[LICM] Drop nsw/nuw flags on affected instructions in hoistMulAddAssociation. #85486
+57
−16
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…ciation. Since we are introducing new multiplies earlier in the arithmetic, the nsw/nuw flags on later instructions are no longer accurate. Fixes llvm#85457.
|
@llvm/pr-subscribers-llvm-transforms Author: Craig Topper (topperc) ChangesSince we are introducing new multiplies earlier in the arithmetic, the nsw/nuw flags on later instructions are no longer accurate. Fixes #85457. Full diff: https://github.com/llvm/llvm-project/pull/85486.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/Scalar/LICM.cpp b/llvm/lib/Transforms/Scalar/LICM.cpp
index 40bc16f9b57516..e50413de46b1b4 100644
--- a/llvm/lib/Transforms/Scalar/LICM.cpp
+++ b/llvm/lib/Transforms/Scalar/LICM.cpp
@@ -2701,6 +2701,7 @@ static bool hoistMulAddAssociation(Instruction &I, Loop &L,
// First, we need to make sure we should do the transformation.
SmallVector<Use *> Changes;
+ SmallVector<BinaryOperator *> Adds;
SmallVector<BinaryOperator *> Worklist;
if (BinaryOperator *VariantBinOp = dyn_cast<BinaryOperator>(VariantOp))
Worklist.push_back(VariantBinOp);
@@ -2713,6 +2714,7 @@ static bool hoistMulAddAssociation(Instruction &I, Loop &L,
isa<BinaryOperator>(BO->getOperand(1))) {
Worklist.push_back(cast<BinaryOperator>(BO->getOperand(0)));
Worklist.push_back(cast<BinaryOperator>(BO->getOperand(1)));
+ Adds.push_back(BO);
continue;
}
if (!isReassociableOp(BO, Instruction::Mul, Instruction::FMul) ||
@@ -2735,6 +2737,12 @@ static bool hoistMulAddAssociation(Instruction &I, Loop &L,
if (Changes.empty())
return false;
+ // Drop the poison flags for any adds we looked through.
+ if (I.getType()->isIntOrIntVectorTy()) {
+ for (auto *Add : Adds)
+ Add->dropPoisonGeneratingFlags();
+ }
+
// We know we should do it so let's do the transformation.
auto *Preheader = L.getLoopPreheader();
assert(Preheader && "Loop is not in simplify form?");
@@ -2743,9 +2751,11 @@ static bool hoistMulAddAssociation(Instruction &I, Loop &L,
assert(L.isLoopInvariant(U->get()));
Instruction *Ins = cast<Instruction>(U->getUser());
Value *Mul;
- if (I.getType()->isIntOrIntVectorTy())
+ if (I.getType()->isIntOrIntVectorTy()) {
Mul = Builder.CreateMul(U->get(), Factor, "factor.op.mul");
- else
+ // Drop the poison flags on the original multiply.
