[Flang][OpenMP] default(firstprivate) failing with parallel do

[NimishMishra]

default(firstprivate) fails to privatize when used with parallel do. Reproducer:

program sample
  
   integer :: x
   x = 10
   !$omp parallel do default(firstprivate) num_threads(50)
      do i = 1, 50
        x = x + 1
      end do
   !$omp end parallel do
   print *, x
end program sample

Current IR (mainly fir.store %10 to %1 : !fir.ref<i32>):

    %0 = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
    %1 = fir.alloca i32 {bindc_name = "x", uniq_name = "_QFEx"}
    %c10_i32 = arith.constant 10 : i32
    fir.store %c10_i32 to %1 : !fir.ref<i32>
    %c50_i32 = arith.constant 50 : i32
    omp.parallel num_threads(%c50_i32 : i32) {
      %8 = fir.alloca i32 {adapt.valuebyref, pinned}
      %c1_i32 = arith.constant 1 : i32
      %c50_i32_1 = arith.constant 50 : i32
      %c1_i32_2 = arith.constant 1 : i32
      omp.wsloop for  (%arg0) : i32 = (%c1_i32) to (%c50_i32_1) inclusive step (%c1_i32_2) {
        fir.store %arg0 to %8 : !fir.ref<i32>
        %9 = fir.load %1 : !fir.ref<i32>
        %c1_i32_3 = arith.constant 1 : i32
        %10 = arith.addi %9, %c1_i32_3 : i32
        fir.store %10 to %1 : !fir.ref<i32>
        omp.yield
      }
      omp.terminator
    }

Expected IR should be as follows (mainly fir.store %12 to %9 : !fir.ref<i32>):

    %0 = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
    %1 = fir.alloca i32 {bindc_name = "x", uniq_name = "_QFEx"}
    %c10_i32 = arith.constant 10 : i32
    fir.store %c10_i32 to %1 : !fir.ref<i32>
    %c50_i32 = arith.constant 50 : i32
    omp.parallel num_threads(%c50_i32 : i32) {
      %8 = fir.alloca i32 {adapt.valuebyref, pinned}
      %9 = fir.alloca i32 {bindc_name = "x", pinned, uniq_name = "_QFEx"}
      %10 = fir.load %1 : !fir.ref<i32>
      fir.store %10 to %9 : !fir.ref<i32>
      %c1_i32 = arith.constant 1 : i32
      %c50_i32_1 = arith.constant 50 : i32
      %c1_i32_2 = arith.constant 1 : i32
      omp.wsloop for  (%arg0) : i32 = (%c1_i32) to (%c50_i32_1) inclusive step (%c1_i32_2) {
        fir.store %arg0 to %8 : !fir.ref<i32>
        %11 = fir.load %9 : !fir.ref<i32>
        %c1_i32_3 = arith.constant 1 : i32
        %12 = arith.addi %11, %c1_i32_3 : i32
        fir.store %12 to %9 : !fir.ref<i32>
        omp.yield
      }
      omp.terminator
    }

[llvmbot]

@llvm/issue-subscribers-bug

Author: None (NimishMishra)

`default(firstprivate)` fails to privatize when used with parallel do. Reproducer:

program sample
  
   integer :: x
   x = 10
   !$omp parallel do default(firstprivate) num_threads(50)
      do i = 1, 50
        x = x + 1
      end do
   !$omp end parallel do
   print *, x
end program sample

Current IR (mainly fir.store %10 to %1 : !fir.ref<i32>):

    %0 = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
    %1 = fir.alloca i32 {bindc_name = "x", uniq_name = "_QFEx"}
    %c10_i32 = arith.constant 10 : i32
    fir.store %c10_i32 to %1 : !fir.ref<i32>
    %c50_i32 = arith.constant 50 : i32
    omp.parallel num_threads(%c50_i32 : i32) {
      %8 = fir.alloca i32 {adapt.valuebyref, pinned}
      %c1_i32 = arith.constant 1 : i32
      %c50_i32_1 = arith.constant 50 : i32
      %c1_i32_2 = arith.constant 1 : i32
      omp.wsloop for  (%arg0) : i32 = (%c1_i32) to (%c50_i32_1) inclusive step (%c1_i32_2) {
        fir.store %arg0 to %8 : !fir.ref<i32>
        %9 = fir.load %1 : !fir.ref<i32>
        %c1_i32_3 = arith.constant 1 : i32
        %10 = arith.addi %9, %c1_i32_3 : i32
        fir.store %10 to %1 : !fir.ref<i32>
        omp.yield
      }
      omp.terminator
    }

Expected IR should be as follows (mainly fir.store %12 to %9 : !fir.ref<i32>):

    %0 = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
    %1 = fir.alloca i32 {bindc_name = "x", uniq_name = "_QFEx"}
    %c10_i32 = arith.constant 10 : i32
    fir.store %c10_i32 to %1 : !fir.ref<i32>
    %c50_i32 = arith.constant 50 : i32
    omp.parallel num_threads(%c50_i32 : i32) {
      %8 = fir.alloca i32 {adapt.valuebyref, pinned}
      %9 = fir.alloca i32 {bindc_name = "x", pinned, uniq_name = "_QFEx"}
      %10 = fir.load %1 : !fir.ref<i32>
      fir.store %10 to %9 : !fir.ref<i32>
      %c1_i32 = arith.constant 1 : i32
      %c50_i32_1 = arith.constant 50 : i32
      %c1_i32_2 = arith.constant 1 : i32
      omp.wsloop for  (%arg0) : i32 = (%c1_i32) to (%c50_i32_1) inclusive step (%c1_i32_2) {
        fir.store %arg0 to %8 : !fir.ref<i32>
        %11 = fir.load %9 : !fir.ref<i32>
        %c1_i32_3 = arith.constant 1 : i32
        %12 = arith.addi %11, %c1_i32_3 : i32
        fir.store %12 to %9 : !fir.ref<i32>
        omp.yield
      }
      omp.terminator
    }

