| Issue |
170782
|
| Summary |
[Flang] [OpenMP] Inconsistent scope handling for unassociated DO variables in OpenMP SIMD Constructs (collapse(1))
|
| Labels |
flang
|
| Assignees |
|
| Reporter |
nakazawa-fj
|
```
Version of flang : 22.0.0git(7761a89e1230d8ccbb8e6ee15dc14ea74792a53c)/AArch64
```
In the attached program (`sample.f90`), when using `!$omp simd collapse(1)`, the `DO variable` of an inner loop that is not associated with　the `simd construct` is treated differently by flang. According to the `OpenMP specification` \
(version 6.0), such variables should be shared according to the `implicit data-sharing attribute rules`, but flang incorrectly treats them as `private`. This results in unexpected behavior
after the `simd region` ends.
Specification References:
OpenMP Application Programming Interface Version 6.0
6.4.5:collapse Clause
12.4:simd Construct
With `collapse(1)`, only the immediately following loop is associated with the `simd Construct`. Other loops remain unassociated
and should follow `implicit data-sharing rules`, which means they should be `shared`.
The following are the test program, Flang, Gfortran and ifx compilation/execution results.
sample.f90
```fortran
program test
implicit none
integer :: n, k
integer :: omp_get_thread_num
!$omp parallel
k = 0
!$omp simd collapse(1)
do n = 3, 3
do k = 4, 4
print '(1x,"In k= ",i2," k-address:",z16.16," thread num: ",i2)', k, loc(k), omp_get_thread_num()
end do
end do
!$omp end simd
if (omp_get_thread_num() == 0) then
print '(1x,"Out k= ",i2," k-address:",z16.16," thread num: 0")', k, loc(k)
if (k /= 5) print *, 'error', k
end if
!$omp end parallel
end program
```
```
$ export OMP_NUM_THREADS=4; flang -fopenmp sample.f90; ./a.out
In k= 4 k-address:0000FFFF91C2C610 thread num: 2
In k= 4 k-address:0000FFFF925CC5D0 thread num: 1
In k= 4 k-address:0000FFFF9128C650 thread num: 3
In k= 4 k-address:0000FFFFF2440CA0 thread num: 0
Out k= 0 k-address:0000FFFFF2440D08 thread num: 0
error 0
```
```
$ export OMP_NUM_THREADS=4; gfortran -fopenmp sample.f90; ./a.out
In k= 4 k-address:0000FFFFE3809588 thread num: 0
Out k= 5 k-address:0000FFFFE3809588 thread num: 0
In k= 4 k-address:0000FFFFB170B768 thread num: 2
In k= 4 k-address:0000FFFFB20AA768 thread num: 1
In k= 4 k-address:0000FFFFB0D6C768 thread num: 3
```
```
$ export OMP_NUM_THREADS=4; ifx -qopenmp sample.f90; ./a.out
In k= 4 k-address:00007FF7CF37EA1C thread num: 1
In k= 4 k-address:00007FF7CE37AA9C thread num: 3
In k= 4 k-address:00007FFCDC13791C thread num: 0
In k= 4 k-address:00007FF7CEB7CA1C thread num: 2
Out k= 0 k-address:00007FFCDC137920 thread num: 0
error 0
```
Expected:
The unassociated inner DO variable `k` is shared.
After the SIMD region ends, `k` should be 5 per Fortran DO rules; `loc(k)` for `Out` is consistent for the shared outer `k`.
Actual:
flang/ifx: `k` behaves as private; `Out k` remains 0 (unexpected), `loc(k)` differs across threads in `In` prints,
indicating per-thread privatization; the `Out` value is not updated to 5.
gfortran: `k` behaves as shared; `Out k` is 5 (expected), `loc(k)` for `Out` matches the shared outer `k`.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
