[InstCombine] Pull extract through broadcast #143380

agorenstein-nvidia · 2025-06-09T13:56:22Z

The change adds a new instcombine pattern, and associated test, for patterns like this:

  %3 = shufflevector <2 x float> %1, <2 x float> poison, <4 x i32> zeroinitializer
  %4 = extractelement <4 x float> %3, i64 %idx

The shufflevector has a splat, or broadcast, mask, so the extractelement simply must be the first element of %1, so we transform this to

  %2 = extractelement <2 x float> %1, i64 0

github-actions · 2025-06-09T13:56:45Z

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 · 2025-06-09T13:57:48Z

@llvm/pr-subscribers-llvm-transforms

Author: None (agorenstein-nvidia)

Changes

The change adds a new instcombine pattern, and associated test, for patterns like this:

  %3 = shufflevector &lt;2 x float&gt; %1, &lt;2 x float&gt; poison, &lt;4 x i32&gt; zeroinitializer
  %4 = extractelement &lt;4 x float&gt; %3, i64 %idx

The shufflevector has an all-zeros mask, so the extractelement simply must be the first element of %1, so we transform this to

  %2 = extractelement &lt;2 x float&gt; %1, i64 0

Full diff: https://github.com/llvm/llvm-project/pull/143380.diff

2 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp (+6)
(added) llvm/test/Transforms/InstCombine/vec_extract_through_broadcast.ll (+12)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index f946c3856948b..f2fb26c3ffae8 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -542,6 +542,12 @@ Instruction *InstCombinerImpl::visitExtractElementInst(ExtractElementInst &EI) {
         }
       }
     } else if (auto *SVI = dyn_cast<ShuffleVectorInst>(I)) {
+      // extractelt (shufflevector %v1, %v2, zeroinitializer) ->
+      // extractelt %v1, 0
+      if (isa<FixedVectorType>(SVI->getType()))
+        if (all_of(SVI->getShuffleMask(), [](int Elt) { return Elt == 0; }))
+          return ExtractElementInst::Create(SVI->getOperand(0), Builder.getInt64(0));
+
       // If this is extracting an element from a shufflevector, figure out where
       // it came from and extract from the appropriate input element instead.
       // Restrict the following transformation to fixed-length vector.
diff --git a/llvm/test/Transforms/InstCombine/vec_extract_through_broadcast.ll b/llvm/test/Transforms/InstCombine/vec_extract_through_broadcast.ll
new file mode 100644
index 0000000000000..ec80fbf01a3c9
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/vec_extract_through_broadcast.ll
@@ -0,0 +1,12 @@
+; RUN: opt -passes=instcombine -S < %s | FileCheck %s
+
+define float @extract_from_zero_init_shuffle(<2 x float> %1, i64 %idx) {
+; CHECK-LABEL: @extract_from_zero_init_shuffle(
+; CHECK-NEXT:    [[TMP1:%.*]] = extractelement <2 x float> [[W:%.*]], i64 0
+; CHECK-NEXT:    ret float [[TMP1]]
+;
+  %3 = shufflevector <2 x float> %1, <2 x float> poison, <4 x i32> zeroinitializer
+  %4 = extractelement <4 x float> %3, i64 %idx
+  ret float %4
+}
+

agorenstein-nvidia · 2025-06-09T14:01:31Z

(Per github-actions bot, it seems I don't have the permissions to add reviewers directly. So...) Adding comment to request also my colleague @Prince781 to review this work. Thanks.

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp

nikic

This seems like a generally profitable vector transform to me, but adding some more people to confirm.

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp

agorenstein-nvidia · 2025-06-09T16:06:17Z

Thanks for the feedback on how to take this further, @nikic and @RKSimon ! I've just pushed some changes with the intention of realizing those extensions. I've added additional tests, including a couple negative cases (inspired by RKSimon's mention of undef).

I'll add that we can assume idx is valid (or at least, not-known-to-be-invalid) by earlier checks in the same instcombine routine, so this shouldn't be over-aggressive in removing poison opportunities when replacing idx with mask[0] (so-to-speak).

github-actions · 2025-06-09T16:24:55Z

✅ With the latest revision this PR passed the undef deprecator.

…sing meaningful global instead of magic constant for shufflemask poison-indicator

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp

llvm/test/Transforms/InstCombine/vec_extract_through_broadcast.ll

…ack removing negative tests

Prince781

LGTM.

