RFR: 8372154: AArch64: Match rule failure with some CompareAndSwap operand shapes [v2]

[Andrew Dinn]

On Thu, 20 Nov 2025 07:55:51 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> See bug for more details. 
>> 
>> Following up on [JDK-8371959](https://bugs.openjdk.org/browse/JDK-8371959) failures, I managed to reproduce the "bad AD" file assert. It is heavily intermittent, and needs hours of runs before we hit the lucky seed, plus [JDK-8360557](https://bugs.openjdk.org/browse/JDK-8360557) to have broader testing scope. 
>> 
>> It looks like `CastII` node accepts the wider operand type (`iRegI`), which fails to match against narrower type in CAS match rules (`iRegINoSp`). It makes sense to use `iRegINoSp` for destination regs, so that we do not start writing to these special registers. But for operand registers, it makes little sense, IMO. I note that cas.m4-generated CAE/WCAS stubs actually already have the wider `iRegI` for operand types. 
>> 
>> So it looks to me the manual CAS match rules should also use `iRegI`. It would be even better to auto-generate these match rules from M4 stencils, and I tried that, but ultimately decided it obscures the actual bug fix. [JDK-8372188](https://bugs.openjdk.org/browse/JDK-8372188) is dedicated to moving the match rules, hopefully without the semantic change.
>> 
>> In this change, I dropped `*NoSp` from CAS operand match rules. It fixes the `iRegI` mismatch, and prepares us for harmonizing these rules with the rest of CAS/CAE generated ones.
>> 
>> Additional testing:
>>  - [x] Linux AArch64 server fastdebug, local `bad AD` assert reproducer no longer fails
>>  - [x] Linux AArch64 server fastdebug, `all`
>>  - [ ] Linux AArch64 server fastdebug, jcstress run
>
> Aleksey Shipilev has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains two additional commits since the last revision:
> 
>  - Merge branch 'master' into JDK-8372154-aarch64-cas-operand-match
>  - Fix

Marked as reviewed by adinn (Reviewer).

Ah, got it. I was confused by the explanation in the description. It's the instruction encodings that are using the wrong input type restrictions (iRegXNoSp). The match rules specify the wider register classes (iRegX). Well spotted!

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

PR Review: https://git.openjdk.org/jdk/pull/28398#pullrequestreview-3487004950
PR Comment: https://git.openjdk.org/jdk/pull/28398#issuecomment-3557185156