[llvmbot]

@llvm/issue-subscribers-flang-ir

Author: None (NimishMishra)

`default(firstprivate)` fails to privatize when used with parallel do. Reproducer:

program sample
  
   integer :: x
   x = 10
   !$omp parallel do default(firstprivate) num_threads(50)
      do i = 1, 50
        x = x + 1
      end do
   !$omp end parallel do
   print *, x
end program sample

Current IR (mainly fir.store %10 to %1 : !fir.ref<i32>):

    %0 = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
    %1 = fir.alloca i32 {bindc_name = "x", uniq_name = "_QFEx"}
    %c10_i32 = arith.constant 10 : i32
    fir.store %c10_i32 to %1 : !fir.ref<i32>
    %c50_i32 = arith.constant 50 : i32
    omp.parallel num_threads(%c50_i32 : i32) {
      %8 = fir.alloca i32 {adapt.valuebyref, pinned}
      %c1_i32 = arith.constant 1 : i32
      %c50_i32_1 = arith.constant 50 : i32
      %c1_i32_2 = arith.constant 1 : i32
      omp.wsloop for  (%arg0) : i32 = (%c1_i32) to (%c50_i32_1) inclusive step (%c1_i32_2) {
        fir.store %arg0 to %8 : !fir.ref<i32>
        %9 = fir.load %1 : !fir.ref<i32>
        %c1_i32_3 = arith.constant 1 : i32
        %10 = arith.addi %9, %c1_i32_3 : i32
        fir.store %10 to %1 : !fir.ref<i32>
        omp.yield
      }
      omp.terminator
    }

Expected IR should be as follows (mainly fir.store %12 to %9 : !fir.ref<i32>):

    %0 = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
    %1 = fir.alloca i32 {bindc_name = "x", uniq_name = "_QFEx"}
    %c10_i32 = arith.constant 10 : i32
    fir.store %c10_i32 to %1 : !fir.ref<i32>
    %c50_i32 = arith.constant 50 : i32
    omp.parallel num_threads(%c50_i32 : i32) {
      %8 = fir.alloca i32 {adapt.valuebyref, pinned}
      %9 = fir.alloca i32 {bindc_name = "x", pinned, uniq_name = "_QFEx"}
      %10 = fir.load %1 : !fir.ref<i32>
      fir.store %10 to %9 : !fir.ref<i32>
      %c1_i32 = arith.constant 1 : i32
      %c50_i32_1 = arith.constant 50 : i32
      %c1_i32_2 = arith.constant 1 : i32
      omp.wsloop for  (%arg0) : i32 = (%c1_i32) to (%c50_i32_1) inclusive step (%c1_i32_2) {
        fir.store %arg0 to %8 : !fir.ref<i32>
        %11 = fir.load %9 : !fir.ref<i32>
        %c1_i32_3 = arith.constant 1 : i32
        %12 = arith.addi %11, %c1_i32_3 : i32
        fir.store %12 to %9 : !fir.ref<i32>
        omp.yield
      }
      omp.terminator
    }

[luporl]

It seems the issue happens whenever there is a variable that is referenced only inside a nested evaluation.
Example:

program test
  integer :: x
  x = 5

  !$omp parallel default(firstprivate)
    if (.true.) then
      x = 10
    end if
  !$omp end parallel
  print *, x
end program

In the program above, x won't be privatized and 10 will be printed, instead of 5.

[NimishMishra]

In the program above, x won't be privatized and 10 will be printed, instead of 5.

Thanks @luporl for the test case. I have a PR at #72510 for which this test case gives 5. Below is the IR (the privatization is proceeding as expected):

    %0 = fir.alloca i32 {bindc_name = "x", uniq_name = "_QFEx"}
    %1:2 = hlfir.declare %0 {uniq_name = "_QFEx"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
    %c5_i32 = arith.constant 5 : i32
    hlfir.assign %c5_i32 to %1#0 : i32, !fir.ref<i32>
    omp.parallel {
      %8 = fir.alloca i32 {bindc_name = "x", pinned, uniq_name = "_QFEx"}
      %9:2 = hlfir.declare %8 {uniq_name = "_QFEx"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
      %10 = fir.load %1#0 : !fir.ref<i32>
      hlfir.assign %10 to %9#0 temporary_lhs : i32, !fir.ref<i32>
      %true = arith.constant true
      fir.if %true {
        %c10_i32_0 = arith.constant 10 : i32
        hlfir.assign %c10_i32_0 to %9#0 : i32, !fir.ref<i32>
      } else {
      }
      omp.terminator
    }

[luporl]

Indeed, PR #72510 seems to be a good fix for this issue, thanks.

For context, I stumbled into this issue while investigating #64480.
It seems the handling of OpenMP implicit data sharing attributes (OpenMP 4.5, subclause 2.15.1.1) is incomplete in Flang.
I'm trying to handle the missing cases.

[NimishMishra]

Fixed with merging the above PR