RFR: 8333890: Fatal error in auto-vectorizer with float16 kernel. [v2]

[Emanuel Peter]

On Sat, 6 Jul 2024 23:55:04 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

>> Hi,
>> 
>> This bug fix patch fixes an assertion failure seen while auto-vectorizing conversion chain involving float16 type.
>> Constraining the container type for ConvF2HF IR to short type upfront , as relying on backpropagating narrow
>> type from memory operation  may not be sufficient if IR is succeeded by ConvHF2F as depicted by the include test case.
>> 
>> Best Regards,
>> Jatin
>
> Jatin Bhateja has updated the pull request incrementally with one additional commit since the last revision:
> 
>   IR rule extension for ANY vector size

Also: Vladimir did not run testing on this, and now it fail in the CI

test/hotspot/jtreg/compiler/vectorization/TestFloat16VectorConvChain.java line 42:

> 40: 
> 41:     @Test
> 42:     @IR(counts = {IRNode.VECTOR_CAST_HF2F, IRNode.VECTOR_SIZE_ANY, ">= 1", IRNode.VECTOR_CAST_F2HF, IRNode.VECTOR_SIZE_ANY, " >= 1"})

You are missing restrictions to platforms where these vector operations are supported.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/20062#issuecomment-2213658496
PR Review Comment: https://git.openjdk.org/jdk/pull/20062#discussion_r1668404974