[all-commits] [llvm/llvm-project] 752647: [mlir][OpenACC] Barrier seq loops that reuse gang-...

[Delaram Talaashrafi via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 752647bcfcbc97c1ffe8db1d0cc50ec71b3957cf
      https://github.com/llvm/llvm-project/commit/752647bcfcbc97c1ffe8db1d0cc50ec71b3957cf
  Author: Delaram Talaashrafi <dtalaashrafi at nvidia.com>
  Date:   2026-09-22 (Tue, 22 Sep 2026)

  Changed paths:
    M mlir/lib/Dialect/OpenACC/Transforms/ACCCGToGPU.cpp
    A mlir/test/Dialect/OpenACC/acc-cg-to-gpu-seq-loop-routine-reuse-barrier.mlir

  Log Message:
  -----------
  [mlir][OpenACC] Barrier seq loops that reuse gang-private slots after worker calls (#224437)

A sequential `scf.for` that stores a gang-private value and then calls a
worker- or vector-level acc routine can let threads overlap iterations
on the same workgroup slot. Insert a workgroup barrier at the top of
that loop body so the next iteration's store waits for the previous
iteration's last read.
Skip sequential gang remnants, loops under thread-level worksharing, and
bodies already inside `scf.if`, where a workgroup barrier would
deadlock.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications