[Flang][OpenMP] Compilation errors in DO SIMD construct with FIRSTPRIVATE clause

[ohno-fj]

Version of flang : 22.0.0(5892a2beec4eaf77922211a9aec70faa0ae287db)/AArch64

Specifying FIRSTPRIVATE clause in DO SIMD construct results in a compilation error (FIRSTPRIVATE cannot be used with SIMD).

According to OpenMP 6.0: 12.4 simd Construct and OpenMP 6.0: 13.6.2 do Construct,

• FIRSTPRIVATE clause cannot be specified in SIMD construct.
• FIRSTPRIVATE clause can be specified in DO construct.

Also, according to OpenMP 6.0: 19.2 Clauses on Compound Constructs,

• !$omp do simd is Compound Construct that combines DO construct (worksharing construct) and SIMD construct.
• DO construct is worksharing construct and FIRSTPRIVATE clause is applied.
• SIMD construct is not a worksharing construct, so FIRSTPRIVATE clause is not applied.

So it seems likely that FIRSTPRIVATE clause can be specified in DO SIMD construct, which is Compound Construct.
Thus, the Flang compilation error messages appear to violate the OpenMP specification.
Please point out any errors in my interpretation of OpenMP.

The following are the test program, Flang, Gfortran and ifx compilation/execution results.

test01.f90:

subroutine sub
  integer a
  a=10
!$omp do simd lastprivate(a) firstprivate(a)
  do i=1,1
     if (a/=10) print *,11
     a=20
  end do
!$omp end do simd
  if (a/=20) print *,21
end subroutine sub

program main
  integer a
  a=10
!$omp do simd lastprivate(a) firstprivate(a)
  do i=1,1
     if (a/=10) print *,11
     a=20
  end do
!$omp end do simd
  if (a/=20) print *,21
  call sub
  print *,'pass'
end program main

$ flang -fopenmp test01.f90
error: loc("/work/home/ohno/CT/test/fort/tp/reproducerJ/FJ-Compiler-Test-Suite/34/test01.f90":4:7): FIRSTPRIVATE cannot be used with SIMD
error: loc("/work/home/ohno/CT/test/fort/tp/reproducerJ/FJ-Compiler-Test-Suite/34/test01.f90":16:7): FIRSTPRIVATE cannot be used with SIMD
error: verification of lowering to FIR failed
$

$ gfortran -fopenmp test01.f90
test01.f90:4:44:

    4 | !$omp do simd lastprivate(a) firstprivate(a)
      |                                            ^
Error: firstprivate variable ‘a’ is private in outer context
test01.f90:4:44: Error: lastprivate variable ‘a’ is private in outer context
test01.f90:16:44:

   16 | !$omp do simd lastprivate(a) firstprivate(a)
      |                                            ^
Error: firstprivate variable ‘a’ is private in outer context
test01.f90:16:44: Error: lastprivate variable ‘a’ is private in outer context
$

$ export OMP_NUM_THREADS=2; ifx -qopenmp test01.f90; ./a.out
 pass
$

[llvmbot]

@llvm/issue-subscribers-flang-ir

Author: None (ohno-fj)

