[all-commits] [llvm/llvm-project] 795a5a: [flang][cuda] Implicitly attribute allocatables/po...

[Vijay Kandiah via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 795a5a9a19bb2b356d82a46426d95f245d1d0049
      https://github.com/llvm/llvm-project/commit/795a5a9a19bb2b356d82a46426d95f245d1d0049
  Author: Vijay Kandiah <vkandiah at nvidia.com>
  Date:   2026-07-13 (Mon, 13 Jul 2026)

  Changed paths:
    M flang/lib/Semantics/resolve-names.cpp
    A flang/test/Lower/CUDA/cuda-implicit-managed-alloc.cuf

  Log Message:
  -----------
  [flang][cuda] Implicitly attribute allocatables/pointers under -gpu=unified (#209292)

When `-gpu=unified` is in effect, allocatables and pointers with no
explicit CUDA data attribute were left with no data attribute, and were allocated
as ordinary host memory. Under unified memory, such objects must be
reachable from the device; leaving them as plain host allocations makes device kernels
fault on systems that do not provide transparent host-memory access.

The existing code here already assigned `Managed` implicitly under
`-gpu=managed` (when CUDA Fortran is enabled). This extends that handling to
`-gpu=unified`: prefer the `Unified` attribute where it is legal (host subprogram, main
program, or derived-type component) so generic resolution still selects the unified
specific, and fall back to `Managed` elsewhere (module scope, device
subprograms). Both attributes route allocation through the same managed
allocator, so this keeps the allocation device-accessible while avoiding
spurious `ATTRIBUTES(UNIFIED)` legality errors.

The attribution stays gated on CUDA Fortran being enabled, preserving
the existing behavior for non-CUDA-Fortran translation units.



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