[all-commits] [llvm/llvm-project] 2f238d: [flang][cuda] Treat unified data as host data in a...

[Zhen Wang via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 2f238db9da1cec04f7e143fb264aa1c0fd4578ef
      https://github.com/llvm/llvm-project/commit/2f238db9da1cec04f7e143fb264aa1c0fd4578ef
  Author: Zhen Wang <zhenw at nvidia.com>
  Date:   2026-07-30 (Thu, 30 Jul 2026)

  Changed paths:
    M flang/include/flang/Evaluate/tools.h
    M flang/test/Lower/CUDA/cuda-managed-assign.cuf
    A flang/test/Lower/CUDA/cuda-unified-assign.cuf

  Log Message:
  -----------
  [flang][cuda] Treat unified data as host data in assignments (#212912)

Unified data was counted as managed data in the data transfer predicate,
so assignments between unified sides, or between unified and host data,
became synchronous copies.

Unified data is host memory that the device can also access, so it now
takes the place of host data there: the assignment is a copy only when
the other side is device or constant data, or a whole managed array, and
is done on the host otherwise.

This matches the reference compiler for whole arrays, sections,
elements, scalars, allocatables, components and expressions.



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