[all-commits] [llvm/llvm-project] 304c66: [flang][OpenMP] Support lowering of metadirective ...

[Chi-Chun, Chen via All-commits]

  Branch: refs/heads/users/cchen/flang-metadirective-loop-lowering
  Home:   https://github.com/llvm/llvm-project
  Commit: 304c66040a4b09f4cb44c1bea96354bb1593e35f
      https://github.com/llvm/llvm-project/commit/304c66040a4b09f4cb44c1bea96354bb1593e35f
  Author: Chi Chun, Chen <chichun.chen at hpe.com>
  Date:   2026-07-10 (Fri, 10 Jul 2026)

  Changed paths:
    M flang/lib/Lower/OpenMP/DataSharingProcessor.cpp
    M flang/lib/Lower/OpenMP/OpenMP.cpp
    M flang/lib/Lower/OpenMP/Utils.cpp
    M flang/lib/Lower/OpenMP/Utils.h
    R flang/test/Lower/OpenMP/Todo/metadirective-loop.f90
    A flang/test/Lower/OpenMP/Todo/metadirective-module-loop.f90
    A flang/test/Lower/OpenMP/Todo/metadirective-target-loop.f90
    A flang/test/Lower/OpenMP/metadirective-loop.f90

  Log Message:
  -----------
  [flang][OpenMP] Support lowering of metadirective (part 3)

Enable lowering of metadirectives that resolve to a loop-associated
variant such as `do`, `simd`, `parallel do`, and `do simd`. The
associated DO construct is made available to the selected variant so the
existing OpenMP loop lowering can process it. A standalone metadirective
uses the following sibling DO, and a begin/end metadirective uses the
first substantive evaluation nested in its block.

For example, when the `vendor(llvm)` selector matches, the `parallel do`
variant is lowered together with the DO loop that follows it:

```fortran
!$omp metadirective when(implementation={vendor(llvm)}: parallel do)
do i = 1, n
  y(i) = a*x(i) + y(i)
end do
```

```mlir
omp.parallel {
  omp.wsloop {
    omp.loop_nest (%iv) = (%lb) to (%ub) ... { ... }
  }
}
```

Because semantic analysis cannot know which variant is selected,
induction-variable data-sharing attributes are marked during lowering and
scoped to the selected variant, so DSA state does not leak between
alternatives. A variant that opens a parallel, task, or teams region also
privatizes the indices of its nested sequential loops. `collapse(n)` and
`ordered(n)` use the normal affected-loop-depth calculation, and SIMD
variants are recognized from their leaf constructs.

Dynamic selection preserves the associated DO evaluation in every arm, so
a runtime `condition(flag)` that may pick a loop-associated variant or a
standalone fallback does not drop the loop body from the fallback arm.

Lowering assumes the associated loop nest is a canonical DO nest that is
deep enough for any `collapse(n)` or `ordered(n)`. The merged metadirective
loop-nest semantic checks
([#207088](https://github.com/llvm/llvm-project/pull/207088)) reject a DO
WHILE, a missing or non-canonical loop, and an oversized `collapse`/`ordered`,
so a malformed nest does not reach lowering in a valid test.

Part of the feature work for #188820.

Assisted with Copilot.



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