add nvptx_target_feature and use unaligned barrier #1805

jedbrown · 2025-05-22T22:35:57Z

Deprecate _syncthreads (the CUDA name) in favor of new _barrier_sync (NVPTX name barrier.sync).
The: barrier.sync instruction is equivalent to barrier.sync.aligned prior to sm_70, and will lead to errors/deadlock if passes (such as MIR JumpThreading) lose the aligned property.
Since: MIR does not currently have a way to apply something like LLVM's convergent attribute (and because convergent does not preserve alignment, which can be broken by inlining), we cannot pre 8000 vent loss of alignment, and thus we require target feature sm_70. In short, we cannot prevent miscompilation of aligned barriers without hard-to-specify preconditions.

This requires rust-lang/rust#138689 (for nvptx_target_feature) and fixes rust-lang/rust#137086.

* Deprecate _syncthreads (the CUDA name) in favor of new _barrier_sync (NVPTX name barrier.sync). * The: barrier.sync instruction is equivalent to barrier.sync.aligned prior to sm_70, and will lead to errors/deadlock if passes (such as MIR JumpThreading) lose the aligned property. rust-lang/rust#137086 https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-bar * Since: MIR does not currently have a way to apply something like LLVM's convergent attribute (and because convergent does not preserve alignment, which can be broken by inlining), we cannot prevent loss of alignment, and thus we require target feature sm_70. https://llvm.org/docs/ConvergentOperations.html

rustbot · 2025-05-22T22:36:01Z

rustbot has assigned @Amanieu.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

jedbrown added 2 commits May 19, 2025 13:33

add nvptx_target_feature

303f0dd

rustbot assigned Amanieu May 22, 2025

jedbrown mentioned this pull request May 23, 2025

add nvptx_target_feature rust-lang/rust#138689

Open

Provide feedback