[InstCombine] Fix a cycle when folding fneg(select) with scalable vector types #112465

ssijaric-nv · 2024-10-16T03:25:25Z

The two folding operations are causing a cycle for the following case with
scalable vector types:

define <vscale x 2 x double> @test_fneg_select_abs(<vscale x 2 x i1> %cond, <vscale x 2 x double> %b) {
%1 = select <vscale x 2 x i1> %cond, <vscale x 2 x double> zeroinitializer, <vscale x 2 x double> %b
%2 = fneg fast <vscale x 2 x double> %1
ret <vscale x 2 x double> %2
}

fold fneg: -(Cond ? C : Y) -> Cond ? -C : -Y
fold select: (Cond ? -X : -Y) -> -(Cond ? X : Y)
results in the following since '<vscale x 2 x double> zeroinitializer' passes
the check for the immediate constant:

%.neg = fneg fast <vscale x 2 x double> zeroinitializer
%b.neg = fneg fast <vscale x 2 x double> %b
%1 = select fast <vscale x 2 x i1> %cond, <vscale x 2 x double> %.neg, <vscale x 2 x double> %b.neg

and so we end up going back and forth between 1) and 2).

Attempt to fold scalable vector constants, so that we end up with a splat instead:

define <vscale x 2 x double> @test_fneg_select_abs(<vscale x 2 x i1> %cond, <vscale x 2 x double> %b) {
%b.neg = fneg fast <vscale x 2 x double> %b
%1 = select fast <vscale x 2 x i1> %cond, <vscale x 2 x double> shufflevector (<vscale x 2 x double> insertelement (<vscale x 2 x double> poison, double -0.000000e+00, i64 0), <vscale x 2 x double> poison, <vscale x 2 x i32> zeroinitializer), <vscale x 2 x double> %b.neg
ret <vscale x 2 x double> %1
}

github-actions · 2024-10-16T03:25:45Z

Thank you for submitting a Pull Request (PR) to the LLVM Project!

This PR will be automatically labeled and the relevant teams will be notified.

If you wish to, you can add reviewers by using the "Reviewers" section on this page.

If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using @ followed by their GitHub username.

If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers.

If you have further questions, they may be answered by the LLVM GitHub User Guide.

You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums.

llvmbot · 2024-10-16T03:26:19Z

@llvm/pr-subscribers-llvm-ir

@llvm/pr-subscribers-llvm-transforms

Author: None (ssijaric-nv)

Changes

The two folding operations are causing a cycle for the following case with
scalable vector types:

define <vscale x 2 x double> @test_fneg_select_abs(<vscale x 2 x i1> %cond, <vscale x 2 x double> %b) {
%1 = select <vscale x 2 x i1> %cond, <vscale x 2 x double> zeroinitializer, <vscale x 2 x double> %b
%2 = fneg fast <vscale x 2 x double> %1
ret <vscale x 2 x double> %2
}

fold fneg: -(Cond ? C : Y) -> Cond ? -C : -Y
fold select: (Cond ? -X : -Y) -> -(Cond ? X : Y)
results in the following since '<vscale x 2 x double> zeroinitializer' passes
the check for the immediate constant:

%.neg = fneg fast <vscale x 2 x double> zeroinitializer
%b.neg = fneg fast <vscale x 2 x double> %b
%1 = select fast <vscale x 2 x i1> %cond, <vscale x 2 x double> %.neg, <vscale x 2 x double> %b.neg

and so we end up going back and forth between 1) and 2).

Don't perform 1) if we are dealing with scalable vector constants as there is no
immediate representation for a negative scalable zeroinitializer.

Full diff: https://github.com/llvm/llvm-project/pull/112465.diff

2 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp (+2-1)
(modified) llvm/test/Transforms/InstCombine/fneg.ll (+12)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index 21588aca512758..645fad14c1999c 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -2878,7 +2878,8 @@ Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
 
