8350835: C2 SuperWord: assert/wrong result when using Float.float16ToFloat with byte instead of short input #23939

sviswa7 · 2025-03-07T01:56:49Z

Float.float16ToFloat generates wrong vectorized code in product build and asserts in fastdebug/debug when argument is of type byte, int, or long array. The short term solution is to not auto vectorize in these cases.

Review comments are welcome.

Best Regards,
Sandhya

Progress

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

Issue

JDK-8350835: C2 SuperWord: assert/wrong result when using Float.float16ToFloat with byte instead of short input (Bug - P2)

Reviewers

Vladimir Kozlov (@vnkozlov - Reviewer) Review applies to 4ebb47a7
Emanuel Peter (@eme64 - Reviewer)

Reviewing

Using git

Checkout this PR locally:
$ git fetch https://git.openjdk.org/jdk.git pull/23939/head:pull/23939
$ git checkout pull/23939

Update a local copy of the PR:
$ git checkout pull/23939
$ git pull https://git.openjdk.org/jdk.git pull/23939/head

Using Skara CLI tools

Checkout this PR locally:
$ git pr checkout 23939

View PR using the GUI difftool:
$ git pr show -t 23939

Using diff file

Download this PR as a diff file:
https://git.openjdk.org/jdk/pull/23939.diff

Using Webrev

Link to Webrev Comment

…Float with byte instead of short input

bridgekeeper · 2025-03-07T01:57:44Z

👋 Welcome back sviswanathan! 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 · 2025-03-07T01:58:06Z

@sviswa7 This change now passes all automated pre-integration checks.

ℹ️ This project also has non-automated pre-integration requirements. Please see the file CONTRIBUTING.md for details.

After integration, the commit message for the final commit will be:

8350835: C2 SuperWord: assert/wrong result when using Float.float16ToFloat with byte instead of short input

Reviewed-by: epeter, kvn

You can use pull request commands such as /summary, /contributor and /issue to adjust it as needed.

At the time when this comment was updated there had been 120 new commits pushed to the master branch:

4c6a523: 8352096: Test jdk/jfr/event/profiling/TestFullStackTrace.java shouldn't be executed with -XX:+DeoptimizeALot
d68775d: 8351995: JFR: Leftovers from removal of Security Manager
e62becc: 8350964: Add an ArtifactResolver.fetch(clazz) method
dbf47d6: 8351876: RISC-V: enable and fix some float round tests
d207ed3: 8352066: JVM.commit() and JVM.flush() exhibit race conditions against JFR epochs
0450ba9: 8351999: JFR: Incorrect scaling of throttled values
e5666f5: 8351976: assert(vthread_epoch == current_epoch) failed: invariant
2eecf15: 8351967: JFR: AnnotationIterator should handle num_annotations = 0
c8913d2: 8345555: Improve layout of search results
e29d405: 8352110: [BACKOUT] C2: Print compilation bailouts with PrintCompilation compile command
... and 110 more: https://git.openjdk.org/jdk/compare/08929134b3533362133139c4e964b1b28de6ebfb...master

As there are no conflicts, your changes will automatically be rebased on top of these commits when integrating. If you prefer to avoid this automatic rebasing, please check the documentation for the /integrate command for further details.

➡️ To integrate this PR with the above commit message to the master branch, type /integrate in a new comment.

openjdk · 2025-03-07T01:58:42Z

@sviswa7 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.

mlbridge · 2025-03-07T02:24:22Z

Webrevs

vnkozlov

Good.

jatin-bhateja

Hi @sviswa7, The Fix looks reasonable to me. Kindly consider including applicable suggestions.

Best Regards

jatin-bhateja · 2025-03-07T07:48:07Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+        goldL = testLongKernel(aL);
+    }
+
+    @Test


Suggested change

@Test

@Test

@IR(failOn = { IRNode.VECTOR_CAST_HF2F }, applyIfCPUFeatureOr = { "avx512vl", "true", "f16c", "true" })

I left out the IR check because we do intend to vectorize this going forward. Instead the bug fix is verified by checkResult. Also the fix is not specific to Intel platform so if we do add IR check it will need to be generic.
@eme64 your thoughts please? Would you like to see an IR check here that vectorization is not happening?

