[TritonIntelGPUToLLVM] Vectorize masked StoreOp when mask isn't statically uniform

[anmyachev]

No description provided.

[Copilot]

Pull request overview

This PR improves TritonIntelGPUToLLVM lowering for masked triton::StoreOp when the mask is not statically uniform across the chosen vectorization width, by emitting a fast wide-store path guarded by a runtime “all-lanes-valid” predicate and falling back to exact per-element stores otherwise.

Changes:

• Add a helper to build two-armed predicated control flow for side-effecting, result-less regions.
• Update StoreOp lowering to keep wide vectorization even with non-uniform masks by dynamically guarding the wide store and introducing a per-element slow path fallback.
• Update the MLIR conversion test to assert the new two-armed guarded structure and expected store widths/alignments.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
third_party/intel/lib/TritonIntelGPUToLLVM/Utility.h	Adds createTwoArmedPredicatedBlock helper used to generate fast/slow guarded control flow.
third_party/intel/lib/TritonIntelGPUToLLVM/LoadStoreOpToLLVM.cpp	Changes masked StoreOp lowering to prefer wide stores with a runtime group mask AND, with a scalar per-lane fallback for correctness.
test/Conversion/intel/load_store_256b_to_llvm.mlir	Updates checks to validate the new guarded wide-store + scalar-fallback lowering.

Comments suppressed due to low confidence (1)

third_party/intel/lib/TritonIntelGPUToLLVM/Utility.h:108

• Same issue as above for the slow-path callback: using std::forward here requires <utility> (not included in this header) and isn't needed for typical lambda callbacks. Calling the functor directly avoids relying on transitive includes.

  rewriter.setInsertionPointToStart(slowBlock);
  (void)std::forward<SlowFn>(slowFn)();
  cf::BranchOp::create(rewriter, loc, endBlock);