     // -(Cond ? X : C) --> Cond ? -X : -C
     // -(Cond ? C : Y) --> Cond ? -C : -Y
-    if (match(X, m_ImmConstant()) || match(Y, m_ImmConstant())) {
+    if ((match(X, m_ImmConstant()) && !isa<ScalableVectorType>(X->getType())) ||
+        (match(Y, m_ImmConstant()) && !isa<ScalableVectorType>(Y->getType()))) {
       Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
       Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
       SelectInst *NewSel = SelectInst::Create(Cond, NegX, NegY);
diff --git a/llvm/test/Transforms/InstCombine/fneg.ll b/llvm/test/Transforms/InstCombine/fneg.ll
index 3c4088832feaaa..643c8fd8d8a742 100644
--- a/llvm/test/Transforms/InstCombine/fneg.ll
+++ b/llvm/test/Transforms/InstCombine/fneg.ll
@@ -1109,4 +1109,16 @@ define float @test_fneg_select_maxnum(float %x) {
   ret float %neg
 }
 
+; Check that there's no infinite loop.
+define <vscale x 2 x double> @test_fneg_select_svec(<vscale x 2 x i1> %cond, <vscale x 2 x double> %b) {
+; CHECK-LABEL: @test_fneg_select_svec(
+; CHECK-NEXT:    [[TMP1:%.*]] = select <vscale x 2 x i1> [[COND:%.*]], <vscale x 2 x double> zeroinitializer, <vscale x 2 x double> [[B:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = fneg fast <vscale x 2 x double> [[TMP1]]
+; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP2]]
+;
+  %1 = select <vscale x 2 x i1> %cond, <vscale x 2 x double> zeroinitializer, <vscale x 2 x double> %b
+  %2 = fneg fast <vscale x 2 x double> %1
+  ret <vscale x 2 x double> %2
+}
+
 !0 = !{}

arsenm · 2024-10-16T04:19:36Z

if we are dealing with scalable vector constants as there is no
immediate representation for a negative scalable zeroinitializer.

I thought there was a splat constant for these, this would just be a splat of -0

arsenm · 2024-10-16T04:19:01Z

llvm/test/Transforms/InstCombine/fneg.ll

+  %2 = fneg fast <vscale x 2 x double> %1
+  ret <vscale x 2 x double> %2
+}
+


Test with zeroinitializer in both positions

dtcxzyw · 2024-10-16T04:19:56Z

llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

@@ -2878,7 +2878,8 @@ Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {

    // -(Cond ? X : C) --> Cond ? -X : -C
    // -(Cond ? C : Y) --> Cond ? -C : -Y
-    if (match(X, m_ImmConstant()) || match(Y, m_ImmConstant())) {
+    if ((match(X, m_ImmConstant()) && !isa<ScalableVectorType>(X->getType())) ||


We can use ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL) instead of using IRBuilder to fix this issue.

if (match(X, m_ImmConstant(XC)) || match(Y, m_ImmConstant(YC))) {
Value *NegX = ConstantFoldUnaryOpOperand(Instruction::FNeg, XC, DL);
Value *NegY = ConstantFoldUnaryOpOperand(Instruction::FNeg, YC, DL);
if (NegX || NegY) {
if (!NegX) NegX = Builder.CreateFNeg...;
if (!NegY) NegY = Builder.CreateFNeg...;
SelectInst *NewSel = SelectInst::Create(Cond, NegX, NegY);
propagateSelectFMF(NewSel, /CommonOperand=/true);
return NewSel;
}
}

Thanks for having a look, Yingwei. I'll test and upload the revised version.

The IRBuilder should do this like in other cases?

ssijaric-nv · 2024-10-16T07:06:39Z

if we are dealing with scalable vector constants as there is no
immediate representation for a negative scalable zeroinitializer.

I thought there was a splat constant for these, this would just be a splat of -0

Thanks for having a look, Matt. Would a better approach be to generate the following, instead of giving up?

define <vscale x 2 x double> @test_fneg_select_svec(<vscale x 2 x i1> %cond, <vscale x 2 x double> %b) {
%1 = fneg fast <vscale x 2 x double> %b
%2 = select <vscale x 2 x i1> %cond, <vscale x 2 x double> splat(double -0.0), <vscale x 2 x double> %1
ret <vscale x 2 x double> %2
}

nikic · 2024-10-16T10:00:02Z

if we are dealing with scalable vector constants as there is no
immediate representation for a negative scalable zeroinitializer.

