8387073: C2: StoreNode::Ideal wrongly optimizes away scalar store before masked vector store by mhaessig · Pull Request #31695 · openjdk/jdk

mhaessig · 2026-06-26T14:11:52Z

StoreNode::Ideal() optimizes back-to-back stores to the same address by removing the earlier store if it is the same size or narrower than the later store when we have UseAVX >= 3 or UseSVE > 0. This optimization is incorrect if the later store is a masked vector store as it might not overwrite all bits of the preceding store.

This PR fixes this issue by plainly disabling the optimization for masked stores. We could conceivably enable the optimization for masked stores if we can prove that the mask overwrites all of the preceding store, but this fix is aimed at correctness before the 27 release. Thus, such optimizations are filed as follow-up RFEs.

Testing:

Github Actions
tier1,tier2,tier3 plus additional stress testing and an additional run on AVX-512 machines only.
tier4,tier5

I confirm that I make this contribution in accordance with the OpenJDK Interim AI Policy.

Progress

Change must not contain extraneous whitespace
Commit message must refer to an issue
Change must be properly reviewed (2 reviews required, with at least 1 Reviewer, 1 Author)

Issue

JDK-8387073: C2: StoreNode::Ideal wrongly optimizes away scalar store before masked vector store (Bug - P4)(⚠️ The fixVersion in this issue is [27] but the fixVersion in .jcheck/conf is 28, a new backport will be created when this pr is integrated.)

Reviewers

Christian Hagedorn (@chhagedorn - Reviewer)

Reviewing

Using git

Checkout this PR locally:
$ git fetch https://git.openjdk.org/jdk.git pull/31695/head:pull/31695
$ git checkout pull/31695

Update a local copy of the PR:
$ git checkout pull/31695
$ git pull https://git.openjdk.org/jdk.git pull/31695/head

Using Skara CLI tools

Checkout this PR locally:
$ git pr checkout 31695

View PR using the GUI difftool:
$ git pr show -t 31695

Using diff file

Download this PR as a diff file:
https://git.openjdk.org/jdk/pull/31695.diff

Using Webrev

Link to Webrev Comment

bridgekeeper · 2026-06-26T14:13:24Z

👋 Welcome back mhaessig! A progress list of the required criteria for merging this PR into master will be added to the body of your pull request. There are additional pull request commands available for use with this pull request.

openjdk · 2026-06-26T14:13:57Z

❗ This change is not yet ready to be integrated.
See the Progress checklist in the description for automated requirements.

openjdk · 2026-06-26T14:15:03Z

@mhaessig The following label will be automatically applied to this pull request:

hotspot-compiler

When this pull request is ready to be reviewed, an "RFR" email will be sent to the corresponding mailing list. If you would like to change these labels, use the /label pull request command.

openjdk · 2026-06-26T14:15:06Z

The total number of required reviews for this PR has been set to 2 based on the presence of this label: hotspot-compiler. This can be overridden with the /reviewers command.

mlbridge · 2026-06-26T14:19:34Z

Webrevs

01: Full - Incremental (28eab9bb)
00: Full (11644fcf)

chhagedorn

That looks reasonable to me. Have you already filed a follow-up RFE to revisit it in JDK 28?

eme64

Thanks for tackling this one @mhaessig 😊

eme64 · 2026-06-29T13:54:58Z

-          st->as_Store()->memory_size() <= this->memory_size()) {
+          st->as_Store()->memory_size() <= this->memory_size() &&
+          !this->is_StoreVectorMasked() && !this->is_StoreVectorScatterMasked() &&
+          !this->is_predicated_vector()) {


Is is_predicated_vector actually required here? Or could it be an assert?

eme64 · 2026-06-29T13:59:48Z

+    static boolean[] boolArr = new boolean[] {true, true};
+    static byte[] byteArr = new byte[] {(byte)42};
+
+    static boolean[] testBool() {
+        return Arrays.copyOf(boolArr, 4);
+    }
+
+    static byte[] testByte() {
+        // Should be zero-padded, but we seem to get non-zero values.
+        return Arrays.copyOf(byteArr, 4);
+    }


It would be nice if we could have a bit more testing in this area.
I was super surprised that we have not caught this earlier.

Idea: look for similar methods, and test them all for random arrays and lengths (constant and variable).

Also: maybe more types, including oop?

eme64 · 2026-06-29T14:01:47Z

+        final List<TemplateToken> tests = new ArrayList<>();
+        tests.addAll(CodeGenerationDataNameType.VECTOR_VECTOR_TYPES
+                        .stream()
+                        .filter(vec -> vec.byteSize() <= maxVecByteSize)


Why do you do this filtering? At least a comment would be nice ;)

What I would recommend: guard the IR rules instead, using MaxVectorSize. That would allow us to still do result verification for the other cases.

eme64 · 2026-06-29T14:03:03Z


      if (st->in(MemNode::Address)->eqv_uncast(address) &&
-          st->as_Store()->memory_size() <= this->memory_size()) {
+          st->as_Store()->memory_size() <= this->memory_size() &&


Optional: Can you also file an RFE to allow overlaps that don't require the address to be the same?
Neither of us probably has time, but I think we could probably find someone for it in the wider community ;)

eme64 · 2026-06-29T14:06:54Z

+                // For IR-verification we need more than two lanes. To prevent flakyness, we skip verification
+                // for vectors of longs because there are other StoreL nodes in this method.
+                if (vec.length <= 2 || vec.elementType.name().equals("long")) {
+                    return scope("");
+                }


Could we get the IR matching to be "sharper" by matching more parts of the dump string, somehow?

eme64 · 2026-06-29T14:08:44Z

+                // IR verification of vectors of int needs special handling due to an additional
+                // node in used for address computation when not using COH.
+                if (vec.elementType.name().equals("int")) {
+                    return scope(
+                        let("ptyIR", ptyIR),
+                        """
+                            @IR(counts = {IRNode.STORE_#ptyIR, "= 1",
+                                          IRNode.VECTOR_STORE_MASK, "= 1"},
+                                applyIf = {"UseCompactObjectHeaders", "true"},
+                                applyIfCPUFeatureOr = {"avx512", "true", "sve", "true"})
+                            @IR(counts = {IRNode.STORE_#ptyIR, "= 2",
+                                          IRNode.VECTOR_STORE_MASK, "= 1"},
+                                applyIf = {"UseCompactObjectHeaders", "false"},
+                                applyIfCPUFeatureOr = {"avx512", "true", "sve", "true"})
+                        """
+                    );
+                }


Can you explain the IR shapes a bit more here? What exactly makes the difference why we end up with one or two stores?

eme64 · 2026-06-29T14:11:03Z

+                        // Mask with one element set to false.
+                        VectorMask<#boxedTy> mask = VectorMask.fromLong(#species, -1 - (1 << #idx));


Having a completely random mask would also be nice, no? Or does that mess with something?

Maybe you need it to not constant fold away? I suppose we could also try with a local constant mask, and a mask that gets loaded from somewhere, i.e. a static field. To both cover possible constant folding and variable mask.

eme64 · 2026-06-29T14:13:20Z

+                        var v = #vecTy.broadcast(#species, broadcastVal);
+                        v.intoArray(a, 0, mask);


Do you know why this is needed? I did not spend enough time on it when I found the reproducer. But it seems a bit strange to me.

mhaessig added 2 commits June 26, 2026 09:00

Exclude masked stores

531adbb

Add test

6f2a9af

openjdk Bot added the hotspot-compiler hotspot-compiler-dev@openjdk.org label Jun 26, 2026

openjdk Bot added the rfr Pull request is ready for review label Jun 26, 2026

mhaessig marked this pull request as draft June 26, 2026 14:59

Add partial arraycopy inlining test

28eab9b

mhaessig force-pushed the JDK-8387073-masked branch from 11644fc to 28eab9b Compare June 26, 2026 15:00

openjdk Bot removed the rfr Pull request is ready for review label Jun 26, 2026

mhaessig marked this pull request as ready for review June 26, 2026 15:00

openjdk Bot added the rfr Pull request is ready for review label Jun 26, 2026

chhagedorn approved these changes Jun 29, 2026

View reviewed changes

eme64 reviewed Jun 29, 2026

View reviewed changes

		// Mask with one element set to false.
		VectorMask<#boxedTy> mask = VectorMask.fromLong(#species, -1 - (1 << #idx));

		var v = #vecTy.broadcast(#species, broadcastVal);
		v.intoArray(a, 0, mask);

Uh oh!

Conversation

mhaessig commented Jun 26, 2026 • edited by openjdk Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Progress

Issue

Reviewers

Reviewing

Uh oh!

bridgekeeper Bot commented Jun 26, 2026

Uh oh!

openjdk Bot commented Jun 26, 2026

Uh oh!

openjdk Bot commented Jun 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

openjdk Bot commented Jun 26, 2026

Uh oh!

mlbridge Bot commented Jun 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Webrevs

Uh oh!

chhagedorn left a comment

Choose a reason for hiding this comment

Uh oh!

eme64 left a comment

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!

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!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

3 participants

mhaessig commented Jun 26, 2026 •

edited by openjdk Bot

Loading

openjdk Bot commented Jun 26, 2026 •

edited

Loading

mlbridge Bot commented Jun 26, 2026 •

edited

Loading