Personally, I generally prefer to have failOn IR rules, if we expect that at least for now there should be no vectorization. But add a comment why we expect no vectorization, so that if it ever does vectorize and the IR rule fails the person has a hint, and does not have to reverse-engineer too much. And if it turns out that we should one day vectorize, then we already have all these tests ready to just flip the failOn into count.

jatin-bhateja · 2025-03-07T07:48:41Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+        return res;
+    }
+
+    @Test


Suggested change

@Test

@Test

@IR(counts = { IRNode.VECTOR_CAST_HF2F, " >0 " }, applyIfCPUFeatureOr = { "avx512vl", "true", "f16c", "true" })

Used the IR test from compiler/vectorization/TestFloatConversionsVector.java to include other architectures as well.

Thanks @sviswa7 , these operations are also supported on PPC.

jatin-bhateja · 2025-03-07T07:50:13Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+        return res;
+    }
+
+    @Test


Suggested change

@Test

/*

* C2 handles i2s conversion by constraining the value range of the integral argument; thus

* argument fed to ConvHF2F is of type T_INT. Fix for JDK-8350835 skips over vectorizing such a case

* for now.

*/

@Test

@IR(failOn = { IRNode.VECTOR_CAST_HF2F }, applyIfCPUFeatureOr = { "avx512vl", "true", "f16c", "true" })

I don't see any harm in including the above suggested comment as you mentioned we plan to support these auto vrctoriizations in future

jatin-bhateja · 2025-03-07T07:51:24Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+        return res;
+    }
+
+    @Test


Suggested change

@Test

/*

* C2 handles this in two steps: l2i handling creates ConvL2I IR ,followed by i2s conversion which onstrains the

* value range of the integral argument; thus, the argument fed to ConvHF2F is of type T_INT. Fix for

* JDK-8350835 skip over vectorizing such a case for now.

*/

@Test

@IR(failOn = { IRNode.VECTOR_CAST_HF2F }, applyIfCPUFeatureOr = { "avx512vl", "true", "f16c", "true" })

Thanks, added the generic failOn tests for byte, int, and long test cases as it is applicable across architectures.

sviswa7 · 2025-03-07T16:43:20Z

Thanks a lot @vnkozlov for the review and approval.

eme64 · 2025-03-10T08:56:17Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+* @key randomness
+* @bug 8350835
+* @summary Test bug fix for JDK-8350835 discovered through Template Framework
+* @requires vm.compiler2.enabled


Suggested change

* @requires vm.compiler2.enabled

Is this restriction necessary? I generally prefer running tests on all platforms, and only restricting IR rules. That way we can get more test coverage with result verification.

The restriction is not necessary, removed.

eme64 · 2025-03-10T08:57:24Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+* @summary Test bug fix for JDK-8350835 discovered through Template Framework
+* @requires vm.compiler2.enabled
+* @library /test/lib /
+* @run main/othervm -XX:-TieredCompilation -XX:CompileOnly=compiler.vectorization.TestFloat16ToFloatConv::test* compiler.vectorization.TestFloat16ToFloatConv


Are the additional flags really necessary for reproducing the bug? I would suspect not really. The IR framework already takes care of ensuring we run with C2 compilation.

Removed the additional flags.

eme64 · 2025-03-10T08:58:03Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+    private static byte[] aB = new byte[SIZE];
+    private static short[] aS = new short[SIZE];
+    private static int[] aI = new int[SIZE];
+    private static long[] aL = new long[SIZE];
+    private static float[] goldB, goldS, goldI, goldL;


Are you testing for char as well?

Added test for char.

eme64 · 2025-03-10T08:59:03Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+        for (int i = 0; i < aB.length; i++) {
+            aB[i] = (byte)RANDOM.nextInt();
+            aS[i] = (short)RANDOM.nextInt();
+            aI[i] = RANDOM.nextInt();
+            aL[i] = RANDOM.nextLong();
+        }


I would prefer if we could start using Generators. There is a fill method for arrays. It generates more "interesting" values. It is not super relevant here, but it would be nice if we made this common practice now ;)

The Generators don't support the bytes, shorts, and chars yet so ended up not using them for this PR.

eme64 · 2025-03-10T10:12:26Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+        goldL = testLongKernel(aL);
+    }
+
+    @Test


Personally, I generally prefer to have failOn IR rules, if we expect that at least for now there should be no vectorization. But add a comment why we expect no vectorization, so that if it ever does vectorize and the IR rule fails the person has a hint, and does not have to reverse-engineer too much. And if it turns out that we should one day vectorize, then we already have all these tests ready to just flip the failOn into count.

