[all-commits] [llvm/llvm-project] 8472bc: [flang][cuda] Don't allocate managed descriptors f...

[khaki3 via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 8472bc79bf157925c87c3acb7dbb4ac4ddb3070d
      https://github.com/llvm/llvm-project/commit/8472bc79bf157925c87c3acb7dbb4ac4ddb3070d
  Author: khaki3 <47756807+khaki3 at users.noreply.github.com>
  Date:   2026-06-08 (Mon, 08 Jun 2026)

  Changed paths:
    M flang/lib/Optimizer/CodeGen/CodeGen.cpp
    M flang/test/Fir/CUDA/cuda-code-gen.mlir

  Log Message:
  -----------
  [flang][cuda] Don't allocate managed descriptors for OpenACC data-clause boxes (#201957)

Example:
```fortran
module m
  real(8), allocatable :: u(:,:,:,:)   ! device-resident allocatable
contains
  subroutine core(uf, n)
    integer :: n
    real(8) :: uf(n,n,n,5)
    !$acc kernels loop present(uf)      ! descriptor built for uf
    ...
  end subroutine
  subroutine wrap(n)
    call core(u, n)
  end subroutine
end module
```

In this code, the descriptor for `present(uf)` is only used by `acc`
data-clause ops (the kernel gets the clause result, not the box), so it
never needs to be device-accessible. But `isDeviceAllocation()` is true,
so it's allocated via `_FortranACUFAllocDescriptor` in managed memory
and never freed — leaking it and leaving stale descriptors when the
address is reused.

Fix: allocate a descriptor in managed memory only when truly
device-accessible — used by a GPU kernel launch
(`isUsedByGPULaunchFunc`), or device-allocated and not consumed by an
`acc` data-clause op. Add `isUsedByOpenACCDataClause` and gate the
`isDeviceAllocation` branch with it.



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