On Mon, 9 Mar 2026 19:36:05 GMT, Mohamed Issa <[email protected]> wrote:
>> Although the scalar AVX10 floating point min/max instructions (VMINMAXSD, >> VMINMAXSS, VMINMAXSH) are compact, it's better not to use them in reduction >> loops. This is because of serial data dependencies that get triggered across >> loop iterations. An alternate implementation using comparisons and jumps >> leverages branch prediction and limits the effects of data dependencies to >> cheaper instructions (e.g, MOV). Please note that this method is already >> used for non-AVX10 min/max reduction loop scenarios. >> >> With that background provided, these changes remove AVX10 floating point >> min/max instructions from single and double precision floating point >> reduction loops. They are replaced by the separate instruction sequence >> described above. Currently, min/max half precision floating point reduction >> loops aren't detectable, so they will be handled in a separate PR. There is >> also some code cleanup to remove unused instruction definitions while also >> adding necessary supporting infrastructure. The JTREG tests listed below >> were used to verify correctness with the recommended JVM options mentioned >> in corresponding source files. All modifications and tests used [OpenJDK >> v27-b12](https://github.com/openjdk/jdk/releases/tag/jdk-27%2B12) as the >> baseline build. >> >> 1. `jtreg:test/jdk/jdk/incubator/vector/DoubleVector64Tests.java` >> 2. `jtreg:test/jdk/jdk/incubator/vector/DoubleVector128Tests.java` >> 3. `jtreg:test/jdk/jdk/incubator/vector/DoubleVector256Tests.java` >> 4. `jtreg:test/jdk/jdk/incubator/vector/DoubleVector512Tests.java` >> 5. `jtreg:test/jdk/jdk/incubator/vector/DoubleVectorMaxTests.java` >> 6. `jtreg:test/jdk/jdk/incubator/vector/FloatVector64Tests.java` >> 7. `jtreg:test/jdk/jdk/incubator/vector/FloatVector128Tests.java` >> 8. `jtreg:test/jdk/jdk/incubator/vector/FloatVector256Tests.java` >> 9. `jtreg:test/jdk/jdk/incubator/vector/FloatVector512Tests.java` >> 10. `jtreg:test/jdk/jdk/incubator/vector/FloatVectorMaxTests.java` >> 11. >> `jtreg:test/hotspot/jtreg/compiler/vectorization/TestFloat16VectorOperations.java` >> 12. >> `jtreg:test/hotspot/jtreg/compiler/c2/irTests/TestFloat16ScalarOperations.java` >> 13. `jtreg:test/hotspot/jtreg/compiler/igvn/TestMinMaxIdentity.java` >> 14. >> `jtreg:test/hotspot/jtreg/compiler/intrinsics/float16/TestFloat16MaxMinSpecialValues.java` >> >> Finally, the JMH micro-benchmarks listed below were updated to ensure all >> code paths are exercised. >> >> 1. >> `micro:test/micro/org/openjdk/bench/jdk/incubator/vector/Float16OperationsBenchmark.java` >> 2. `mi... > > Mohamed Issa has updated the pull request with a new target base due to a > merge or a rebase. The pull request now contains five commits: > > - Merge branch 'master' into user/missa-prime/avx10_2 > - Remove half precision min/max reduction definitions and adjust > corresponding benchmarks. > - Use alternative instruction flow for half precision reduction loops and > add supporting infrastructure. > - Merge branch 'master' into user/missa-prime/avx10_2 > - Replace scalar AVX10.2 floating point min/max instructions with more > efficient sequence src/hotspot/cpu/x86/macroAssembler_x86.cpp line 1969: > 1967: vmovw(rscratch, src); > 1968: vmovw(dst, rscratch); > 1969: } There is a direct vmovw xmm1, xmm2 instruction in AVX10_2. We should use that here. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/29831#discussion_r2928045394
