[stdlib] Add Apple SIMDGROUP barrier implementation for syncwarp

[YichengDWu]

Closes #5471

[Copilot]

Pull Request Overview

This PR adds Apple GPU support for the syncwarp function by implementing a SIMDGROUP barrier. The implementation provides execution synchronization for Apple GPUs, complementing the existing NVIDIA and AMD GPU support.

• Adds Apple GPU-specific barrier implementation using the llvm.air.simdgroup.barrier intrinsic
• Updates documentation to clarify Apple GPU behavior and limitations
• Notes that lane masks are not supported on Apple GPUs and only execution synchronization (not memory ordering) is provided

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[npanchen]

Thanks for adding this!
Can you please add simple test for syncward similar to test_barrier.mojo ?

Also, since nightly compiler by default mangles AIR intrinsics, I'll need to update compiler internals not do to this for air.simdgroup.barrier. I'll share an update when it will be safe to merge

[YichengDWu]

Thanks! A deterministic syncwarp-only test (no memory fence) relies on a warp-wide register op like shuffle, otherwise the test either becomes flaky or ends up testing a different primitive (e.g., threadgroup barrier + fence). I'll follow up with the test once shuffle lands.

[npanchen]

Thanks for adding this! Can you please add simple test for syncward similar to test_barrier.mojo ?

Also, since nightly compiler by default mangles AIR intrinsics, I'll need to update compiler internals not do to this for air.simdgroup.barrier. I'll share an update when it will be safe to merge

Alright. I've merged compiler changes to avoid mangling of this intrinsic. After remaining things are done it's good for me to be merged

[JoeLoser]

Thanks for adding this! Can you please add simple test for syncward similar to test_barrier.mojo ?
Also, since nightly compiler by default mangles AIR intrinsics, I'll need to update compiler internals not do to this for air.simdgroup.barrier. I'll share an update when it will be safe to merge

Alright. I've merged compiler changes to avoid mangling of this intrinsic. After remaining things are done it's good for me to be merged

@YichengDWu I just released a new nightly release that has Kolya's change, can you try it out please? Mojo compiler won't mangle this AIR intrinsic now.

[BradLarson]

I tried this out locally, and it builds correctly on the latest nightly. We unfortunately don't have any direct tests of syncwarp(), but I tried a manual test of it and the intrinsic did compile and the code ran correctly as far as I can tell.

[BradLarson]

!sync

[YichengDWu]

Glad to hear that! I just pulled the latest nightly and can confirm syncwarp() now builds and runs on my Mac.

[modularbot]

✅🟣 This contribution has been merged 🟣✅

Your pull request has been merged to the internal upstream Mojo sources. It will be reflected here in the Mojo repository on the main branch during the next Mojo nightly release, typically within the next 24-48 hours.

We use Copybara to merge external contributions, click here to learn more.

[JoeLoser]

Glad to hear that! I just pulled the latest nightly and can confirm syncwarp() now builds and runs on my Mac.

Nice! We just synced this in and landed it internally: it'll be available in tomorrow's nightly release. Great contribution!

If you're interested in more Apple Metal work, #5472 is an interesting one and would unlock several of the GPU puzzles to work on Apple Metal from https://github.com/modular/mojo-gpu-puzzles.

[modularbot]

Landed in 98447e5! Thank you for your contribution 🎉