================
@@ -172,7 +170,8 @@ static bool isStrictSubset(const VariantMatchInfo &VMI0,
for (unsigned Bit : VMI0.RequiredTraits.set_bits())
if (!VMI1.RequiredTraits.test(Bit))
return false;
- if (!isSubset<TraitProperty>(VMI0.ConstructTraits, VMI1.ConstructTraits))
+ if (!isOrderedSubset<TraitProperty>(VMI0.ConstructTraits,
----------------
MattPD wrote:
The `RequiredTraits.count()` guard makes `isStrictSubset` return false when
only a repeated construct distinguishes the selectors. The new highest-valued
match scoring can make such selectors tie, so the guard now determines whether
the strict-subset tie-break applies.
Compiling the following reproducer with `flang -fopenmp -fopenmp-version=52 -S
-emit-llvm repro.f90 -o -` emits TASKYIELD at b165f599 instead of the TASKWAIT
selected by the [parent
commit](https://github.com/llvm/llvm-project/commit/81b82b8012a2bf9a1f1822af10778922b4646fef):
```fortran
subroutine subset_first()
integer :: i
!$omp parallel
!$omp parallel do
do i = 1, 2
!$omp metadirective &
!$omp& when(construct={parallel do}, &
!$omp& implementation={vendor(score(1): llvm)}: taskyield) &
!$omp& when(construct={parallel, parallel do}, &
!$omp& implementation={vendor(llvm)}: taskwait)
end do
!$omp end parallel do
!$omp end parallel
end subroutine
```
Flang expands the construct selectors to `[parallel, do]` and `[parallel,
parallel, do]`. Both now score 8. Their `RequiredTraits` counts are equal, so
clause order decides the result. Reversing the clauses selects TASKWAIT.
Could the existing tie-break also recognize a strict subset from the ordered
construct sequence, with tests covering both clause orders? The fix need not
apply the strict-subset rule outside ties.
https://github.com/llvm/llvm-project/pull/231668
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