Details

``` Version of flang : 22.0.0(5892a2b)/AArch64 ```

Specifying FIRSTPRIVATE clause in DO SIMD construct results in a compilation error (FIRSTPRIVATE cannot be used with SIMD).

According to OpenMP 6.0: 12.4 simd Construct and OpenMP 6.0: 13.6.2 do Construct,

• FIRSTPRIVATE clause cannot be specified in SIMD construct.
• FIRSTPRIVATE clause can be specified in DO construct.

Also, according to OpenMP 6.0: 19.2 Clauses on Compound Constructs,

• !$omp do simd is Compound Construct that combines DO construct (worksharing construct) and SIMD construct.
• DO construct is worksharing construct and FIRSTPRIVATE clause is applied.
• SIMD construct is not a worksharing construct, so FIRSTPRIVATE clause is not applied.

So it seems likely that FIRSTPRIVATE clause can be specified in DO SIMD construct, which is Compound Construct.
Thus, the Flang compilation error messages appear to violate the OpenMP specification.
Please point out any errors in my interpretation of OpenMP.

The following are the test program, Flang, Gfortran and ifx compilation/execution results.

test01.f90:

subroutine sub
  integer a
  a=10
!$omp do simd lastprivate(a) firstprivate(a)
  do i=1,1
     if (a/=10) print *,11
     a=20
  end do
!$omp end do simd
  if (a/=20) print *,21
end subroutine sub

program main
  integer a
  a=10
!$omp do simd lastprivate(a) firstprivate(a)
  do i=1,1
     if (a/=10) print *,11
     a=20
  end do
!$omp end do simd
  if (a/=20) print *,21
  call sub
  print *,'pass'
end program main

$ flang -fopenmp test01.f90
error: loc("/work/home/ohno/CT/test/fort/tp/reproducerJ/FJ-Compiler-Test-Suite/34/test01.f90":4:7): FIRSTPRIVATE cannot be used with SIMD
error: loc("/work/home/ohno/CT/test/fort/tp/reproducerJ/FJ-Compiler-Test-Suite/34/test01.f90":16:7): FIRSTPRIVATE cannot be used with SIMD
error: verification of lowering to FIR failed
$

$ gfortran -fopenmp test01.f90
test01.f90:4:44:

    4 | !$omp do simd lastprivate(a) firstprivate(a)
      |                                            ^
Error: firstprivate variable ‘a’ is private in outer context
test01.f90:4:44: Error: lastprivate variable ‘a’ is private in outer context
test01.f90:16:44:

   16 | !$omp do simd lastprivate(a) firstprivate(a)
      |                                            ^
Error: firstprivate variable ‘a’ is private in outer context
test01.f90:16:44: Error: lastprivate variable ‘a’ is private in outer context
$

$ export OMP_NUM_THREADS=2; ifx -qopenmp test01.f90; ./a.out
 pass
$

[kparzysz]

Hey Sergio, could you take a look at this? We end up with private_syms (with the firstprivate sym) on the SimdOp, I'm not sure if that's correct.

"omp.wsloop"(%2#0) <{operandSegmentSizes = array<i32: 0, 0, 0, 0, 1, 0, 0>, private_needs_barrier, private_syms = [@_QFsubEa_firstprivate_i32]}> ({
^bb0(%arg0: !fir.ref<i32>):
  "omp.simd"(%arg0, %4#0) <{operandSegmentSizes = array<i32: 0, 0, 0, 0, 0, 2, 0>, private_needs_barrier, private_syms = [@_QFsubEa_firstprivate_i32, @_QFsubEi_private_i32]}> ({
  [...]

[skatrak]

There's been some recent work enabling privatization for composite DO SIMD constructs (#150979). This does look like a potential edge case that's not being handled properly in lowering. My guess is the problem might be with the data sharing processor for the SIMD leaf construct in genCompositeDoSimd. CC: @mrkajetanp, @tblah.

[mrkajetanp]

I think the issue is even narrower than the description suggests, namely if I write something like

 integer a, b
!$omp do simd lastprivate(a) firstprivate(b)

then I get the result I expect, i.e. the program works and firstprivate(b) is only attached to the outer wsloop but not the inner simd.

    omp.wsloop private(@_QFEa_private_i32 %2 -> %arg0, @_QFEb_firstprivate_f32 %4 -> %arg1 : !fir.ref<i32>, !fir.ref<f32>) {
      omp.simd private(@_QFEa_private_i32 %arg0 -> %arg2, @_QFEi_private_i32 %6 -> %arg3 : !fir.ref<i32>, !fir.ref<i32>) {
        omp.loop_nest (%arg4) : i32 = (%c1_i32) to (%c1_i32) inclusive step (%c1_i32) {

I think there must be a bug in the logic for collecting symbols for privatisation. If the symbol names overlap as they do in the test program (both 'a'), then even though the only clause attached to simd at the point of the lowering is lastprivate(a) (I checked simdItem->clauses), the symbol 'a' gets collected by simdItemDSP.getDelayedPrivSymbols() as lastprivate.
For the test program in question a call to simdItemDSP.getDelayedPrivSymbols() returns:

a (OmpFirstPrivate, OmpLastPrivate, OmpExplicit): HostAssoc
i (OmpLinear, OmpPreDetermined): HostAssoc

while for the test program with non-overlapping variable names it's:

a (OmpLastPrivate, OmpExplicit): HostAssoc
i (OmpLinear, OmpPreDetermined): HostAssoc

[mrkajetanp]

I've looked into it a bit more, the problem is exactly that we have one symbol 'a' which is both firstprivate and lastprivate.
Because it's lastprivate, it gets picked up for delayed privatisation by the logic in DataSharingProcessor.
Because it's also firstprivate, it then fails the verify step in OpenMP mlir because firstprivate symbols are not allowed on simd.

On a different note, is having a symbol like this even supposed to be allowed? gfortran doesn't seem to compile this either.