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp

agorenstein-nvidia · 2025-06-10T15:02:37Z

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp

+        // getSplatIndex doesn't distinguish between the all-poison splat and
+        // a non-splat mask. However, if Index is -1, we still want to propagate
+        // that poison value.
+        int SrcIdx = -2;


I'm not entirely happy with this. I've also considered either a tiny helper function like bool test(int* outParam), or carrying along bool ValidSrcIdx, or rewriting getSplatIndex to be able to meaningfully return the all-poison splat indicator, and this seemed the best option.

use std::optional<int> SrcIdx instead?

@agorenstein-nvidia did std::optional not help simplify this logic?

Thanks for the suggestion and ping! I was out from Thursday until yesterday [edit: inclusive]; this remains top of my "priority queue" and I'm integrating your suggestions now. Provisionally it looks like that exactly captures the intent, and clearly I hadn't considered this approach before. I'll finish putting this change in, and your test-improvement-suggestions (and run the tests/linter!) and push ASAP.

This has been done in the latest push, please take a look when you get the chance. It passes my local versions of git-clang-format and the lit tests, so hopefully that carries through to the CI here.

agorenstein-nvidia · 2025-06-10T15:14:50Z

llvm/test/Transforms/InstCombine/vscale_extractelement-inseltpoison.ll

 }

 define i32 @extractelement_shuffle_invalid_index(i32 %v) {
 ; CHECK-LABEL: @extractelement_shuffle_invalid_index(


Here we optimize, but also in doing so miss the poison "opportunity" (in that the illegal -1 index-usage is removed). So I believe this is legal, but perhaps not optimal. However, I'm a bit confused: it is the case that earlier in our transformation we explicitly look for invalid indices in extractelements:

// InstSimplify should handle cases where the index is invalid. // For fixed-length vector, it's invalid to extract out-of-range element.

Without tracing through, it looks like this may "just" be an order-of-transformations issue. I'm inclined to preserve the behavior. I've simply changed the test to reflect the new output; is there a preferred way of updating/changing these tests in this sort of situation?

agorenstein-nvidia · 2025-06-10T15:16:40Z

llvm/test/Transforms/InstCombine/vscale_extractelement-inseltpoison.ll

-; CHECK-NEXT:    [[IN:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[V:%.*]], i64 0
-; CHECK-NEXT:    [[SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[IN]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
-; CHECK-NEXT:    [[R:%.*]] = extractelement <vscale x 4 x i32> [[SPLAT]], i64 4
-; CHECK-NEXT:    ret i32 [[R]]


This is analogous to the other comment; it may be that the extractelement ... i32 4 is out of bounds (when vscale=1, so to speak). However, we are able to know that any valid index into %splat must be %v, so we just return %v. So, similar questions as the other comment.

might be useful to add a comment above the test explaining the reasoning,similar to your comment here

Did so! Thanks.

agorenstein-nvidia · 2025-06-16T13:34:52Z

Ping

github-actions · 2025-06-16T15:32:23Z

✅ With the latest revision this PR passed the C/C++ code formatter.

agorenstein-nvidia · 2025-06-16T20:29:11Z

Ping

I noticed shortly after this ping someone (?) triggered some test actions, which revealed clang-format errors. I fixed them, and haven't seen any other issues. Now I believe I've addressed all outstanding comments/issues, and while I have questions about some code, they're of the sort "can this code be written better?" rather than change-of-behavior. For these reasons I believe this change is all-set, pending approvals---is there anything more I can do? Happy to take on as much as possible, please let me know.

RKSimon · 2025-06-24T13:35:00Z

llvm/test/Transforms/InstCombine/vec_extract_through_broadcast.ll

+; CHECK-NEXT:    [[TMP1:%.*]] = extractelement <2 x float> [[W:%.*]], i64 0
+; CHECK-NEXT:    ret float [[TMP1]]
+;
+  %3 = shufflevector <2 x float> %1, <2 x float> poison, <4 x i32> zeroinitializer


(style) avoid using numbers for variable names (%vec, %shuffle and %extract would be better)

Done in latest push, please take a look. Thanks.

RKSimon

LGTM @nikic any more comments? whether we're happy with out of bounds extraction from a known splat (possibly with undef shuffle indices) seems to be the only outstanding issue

define i4 @src(i4 %x) {
  %v = insertelement <4 x i4> poison, i4 %x, i32 0
  %s = shufflevector <4 x i4> %v, <4 x i4> poison, 0, 0, 4294967295, 0
  %e = extractelement <4 x i4> %s, i32 4294967295
  ret i4 %e
}
=>
define i4 @tgt(i4 %x) {
  ret i4 %x
}
Transformation seems to be correct!

