIDX_2]], align 4
86-
; CHECK-NEXT: [[MUL_2:%.*]] = fmul float [[X]], [[LOAD_2]]
87-
; CHECK-NEXT: store float [[MUL_2]], ptr [[ARRAYIDX_2]], align 4
88-
; CHECK-NEXT: [[IV_NEXT_2:%.*]] = add nuw nsw i64 [[IV]], 3
89-
; CHECK-NEXT: [[ARRAYIDX_3:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV_NEXT_2]]
90-
; CHECK-NEXT: [[LOAD_3:%.*]] = load float, ptr [[ARRAYIDX_3]], align 4
91-
; CHECK-NEXT: [[MUL_3:%.*]] = fmul float [[X]], [[LOAD_3]]
92-
; CHECK-NEXT: store float [[MUL_3]], ptr [[ARRAYIDX_3]], align 4
93-
; CHECK-NEXT: [[IV_NEXT_3:%.*]] = add nuw nsw i64 [[IV]], 4
94-
; CHECK-NEXT: [[ARRAYIDX_4:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV_NEXT_3]]
95-
; CHECK-NEXT: [[LOAD_4:%.*]] = load float, ptr [[ARRAYIDX_4]], align 4
96-
; CHECK-NEXT: [[MUL_4:%.*]] = fmul float [[X]], [[LOAD_4]]
97-
; CHECK-NEXT: store float [[MUL_4]], ptr [[ARRAYIDX_4]], align 4
98-
; CHECK-NEXT: [[IV_NEXT_4:%.*]] = add nuw nsw i64 [[IV]], 5
99-
; CHECK-NEXT: [[ARRAYIDX_5:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV_NEXT_4]]
100-
; CHECK-NEXT: [[LOAD_5:%.*]] = load float, ptr [[ARRAYIDX_5]], align 4
101-
; CHECK-NEXT: [[MUL_5:%.*]] = fmul float [[X]], [[LOAD_5]]
102-
; CHECK-NEXT: store float [[MUL_5]], ptr [[ARRAYIDX_5]], align 4
103-
; CHECK-NEXT: [[IV_NEXT_5:%.*]] = add nuw nsw i64 [[IV]], 6
104-
; CHECK-NEXT: [[ARRAYIDX_6:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV_NEXT_5]]
105-
; CHECK-NEXT: [[LOAD_6:%.*]] = load float, ptr [[ARRAYIDX_6]], align 4
106-
; CHECK-NEXT: [[MUL_6:%.*]] = fmul float [[X]], [[LOAD_6]]
107-
; CHECK-NEXT: store float [[MUL_6]], ptr [[ARRAYIDX_6]], align 4
108-
; CHECK-NEXT: [[IV_NEXT_6:%.*]] = add nuw nsw i64 [[IV]], 7
109-
; CHECK-NEXT: [[ARRAYIDX_7:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV_NEXT_6]]
110-
; CHECK-NEXT: [[LOAD_7:%.*]] = load float, ptr [[ARRAYIDX_7]], align 4
111-
; CHECK-NEXT: [[MUL_7:%.*]] = fmul float [[X]], [[LOAD_7]]
112-
; CHECK-NEXT: store float [[MUL_7]], ptr [[ARRAYIDX_7]], align 4
113-
; CHECK-NEXT: [[IV_NEXT_7]] = add nuw nsw i64 [[IV]], 8
114-
; CHECK-NEXT: [[COMP_7:%.*]] = icmp eq i64 [[IV_NEXT_7]], 1024
115-
; CHECK-NEXT: br i1 [[COMP_7]], label %[[EXIT_LOOPEXIT:.*]], label %[[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
49+
; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
50+
; CHECK-NEXT: [[COMP:%.*]] = icmp eq i64 [[IV_NEXT]], 1024
51+
; CHECK-NEXT: br i1 [[COMP]], label %[[EXIT_LOOPEXIT:.*]], label %[[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
11652
; CHECK: [[EXIT_LOOPEXIT]]:
11753
; CHECK-NEXT: br label %[[EXIT]]
11854
; CHECK: [[EXIT]]:
@@ -142,8 +78,10 @@ exit:
14278
!4 = !{!"llvm.loop.isvectorized"}
14379
!5 = !{!"llvm.loop.unroll.count", i32 8}
14480
;.
145-
; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
146-
; CHECK: [[META1]] = !{!"llvm.loop.isvectorized"}
147-
; CHECK: [[META2]] = !{!"llvm.loop.unroll.disable"}
148-
; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META1]], [[META2]]}
81+
; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META4:![0-9]+]]}
82+
; CHECK: [[META1]] = distinct !{[[META1]], [[META2:![0-9]+]], [[META3:![0-9]+]]}
83+
; CHECK: [[META2]] = !{!"llvm.loop.isvectorized"}
84+
; CHECK: [[META3]] = !{!"llvm.loop.unroll.count", i32 8}
85+
; CHECK: [[META4]] = !{!"llvm.loop.unroll.runtime.disable"}
86+
; CHECK: [[LOOP5]] = distinct !{[[LOOP5]], [[META1]]}
14987
;.

0 commit comments

Comments
 (0)