I thought there was a splat constant for these, this would just be a splat of -0

Thanks for having a look, Matt. Would a better approach be to generate the following, instead of giving up?

define <vscale x 2 x double> @test_fneg_select_svec(<vscale x 2 x i1> %cond, <vscale x 2 x double> %b) { %1 = fneg fast <vscale x 2 x double> %b %2 = select <vscale x 2 x i1> %cond, <vscale x 2 x double> splat(double -0.0), <vscale x 2 x double> %1 ret <vscale x 2 x double> %2 }

Yes. It looks like ConstantFoldUnaryInstruction is currently missing splat support for scalable vectors. Adding that should resolve the issue.

arsenm · 2024-10-17T06:44:27Z

llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

+      if (XC)
+        NegX = ConstantFoldUnaryOpOperand(Instruction::FNeg, XC, DL);
+      if (YC)
+        NegY = ConstantFoldUnaryOpOperand(Instruction::FNeg, YC, DL);


Generally the attempt to constant fold should be hidden inside the IRBuilder Create* functions

Right, just the changes to ConstantFold should be sufficient, after that there should not be a need to touch this InstCombine code anymore...

nikic

LGTM

ssijaric-nv · 2024-10-24T18:21:24Z

Thank you, all. Can someone please merge this change, as I don't have write access? I'll rebase and push again, to get a clean build before merging.

I ran the test-suite locally with -O3 on AArch64 and x86 machines:

AArch64:
Total Discovered Tests: 3425
Passed: 3425 (100.00%)

With -mcpu=neoverse-v2
Total Discovered Tests: 3425
Passed: 3425 (100.00%)

x86:
Total Discovered Tests: 2441
Passed: 2441 (100.00%)

ninja check-llvm passed as well.

…tor types The two folding operations are causing a cycle for the following case with scalable vector types: define <vscale x 2 x double> @test_fneg_select_abs(<vscale x 2 x i1> %cond, <vscale x 2 x double> %b) { %1 = select <vscale x 2 x i1> %cond, <vscale x 2 x double> zeroinitializer, <vscale x 2 x double> %b %2 = fneg fast <vscale x 2 x double> %1 ret <vscale x 2 x double> %2 } 1) fold fneg: -(Cond ? C : Y) -> Cond ? -C : -Y 2) fold select: (Cond ? -X : -Y) -> -(Cond ? X : Y) 1) results in the following since '<vscale x 2 x double> zeroinitializer' passes the check for the immediate constant: %.neg = fneg fast <vscale x 2 x double> zeroinitializer %b.neg = fneg fast <vscale x 2 x double> %b %1 = select fast <vscale x 2 x i1> %cond, <vscale x 2 x double> %.neg, <vscale x 2 x double> %b.neg and so we end up going back and forth between 1) and 2). Don't perform 1) if we are dealing with scalable vector constants as there is no immediate representation for a negative scalable zeroinitializer.

…hange

github-actions · 2024-10-25T17:48:01Z

@ssijaric-nv Congratulations on having your first Pull Request (PR) merged into the LLVM Project!

Your changes will be combined with recent changes from other authors, then tested by our build bots. If there is a problem with a build, you may receive a report in an email or a comment on this PR.

Please check whether problems have been caused by your change specifically, as the builds can include changes from many authors. It is not uncommon for your change to be included in a build that fails due to someone else's changes, or infrastructure issues.

How to do this, and the rest of the post-merge process, is covered in detail here.

If your change does cause a problem, it may be reverted, or you can revert it yourself. This is a normal part of LLVM development. You can fix your changes and open a new PR to merge them again.

If you don't get any reports, no action is required from you. Your changes are working as expected, well done!

…tor types (llvm#112465) The two folding operations are causing a cycle for the following case with scalable vector types: define <vscale x 2 x double> @test_fneg_select_abs(<vscale x 2 x i1> %cond, <vscale x 2 x double> %b) { %1 = select <vscale x 2 x i1> %cond, <vscale x 2 x double> zeroinitializer, <vscale x 2 x double> %b %2 = fneg fast <vscale x 2 x double> %1 ret <vscale x 2 x double> %2 } 1) fold fneg: -(Cond ? C : Y) -> Cond ? -C : -Y 2) fold select: (Cond ? -X : -Y) -> -(Cond ? X : Y) 1) results in the following since '<vscale x 2 x double> zeroinitializer' passes the check for the immediate constant: %.neg = fneg fast <vscale x 2 x double> zeroinitializer %b.neg = fneg fast <vscale x 2 x double> %b %1 = select fast <vscale x 2 x i1> %cond, <vscale x 2 x double> %.neg, <vscale x 2 x double> %b.neg and so we end up going back and forth between 1) and 2). Attempt to fold scalable vector constants, so that we end up with a splat instead: define <vscale x 2 x double> @test_fneg_select_abs(<vscale x 2 x i1> %cond, <vscale x 2 x double> %b) { %b.neg = fneg fast <vscale x 2 x double> %b %1 = select fast <vscale x 2 x i1> %cond, <vscale x 2 x double> shufflevector (<vscale x 2 x double> insertelement (<vscale x 2 x double> poison, double -0.000000e+00, i64 0), <vscale x 2 x double> poison, <vscale x 2 x i32> zeroinitializer), <vscale x 2 x double> %b.neg ret <vscale x 2 x double> %1 }