eme64 · 2025-03-10T10:13:55Z

@sviswa7 thanks for looking at this! The fix looks good, there are just a few comments about the test :)

sviswa7 · 2025-03-10T21:23:15Z

@eme64 @jatin-bhateja Your review comments are addressed, please take a look.

jatin-bhateja · 2025-03-12T03:37:23Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+    @Test
+    @IR(counts = {IRNode.VECTOR_CAST_HF2F, "> 0"},
+        applyIfOr = {"UseCompactObjectHeaders", "false", "AlignVector", "false"},
+        applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"},


Can you kindly justify the need for compressed object header usage, it will mainly impact the pre-loop trip count compuation. AlignVector should be sufficient since it's a whitelisted option

This check is taken from compiler/vectorization/TestFloatConversionsVector.java which also has float16 conversion tests to be in sync.

If I remove UseCompressedHeaders check then the test will start failing for folks working on compressed headers so good to keep it there and as I mentioned before it is good to be in sync with other Float16ToFloat conversion test.

eme64

Looks much better :)

You are right Generators are missing cases for short, byte, char. You could leave those cases with regular Random, but the int and long cases with Generators, to make sure interesing values are added to the mix more frequently.

eme64 · 2025-03-13T07:37:17Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+    }
+
+    @Test
+    // Not vectorized due to JDK-8350835


That's very non-descriptive. Actually, that is the current bug, so this is not even a future RFE that intends to fix it.
Can you please say why it is not vectorizing now, and what might be possible conditions when it would be ok to vectorize in the future? Could we even file an RFE for this?

Added descriptive comments. RFE filed at https://bugs.openjdk.org/browse/JDK-8352093.

eme64 · 2025-03-13T07:38:41Z

test/hotspot/jtreg/compiler/vectorization/TestFloat16ToFloatConv.java

+* @bug 8350835
+* @summary Test bug fix for JDK-8350835 discovered through Template Framework
+* @library /test/lib /
+* @run main/othervm compiler.vectorization.TestFloat16ToFloatConv


Suggested change

* @run main/othervm compiler.vectorization.TestFloat16ToFloatConv

* @run driver compiler.vectorization.TestFloat16ToFloatConv

I don't think you need a new VM if you have no additional flags ;)

sviswa7 · 2025-03-15T00:53:44Z

@eme64 @jatin-bhateja Your review comments are handled. Please take a look.

eme64 · 2025-03-17T06:58:39Z

@sviswa7 The code and test looks good to me now. I'm re-running testing. Please ping me in a day for the results :)

eme64

Tests are passing. Approved.
@sviswa7 Thanks for the work 😊

sviswa7 · 2025-03-17T16:59:14Z

Thanks a lot @eme64.

sviswa7 · 2025-03-17T17:49:39Z

/integrate

openjdk · 2025-03-17T17:50:35Z

Going to push as commit 3239919.
Since your change was applied there have been 123 commits pushed to the master branch:

47c1960: 8351689: -Xshare:dump with default classlist fails on static JDK
6b82b42: 8348598: Update Libpng to 1.6.47
2674a31: 8351891: Disable TestBreakSignalThreadDump.java#with_jsig and XCheckJSig.java on static JDK
4c6a523: 8352096: Test jdk/jfr/event/profiling/TestFullStackTrace.java shouldn't be executed with -XX:+DeoptimizeALot
d68775d: 8351995: JFR: Leftovers from removal of Security Manager
e62becc: 8350964: Add an ArtifactResolver.fetch(clazz) method
dbf47d6: 8351876: RISC-V: enable and fix some float round tests
d207ed3: 8352066: JVM.commit() and JVM.flush() exhibit race conditions against JFR epochs
0450ba9: 8351999: JFR: Incorrect scaling of throttled values
e5666f5: 8351976: assert(vthread_epoch == current_epoch) failed: invariant
... and 113 more: https://git.openjdk.org/jdk/compare/08929134b3533362133139c4e964b1b28de6ebfb...master

Your commit was automatically rebased without conflicts.

openjdk · 2025-03-17T17:50:41Z

@sviswa7 Pushed as commit 3239919.

