[llvm] [TableGen] Emit better error message for duplicate Subtarget features. (PR #102090)

[Rahul Joshi via llvm-commits]

jurahul wrote:

Thanks for reporting this. I suspect the issue is that llvm::sort is not
stable, so the 2 records can end up in any order in the sorted list,
leading to non determinism in the error message. The fix would be to use a
secondary key for sorting to make it deterministic (Record name if “Name”
field values are same).

I am working on a follow on patch for this and will incorporate that fix in
that. Meanwhile let me disable the test.

Thanks,
Rahul

On Wed, Aug 7, 2024 at 2:56 AM Haojian Wu ***@***.***> wrote:

> Hi, I see a flaky issue when integrating this change into our internal
> codebase.
>
> The newly-added test is flaky under ASAN. Sometimes it passes, sometimes
> it fails.
> The symptom I observed is that the line number of the emitted warning is
> non-deterministic:
>
>    1. Line 10: ***@***.***+2]]:5: error: Feature 'NameA' already defined.
>    (expected: 12)
>    2. Line 11: ***@***.***]:5: note: Previous definition here. (expected:
>    8)
>
> Under ASAN, the line number of the first diagnostic can sometimes be 8,
> and the second one can be 12—they just swap. It is very subtle, and I don't
> see any problem in your patch.
>
> —
> Reply to this email directly, view it on GitHub
> <https://github.com/llvm/llvm-project/pull/102090#issuecomment-2273084124>,
> or unsubscribe
> <https://github.com/notifications/unsubscribe-auth/APRMUB7UWTPJ6XDHUHMTEOLZQHVMBAVCNFSM6AAAAABMBKTXOCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENZTGA4DIMJSGQ>
> .
> You are receiving this because you were assigned.Message ID:
> ***@***.***>
>


https://github.com/llvm/llvm-project/pull/102090