ssijaric-nv requested a review from nikic as a code owner October 16, 2024 03:25

llvmbot added the llvm:transforms label Oct 16, 2024

goldsteinn requested review from arsenm and dtcxzyw October 16, 2024 04:14

arsenm reviewed Oct 16, 2024

View reviewed changes

dtcxzyw reviewed Oct 16, 2024

View reviewed changes

llvmbot added the llvm:ir label Oct 17, 2024

arsenm reviewed Oct 17, 2024

View reviewed changes

nikic approved these changes Oct 17, 2024

View reviewed changes

ssijaric-nv added 3 commits October 24, 2024 12:23

Handle folding of constant splats for scalable vector types

e871975

Only the constant folding change is needed; revert the fneg folding c…

f2e4e8b

…hange

ssijaric-nv force-pushed the fneg_select_sve_fold_imm branch from a0944b5 to f2e4e8b Compare October 24, 2024 21:06

arsenm merged commit 14db069 into llvm:main Oct 25, 2024
8 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[InstCombine] Fix a cycle when folding fneg(select) with scalable vector types #112465

[InstCombine] Fix a cycle when folding fneg(select) with scalable vector types #112465

Uh oh!

ssijaric-nv commented Oct 16, 2024 •

edited

Loading

Uh oh!

github-actions bot commented Oct 16, 2024

Uh oh!

llvmbot commented Oct 16, 2024 •

edited

Loading

Uh oh!

arsenm commented Oct 16, 2024

Uh oh!

arsenm Oct 16, 2024

Uh oh!

dtcxzyw Oct 16, 2024

Uh oh!

ssijaric-nv Oct 16, 2024

Uh oh!

arsenm Oct 16, 2024

Uh oh!

ssijaric-nv commented Oct 16, 2024

Uh oh!

nikic commented Oct 16, 2024

Uh oh!

arsenm Oct 17, 2024

Uh oh!

nikic Oct 17, 2024

Uh oh!

nikic left a comment

Uh oh!

ssijaric-nv commented Oct 24, 2024 •

edited

Loading

Uh oh!

Uh oh!

github-actions bot commented Oct 25, 2024

Uh oh!

Uh oh!

[InstCombine] Fix a cycle when folding fneg(select) with scalable vector types #112465

[InstCombine] Fix a cycle when folding fneg(select) with scalable vector types #112465

Uh oh!

Conversation

ssijaric-nv commented Oct 16, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Oct 16, 2024

Uh oh!

llvmbot commented Oct 16, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

arsenm commented Oct 16, 2024

Uh oh!

arsenm Oct 16, 2024

Choose a reason for hiding this comment

Uh oh!

dtcxzyw Oct 16, 2024

Choose a reason for hiding this comment

Uh oh!

ssijaric-nv Oct 16, 2024

Choose a reason for hiding this comment

Uh oh!

arsenm Oct 16, 2024

Choose a reason for hiding this comment

Uh oh!

ssijaric-nv commented Oct 16, 2024

Uh oh!

nikic commented Oct 16, 2024

Uh oh!

arsenm Oct 17, 2024

Choose a reason for hiding this comment

Uh oh!

nikic Oct 17, 2024

Choose a reason for hiding this comment

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

ssijaric-nv commented Oct 24, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

github-actions bot commented Oct 25, 2024

Uh oh!

Uh oh!

ssijaric-nv commented Oct 16, 2024 •

edited

Loading

llvmbot commented Oct 16, 2024 •

edited

Loading

ssijaric-nv commented Oct 24, 2024 •

edited

Loading