+ Ins->dropPoisonGeneratingFlags();
+ } else
Mul = Builder.CreateFMulFMF(U->get(), Factor, Ins, "factor.op.fmul");
U->set(Mul);
}
diff --git a/llvm/test/Transforms/LICM/expr-reassociate-int.ll b/llvm/test/Transforms/LICM/expr-reassociate-int.ll
index 63548974fb3185..70288731e31c6d 100644
--- a/llvm/test/Transforms/LICM/expr-reassociate-int.ll
+++ b/llvm/test/Transforms/LICM/expr-reassociate-int.ll
@@ -23,7 +23,7 @@ define void @innermost_loop_1d_shouldhoist(i32 %i, i64 %d1, i64 %delta, ptr %cel
; CHECK-LABEL: define void @innermost_loop_1d_shouldhoist
; CHECK-SAME: (i32 [[I:%.*]], i64 [[D1:%.*]], i64 [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
; CHECK-NEXT: entry:
-; CHECK-NEXT: [[MUL_1:%.*]] = mul i64 [[DELTA]], [[D1]]
+; CHECK-NEXT: [[MUL_1:%.*]] = mul nuw nsw i64 [[DELTA]], [[D1]]
; CHECK-NEXT: br label [[FOR_COND:%.*]]
; CHECK: for.cond:
; CHECK-NEXT: [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
@@ -55,7 +55,7 @@ for.body:
%idxprom.j.1 = zext i32 %add.j.1 to i64
%arrayidx.j.1 = getelementptr inbounds i64, ptr %cells, i64 %idxprom.j.1
%cell.1 = load i64, ptr %arrayidx.j.1, align 8
- %mul.1 = mul i64 %delta, %d1
+ %mul.1 = mul nsw nuw i64 %delta, %d1
%mul.2 = mul i64 %mul.1, %cell.1
%idxprom.j = zext i32 %j to i64
%arrayidx.j = getelementptr inbounds i64, ptr %cells, i64 %idxprom.j
@@ -130,8 +130,8 @@ define void @innermost_loop_2d(i32 %i, i64 %d1, i64 %d2, i64 %delta, ptr %cells)
; CONSTRAINED-NEXT: [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
; CONSTRAINED-NEXT: [[ARRAYIDX_J:%.*]] = getelementptr inbounds i64, ptr [[CELLS]], i64 [[IDXPROM_J]]
; CONSTRAINED-NEXT: [[CELL_2:%.*]] = load i64, ptr [[ARRAYIDX_J]], align 8
-; CONSTRAINED-NEXT: [[MUL_2:%.*]] = mul i64 [[CELL_2]], [[D2]]
-; CONSTRAINED-NEXT: [[REASS_ADD:%.*]] = add i64 [[MUL_2]], [[MUL_1]]
+; CONSTRAINED-NEXT: [[MUL_2:%.*]] = mul nuw nsw i64 [[CELL_2]], [[D2]]
+; CONSTRAINED-NEXT: [[REASS_ADD:%.*]] = add nuw nsw i64 [[MUL_2]], [[MUL_1]]
; CONSTRAINED-NEXT: [[REASS_MUL:%.*]] = mul i64 [[REASS_ADD]], [[DELTA]]
; CONSTRAINED-NEXT: store i64 [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
; CONSTRAINED-NEXT: br label [[FOR_COND]]
@@ -155,8 +155,8 @@ for.body:
%idxprom.j = zext i32 %j to i64
%arrayidx.j = getelementptr inbounds i64, ptr %cells, i64 %idxprom.j
%cell.2 = load i64, ptr %arrayidx.j, align 8
- %mul.2 = mul i64 %cell.2, %d2
- %reass.add = add i64 %mul.2, %mul.1
+ %mul.2 = mul nsw nuw i64 %cell.2, %d2
+ %reass.add = add nsw nuw i64 %mul.2, %mul.1
%reass.mul = mul i64 %reass.add, %delta
store i64 %reass.mul, ptr %arrayidx.j, align 8
br label %for.cond
@@ -243,10 +243,10 @@ define void @innermost_loop_3d(i32 %i, i64 %d1, i64 %d2, i64 %d3, i64 %delta, pt
; CONSTRAINED-NEXT: [[IDXPROM_J_2:%.*]] = zext i32 [[ADD_J_2]] to i64
; CONSTRAINED-NEXT: [[ARRAYIDX_J_2:%.*]] = getelementptr inbounds i64, ptr [[CELLS]], i64 [[IDXPROM_J_2]]
; CONSTRAINED-NEXT: [[CELL_3:%.*]] = load i64, ptr [[ARRAYIDX_J_2]], align 8
-; CONSTRAINED-NEXT: [[MUL_3:%.*]] = mul i64 [[CELL_3]], [[D3]]
-; CONSTRAINED-NEXT: [[REASS_ADD:%.*]] = add i64 [[MUL_2]], [[MUL_1]]
-; CONSTRAINED-NEXT: [[REASS_ADD1:%.*]] = add i64 [[REASS_ADD]], [[MUL_3]]
-; CONSTRAINED-NEXT: [[REASS_MUL:%.*]] = mul i64 [[REASS_ADD1]], [[DELTA]]
