[all-commits] [llvm/llvm-project] b76485: [Flang][OpenMP] Heap-allocate GPU dynamic private ...

[Akash Banerjee via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: b764858749de56611794a7e92b32ee3008936337
      https://github.com/llvm/llvm-project/commit/b764858749de56611794a7e92b32ee3008936337
  Author: Akash Banerjee <akash.banerjee at amd.com>
  Date:   2026-06-05 (Fri, 05 Jun 2026)

  Changed paths:
    M flang/include/flang/Lower/Support/PrivateReductionUtils.h
    M flang/include/flang/Lower/Support/Utils.h
    M flang/lib/Lower/OpenMP/DataSharingProcessor.cpp
    M flang/lib/Lower/OpenMP/DataSharingProcessor.h
    M flang/lib/Lower/OpenMP/OpenMP.cpp
    M flang/lib/Lower/Support/PrivateReductionUtils.cpp
    M flang/lib/Lower/Support/Utils.cpp
    A flang/test/Lower/OpenMP/DelayedPrivatization/target-teams-distribute-parallel-do-simd-private-adjustable-array.f90
    A flang/test/Lower/OpenMP/DelayedPrivatization/target-teams-distribute-private-adjustable-array.f90
    A flang/test/Lower/OpenMP/DelayedPrivatization/target-teams-nested-distribute-private-adjustable-array.f90

  Log Message:
  -----------
  [Flang][OpenMP] Heap-allocate GPU dynamic private arrays in distribute parallel do (#200841)

Fixes GPU offload crashes for Fortran automatic arrays privatised in
target teams distribute parallel do.

For delayed privatisation on GPU, dynamically sized boxed array privates
are now routed through the existing heap-allocation path, with matching
cleanup emitted in the privatiser dealloc region. This avoids lowering
such arrays to runtime-sized scratch allocas whose descriptors can be
captured across the distribute callback boundary.

Fixes [#2419](https://github.com/ROCm/llvm-project/issues/2419).
Co-authored-by: Codex <codex at openai.com>



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