💡 You may see a message that your pull request was closed with unmerged commits. This can be safely ignored.

sviswa7 added 2 commits March 6, 2025 17:17

8350835: C2 SuperWord: assert/wrong result when using Float.float16To…

24669ac

…Float with byte instead of short input

some updates

Loading
Loading status checks…

62bef25

openjdk bot added the hotspot-compiler label Mar 7, 2025

whitespace

Loading
Loading status checks…

4ebb47a

sviswa7 marked this pull request as ready for review March 7, 2025 02:20

openjdk bot added the rfr label Mar 7, 2025

vnkozlov approved these changes Mar 7, 2025

View reviewed changes

openjdk bot added the ready label Mar 7, 2025

jatin-bhateja reviewed Mar 7, 2025

View reviewed changes

eme64 suggested changes Mar 10, 2025

View reviewed changes

review comments

Loading
Loading status checks…

70ab0ac

openjdk bot removed the ready label Mar 10, 2025

jatin-bhateja reviewed Mar 12, 2025

View reviewed changes

eme64 suggested changes Mar 13, 2025

View reviewed changes

Review comments from Emanuel

Loading
Loading status checks…

8069f4b

eme64 approved these changes Mar 17, 2025

View reviewed changes

openjdk bot added the ready label Mar 17, 2025

openjdk bot added the integrated label Mar 17, 2025

openjdk bot closed this Mar 17, 2025

openjdk bot removed ready rfr labels Mar 17, 2025

	@Test
	@Test
	@IR(failOn = { IRNode.VECTOR_CAST_HF2F }, applyIfCPUFeatureOr = { "avx512vl", "true", "f16c", "true" })

	@Test
	@Test
	@IR(counts = { IRNode.VECTOR_CAST_HF2F, " >0 " }, applyIfCPUFeatureOr = { "avx512vl", "true", "f16c", "true" })

-    @Test
+    /*
+     *  C2 handles i2s conversion by constraining the value range of the integral argument; thus
+     *  argument fed to ConvHF2F is of type T_INT. Fix for JDK-8350835 skips over vectorizing such a case
+     *  for now.
+     */
+     @Test
+     @IR(failOn = { IRNode.VECTOR_CAST_HF2F }, applyIfCPUFeatureOr = { "avx512vl", "true", "f16c", "true" })

-    @Test
+    /*
+     * C2 handles this in two steps: l2i handling creates ConvL2I IR ,followed by i2s conversion which onstrains the
+     * value range of the integral argument; thus, the argument fed to ConvHF2F is of type T_INT. Fix for
+     * JDK-8350835 skip over vectorizing such a case for now.
+     */
+    @Test
+    @IR(failOn = { IRNode.VECTOR_CAST_HF2F }, applyIfCPUFeatureOr = { "avx512vl", "true", "f16c", "true" })

	* @run main/othervm compiler.vectorization.TestFloat16ToFloatConv
	* @run driver compiler.vectorization.TestFloat16ToFloatConv

8350835: C2 SuperWord: assert/wrong result when using Float.float16ToFloat with byte instead of short input #23939

8350835: C2 SuperWord: assert/wrong result when using Float.float16ToFloat with byte instead of short input #23939

Conversation

sviswa7 commented Mar 7, 2025 • edited by openjdk bot Loading

Progress

Issue

Reviewers

Reviewing

bridgekeeper bot commented Mar 7, 2025

openjdk bot commented Mar 7, 2025 • edited Loading

openjdk bot commented Mar 7, 2025

mlbridge bot commented Mar 7, 2025 • edited Loading

Webrevs

vnkozlov left a comment

Choose a reason for hiding this comment

jatin-bhateja left a comment • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sviswa7 commented Mar 7, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eme64 commented Mar 10, 2025

sviswa7 commented Mar 10, 2025

jatin-bhateja Mar 12, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eme64 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sviswa7 commented Mar 15, 2025

eme64 commented Mar 17, 2025

eme64 left a comment

Choose a reason for hiding this comment

sviswa7 commented Mar 17, 2025

sviswa7 commented Mar 17, 2025

openjdk bot commented Mar 17, 2025

openjdk bot commented Mar 17, 2025

sviswa7 commented Mar 7, 2025 •

edited by openjdk bot

Loading

openjdk bot commented Mar 7, 2025 •

edited

Loading

mlbridge bot commented Mar 7, 2025 •

edited

Loading

jatin-bhateja left a comment •

edited

Loading

jatin-bhateja Mar 12, 2025 •

edited

Loading