+; CONSTRAINED-NEXT: [[MUL_3:%.*]] = mul nuw nsw i64 [[CELL_3]], [[D3]]
+; CONSTRAINED-NEXT: [[REASS_ADD:%.*]] = add nuw nsw i64 [[MUL_2]], [[MUL_1]]
+; CONSTRAINED-NEXT: [[REASS_ADD1:%.*]] = add nuw nsw i64 [[REASS_ADD]], [[MUL_3]]
+; CONSTRAINED-NEXT: [[REASS_MUL:%.*]] = mul nuw nsw i64 [[REASS_ADD1]], [[DELTA]]
; CONSTRAINED-NEXT: store i64 [[REASS_MUL]], ptr [[ARRAYIDX_J_2]], align 8
; CONSTRAINED-NEXT: br label [[FOR_COND]]
; CONSTRAINED: for.end:
@@ -274,10 +274,10 @@ for.body:
%idxprom.j.2 = zext i32 %add.j.2 to i64
%arrayidx.j.2 = getelementptr inbounds i64, ptr %cells, i64 %idxprom.j.2
%cell.3 = load i64, ptr %arrayidx.j.2, align 8
- %mul.3 = mul i64 %cell.3, %d3
- %reass.add = add i64 %mul.2, %mul.1
- %reass.add1 = add i64 %reass.add, %mul.3
- %reass.mul = mul i64 %reass.add1, %delta
+ %mul.3 = mul nsw nuw i64 %cell.3, %d3
+ %reass.add = add nsw nuw i64 %mul.2, %mul.1
+ %reass.add1 = add nsw nuw i64 %reass.add, %mul.3
+ %reass.mul = mul nsw nuw i64 %reass.add1, %delta
store i64 %reass.mul, ptr %arrayidx.j.2, align 8
br label %for.cond
@@ -362,3 +362,34 @@ for.body:
for.end:
ret void
}
+
+; Make sure we drop poison flags on the mul in the loop.
+define i32 @pr85457(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @pr85457
+; CHECK-SAME: (i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[FACTOR_OP_MUL:%.*]] = mul i32 [[X]], [[Y]]
+; CHECK-NEXT: br label [[LOOP:%.*]]
+; CHECK: loop:
+; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 1, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
+; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i32 [[IV]], 1
+; CHECK-NEXT: [[MUL0:%.*]] = mul i32 [[FACTOR_OP_MUL]], [[IV]]
+; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[MUL0]], 1
+; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[LOOP]]
+; CHECK: exit:
+; CHECK-NEXT: ret i32 0
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i32 [ 1, %entry ], [ %iv.next, %loop ]
+ %iv.next = add nuw nsw i32 %iv, 1
+ %mul0 = mul nuw nsw i32 %x, %iv
+ %mul1 = mul nuw i32 %mul0, %y
+ %cmp = icmp slt i32 %mul1, 1
+ br i1 %cmp, label %exit, label %loop
+
+exit:
+ ret i32 0
+}
|
nikic
reviewed
Mar 18, 2024
| // Drop the poison flags for any adds we looked through. | ||
| if (I.getType()->isIntOrIntVectorTy()) { | ||
| for (auto *Add : Adds) | ||
| Add->dropPoisonGeneratingFlags(); |
There was a problem hiding this comment.
I think the flag drop on the adds does not have test coverage?
There was a problem hiding this comment.
Aren't they covered by the NOT_CONSTRAINED-NEXT lines? On the test case where I added nsw nuw to adds?
There was a problem hiding this comment.
Sorry, I got confused by the two sets of check lines.
|
This should also fix #85448 |
chencha3
pushed a commit
to chencha3/llvm-project
that referenced
this pull request
Mar 23, 2024
…ciation. (llvm#85486) Since we are introducing new multiplies earlier in the arithmetic, the nsw/nuw flags on later instructions are no longer accurate. Fixes llvm#85457.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Since we are introducing new multiplies earlier in the arithmetic, the nsw/nuw flags on later instructions are no longer accurate.
Fixes #85457.