[flang][openmp] Incorrect perfect nest detection if using continues instead of enddo.

[scamp-nvidia]

Consider the following example, carved down from an OpenMP SPEC test:

! omp_compile_test.f90
program omp_compile_test
  implicit none
  integer, parameter :: N = 8
  real(8) :: PN(N,N) = 1.0d0, UN(N,N) = 2.0d0, VN(N,N) = 3.0d0
  real(8) :: P = 0.0d0, U = 0.0d0, V = 0.0d0
  integer :: i, j

!$omp target teams map(to:PN,UN,VN) map(tofrom:P,U,V)
!$omp loop collapse(2) reduction(+:P,U,V)
  do 4500 j = 1, N
    do 3500 i = 1, N
      P = P + abs(PN(i,j))
      U = U + abs(UN(i,j))
      V = V + abs(VN(i,j))
    3500 CONTINUE
   4500 CONTINUE
!$omp end target teams
end program

This has prevously worked, but began failing between d96c32c and fad2a4c with the following message:

scamp$ flang test.F90 -fopenmp -o test
error: Semantic errors in test.F90
./test.F90:10:7: error: Canonical loop nest must be perfectly nested.
  !$omp loop collapse(2) reduction(+:P,U,V)
        ^^^^

Reading the changes between those two hashes, I immediately key in on #160283 - which appears to work on perfectly nested OpenMP Flang structures. Tagging @Meinersbur for investigation, please.

[eugeneepshteyn]

Similar error happens with https://github.com/fujitsu/compiler-test-suite/blob/main/Fortran/0966/0966_0000.f90

[Meinersbur]

Fixed by #161554

@eugeneepshteyn https://github.com/fujitsu/compiler-test-suite/blob/main/Fortran/0966/0966_0000.f90?rgh-link-date=2025-10-01T15%3A25%3A31.000Z is indeed a non-perfectly nested loop, which is currently unsupported. Without the error, Flang will just not emit the write at all, which is a miscompilation.

[klausler]

Should/do cases like this work, with multiple loops ending on the same statement?

  do 10 j = 1, N
    do 10 i = 1, N
      P = P + abs(PN(i,j))
      U = U + abs(UN(i,j))
10    V = V + abs(VN(i,j))

[scamp-nvidia]

@klausler - quick testing shows that example at least doesn't get complained about, so I think it works.

[llvmbot]

@llvm/issue-subscribers-flang-frontend

Author: None (scamp-nvidia)

Consider the following example, carved down from an OpenMP SPEC test: ``` ! omp_compile_test.f90 program omp_compile_test implicit none integer, parameter :: N = 8 real(8) :: PN(N,N) = 1.0d0, UN(N,N) = 2.0d0, VN(N,N) = 3.0d0 real(8) :: P = 0.0d0, U = 0.0d0, V = 0.0d0 integer :: i, j

!$omp target teams map(to:PN,UN,VN) map(tofrom:P,U,V)
!$omp loop collapse(2) reduction(+:P,U,V)
do 4500 j = 1, N
do 3500 i = 1, N
P = P + abs(PN(i,j))
U = U + abs(UN(i,j))
V = V + abs(VN(i,j))
3500 CONTINUE
4500 CONTINUE
!$omp end target teams
end program

This has prevously worked, but began failing between d96c32cdaf588903917a9e7db172729759e34c9d and fad2a4c00db68ee10b506b05531cdc6458d9be37 with the following message: 

scamp$ flang test.F90 -fopenmp -o test
error: Semantic errors in test.F90
./test.F90:10:7: error: Canonical loop nest must be perfectly nested.
!$omp loop collapse(2) reduction(+:P,U,V)
^^^^

Reading the changes between those two hashes, I immediately key in on #<!-- -->160283 - which appears to work on perfectly nested OpenMP Flang structures. Tagging @<!-- -->Meinersbur for investigation, please. 


</details>