x86: on targets that requires SSE, use those registers for ABI

[RalfJung]

View all comments

Now that llvm/llvm-project#142321 is fixed, we can use the vector ABI without re-introducing #139029. So this effectively reverts #141309.

This can lead to new hard errors on code that previously compiled if all of the following apply:

• One is compiling for an i686 target (or for i386-apple-ios which is effectively an i686 target despite the name) -- note that i586 targets are not affected
• One has disabled SSE (which has emitted a non-silencable warning since Rust 1.86 as these targets require SSE)
• The program uses a "Rust" ABI function that passes a 128bit SIMD vector by-value

Also this will cause performance regressions on LLVM 22 and 21 as those don't have the fix for llvm/llvm-project#142321. IMO that's fine. Programs will do the right thing, they will just be a bit slower due to #139029.

@rust-lang/lang please let me know if you're fine with just landing this or if you want to apply some sort of process here. (We have just landed this the first time we tried to use SSE in the ABI, but back then I didn't think of the fact that in some unsupported corner cases, this can lead to new hard errors.) If you do think that this needs FCP, please let me know and I will propose that we turn the target-feature ABI check into a hard error on x86-32 (similar to #161280).

TODO:

• Once make target feature ABI check a hard error on ARM #161280 lands, also make the check a hard error on x86-32.

r? @nikic
Cc @tgross35

[RalfJung]

@bors try jobs=i686,various

[rust-bors]

☀️ Try build successful (CI)
Build commit: a32c597 (a32c597e182d95bb782185c7b388721b365ca029)
Base parent: be0ea33 (be0ea33bba5000ac134e58acde780af83d908e44)

[RalfJung]

@bors try jobs=aarch64-linux

[rust-bors]

☀️ Try build successful (CI)
Build commit: 6212267 (6212267738008defc37104e01a0afc3421c7d50e)
Base parent: 9a4ad59 (9a4ad59ae3073b013cd62f53f8349ddc61a012e8)

[RalfJung]

It is strange that PR CI passed. Looks like even on LLVM 21, tests/codegen-llvm/pclmulqdq-target-feature-inlining.rs works? How can that be if it does not have llvm/llvm-project#205106 ?

[RalfJung]

Oh wait nvm, there was a t-lang ask here that I forgot about. :)

[tmandry]

I think we should make this change, including the breaking change and turning the non-silenceable warning for disabling SSE on i686 targets into a hard error.

@rfcbot fcp merge lang

[rust-rfcbot]

@tmandry has proposed to merge this. The next step is review by the rest of the tagged team members:

• @joshtriplett
• @nikomatsakis
• @scottmcm
• @tmandry
• @traviscross

No concerns currently listed.

Once a majority of reviewers approve (and at most 2 approvals are outstanding), this will enter its final comment period. If you spot a major issue that hasn't been raised at any point in this process, please speak up!

cc @rust-lang/lang-advisors: FCP proposed for lang, please feel free to register concerns.
See this document for info about what commands tagged team members can give me.

[traviscross]

Thanks @RalfJung.

@rfcbot reviewed

If you do think that this needs FCP...

It has been our practice to FCP restabilizations (see, e.g., #145954 (comment)).

[RalfJung]

@bors try jobs=llvm

[rust-rfcbot]

🔔 This is now entering its final comment period, as per the review above. 🔔

[RalfJung]

turning the non-silenceable warning for disabling SSE on i686 targets into a hard error.

All right, will do. This is blocked on #161280 (FCP finished there but it hasn't been reviewed yet).

[RalfJung]

It is strange that PR CI passed. Looks like even on LLVM 21, tests/codegen-llvm/pclmulqdq-target-feature-inlining.rs works? How can that be if it does not have llvm/llvm-project#205106 ?

Ah, it doesn't work in LLVM 21, we just don't run that in PR CI. :) (We run it on aarch64, but not on x86.)

[RalfJung]

@bors try jobs=x86llvm

[rust-bors]

☀️ Try build successful (CI)
Build commit: 88b6b60 (88b6b6084feceee6c66834108ee44fd8cca324e9)
Base parent: 824336a (824336ad4127ce295849937a24c08a4aeff6ada7)