[llvm-branch-commits] [llvm] [mlir] [MLIR][OpenMP] Support device shared memory for recipe ops (PR #238364)

Akash Banerjee via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Tue Nov 3 09:50:37 PST 2026


================
@@ -2484,6 +2752,10 @@ cleanupPrivateVars(T op, llvm::IRBuilderBase &builder,
   for (PrivateVarsInfo::VarInfo &var : privateVarsInfo.vars) {
     if (mightUseDeviceSharedMem &&
         omp::allocaUsesRequireSharedMem(var.blockArg)) {
+      // Free the init-region internal allocations before the top-level private
+      // storage, in reverse order.
+      for (const DeviceSharedAlloc &alloc : llvm::reverse(var.sharedAllocs))
----------------
TIFitis wrote:

The new shared-allocation cleanup pattern breaks default AMDGPU compilation of a runtime-sized array privatised on target and accessed inside a nested parallel. I've attached a minimal reproducer below:

```
subroutine test(n, y)
  use omp_lib
  integer :: n, y, x(n)
  !$omp target private(x) map(from:y)
    x(1) = 42
    !$omp parallel num_threads(128)
      if (omp_get_thread_num() == 0) y = x(1)
    !$omp end parallel
  !$omp end target
end subroutine

program main
  integer :: y
  call test(4, y)
  if (y /= 42) stop 1
end program
```

https://github.com/llvm/llvm-project/pull/238364


More information about the llvm-branch-commits mailing list