RKSimon · 2025-06-25T08:47:58Z

llvm/test/Transforms/InstCombine/vscale_extractelement-inseltpoison.ll

-; CHECK-NEXT:    [[IN:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[V:%.*]], i64 0
-; CHECK-NEXT:    [[SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[IN]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
-; CHECK-NEXT:    [[R:%.*]] = extractelement <vscale x 4 x i32> [[SPLAT]], i64 4
-; CHECK-NEXT:    ret i32 [[R]]


might be useful to add a comment above the test explaining the reasoning,similar to your comment here

nikic

LGTM

github-actions · 2025-07-04T16:20:21Z

@agorenstein-nvidia 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!

agorenstein-nvidia added 2 commits June 9, 2025 09:30

Initial changes

34c1222

Merge branch 'main' into instcombine-extract-through-broadcast

31cb512

agorenstein-nvidia requested a review from nikic as a code owner June 9, 2025 13:56

llvmbot added llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes llvm:transforms labels Jun 9, 2025

nikic reviewed Jun 9, 2025

View reviewed changes

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp Outdated Show resolved Hide resolved

nikic requested review from Prince781, RKSimon and davemgreen June 9, 2025 14:03

nikic reviewed Jun 9, 2025

View reviewed changes

RKSimon reviewed Jun 9, 2025

View reviewed changes

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp Outdated Show resolved Hide resolved

agorenstein-nvidia added 2 commits June 9, 2025 11:57

PR feedback 1

e315c89

clang-format

274cec9

agorenstein-nvidia added 2 commits June 9, 2025 12:09

Fixing edit damage in test file

4fda162

Removed unneeded check

5d9e4d9

Improve undef/poison usage in test (per PR automation failure), and u…

a1aec2e

…sing meaningful global instead of magic constant for shufflemask poison-indicator

nikic reviewed Jun 9, 2025

View reviewed changes

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp Outdated Show resolved Hide resolved

llvm/test/Transforms/InstCombine/vec_extract_through_broadcast.ll Show resolved Hide resolved

Fixed varname convention per PR feedback; on reflection from PR feedb…

70d3c01

…ack removing negative tests

Prince781 approved these changes Jun 9, 2025

View reviewed changes

nikic reviewed Jun 9, 2025

View reviewed changes

llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp Outdated Show resolved Hide resolved

Restore old tests, reframe optimization as extension of existing one.

52f5c4b

agorenstein-nvidia commented Jun 10, 2025

View reviewed changes

Prince781 self-requested a review June 10, 2025 15:58

agorenstein-nvidia requested review from RKSimon and nikic June 10, 2025 20:16

fix formatting errors, sorry

e21bde6

farzonl mentioned this pull request Jun 20, 2025

[DirectX] add support for i64 buffer load/stores #145047

Merged

RKSimon reviewed Jun 24, 2025

View reviewed changes

agorenstein-nvidia added 2 commits June 24, 2025 11:48

Improvement via optional from RKSimon's suggestion

09d5e4c

Improved tests with semantic names per RKSimon's PR feedback

ce809f6

RKSimon approved these changes Jun 25, 2025

View reviewed changes

Improved comments to explicitly-named tests per PR suggestion

21b8638

nikic approved these changes Jul 4, 2025

View reviewed changes

nikic merged commit b0473c5 into llvm:main Jul 4, 2025
7 checks passed

[InstCombine] Pull extract through broadcast #143380

[InstCombine] Pull extract through broadcast #143380

Uh oh!

Conversation

agorenstein-nvidia commented Jun 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Jun 9, 2025

Uh oh!

llvmbot commented Jun 9, 2025

Uh oh!

agorenstein-nvidia commented Jun 9, 2025

Uh oh!

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

agorenstein-nvidia commented Jun 9, 2025

Uh oh!

github-actions bot commented Jun 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Prince781 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

agorenstein-nvidia Jun 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

agorenstein-nvidia commented Jun 16, 2025

Uh oh!

github-actions bot commented Jun 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

agorenstein-nvidia commented Jun 16, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

RKSimon left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions bot commented Jul 4, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

agorenstein-nvidia commented Jun 9, 2025 •

edited

Loading

github-actions bot commented Jun 9, 2025 •

edited

Loading

agorenstein-nvidia Jun 24, 2025 •

edited

Loading

github-actions bot commented Jun 16, 2025 •

edited

Loading