[flang-commits] [flang] [flang][OpenMP] Lower DO and SIMD variants in metadirectives (PR #210810)

[via flang-commits]

================
@@ -1352,6 +1358,14 @@ void OmpStructureChecker::Enter(const parser::OpenMPDeclarativeConstruct &x) {
   EnterDirectiveNest(DeclarativeNest);
 }
 
+void OmpStructureChecker::Enter(const parser::OpenACCDeclarativeConstruct &) {
----------------
MattPD wrote:

The new `Enter(parser::OpenACCDeclarativeConstruct)` and `Enter(parser::OpenACCRoutineConstruct)` hooks diagnose the OpenACC cases, but plain Fortran reaches the same abort. A `FORMAT` statement between the metadirective and its loop is accepted by `-fsyntax-only` and then aborts in lowering:

```fortran
subroutine fmt_sib(n, a)
  integer :: n, a(n), i
  !$omp metadirective when(implementation={vendor(llvm)}: do) otherwise(nothing)
100 format(1x,i0)
  do i = 1, n
    a(i) = i
  end do
end subroutine
```

```
not yet implemented: loop-associated METADIRECTIVE without associated DO   (OpenMP.cpp:6316)
```

Two conditions combine:

- `FORMAT` is legal in a specification part, so the parser classifies the metadirective as an `OpenMPDeclarativeConstruct` there. `Enter(parser::ExecutionPartConstruct)` then never sees an interposed construct.
- `FORMAT` also produces an evaluation of its own, which `spliceAssociatedDoEval` does not skip. It skips only end statements and `parser::CompilerDirective`.

`-fdebug-dump-pft` shows the difference against an interposed `DATA` statement, which is equally legal in a specification part but lowers cleanly. With `FORMAT`:

```
  OpenMPDeclarativeConstruct
  FormatStmt: 100 format(1x,i0)
  <<DoConstruct>> -> 4
```

With `DATA`:

```
  OpenMPDeclarativeConstruct
  <<DoConstruct>> -> 4
```

So specification-part legality alone is not the discriminator. The construct also has to produce an evaluation of its own, which is what puts it between the metadirective and the loop. `ENTRY` meets both conditions and behaves like `FORMAT`, while an ordinary assignment is diagnosed correctly.

Would deriving the semantic check and the splice from one shared predicate close the category at once? `FORMAT` looks like it should join the skip list, whereas `ENTRY` is a real entry point, so a diagnostic may be the right answer there.

https://github.com/llvm/llvm-project/pull/210810