[flang][hlfir] get rid of box when translating scalars to extented value

[jeanPerier]

The code in translateToExtendedValue(hlfir::Entity) was not getting rid of the fir.box for scalars because isSimplyContiguous() returned false for them.

This created issues downstream because utilities using fir::ExtendedValue were not implemented to work with intrinsic scalars fir.box.

fir.box of intrinsic scalars are not very commonly used as hlfir::Entity but they are allowed and should work where accepted.

[llvmbot]

@llvm/pr-subscribers-flang-fir-hlfir

Author: None (jeanPerier)

Changes

The code in translateToExtendedValue(hlfir::Entity) was not getting rid of the fir.box for scalars because isSimplyContiguous() returned false for them.

This created issues downstream because utilities using fir::ExtendedValue were not implemented to work with intrinsic scalars fir.box.

fir.box of intrinsic scalars are not very commonly used as hlfir::Entity but they are allowed and should work where accepted.

---

Full diff: https://github.com/llvm/llvm-project/pull/125059.diff

2 Files Affected:

• (modified) flang/include/flang/Optimizer/Builder/HLFIRTools.h (+1-1)
• (modified) flang/test/HLFIR/assign-codegen.fir (+10)

diff --git a/flang/include/flang/Optimizer/Builder/HLFIRTools.h b/flang/include/flang/Optimizer/Builder/HLFIRTools.h
index 0684ad0f926ec9..d8785969bb7247 100644
--- a/flang/include/flang/Optimizer/Builder/HLFIRTools.h
+++ b/flang/include/flang/Optimizer/Builder/HLFIRTools.h
@@ -125,7 +125,7 @@ class Entity : public mlir::Value {
   bool isSimplyContiguous() const {
     // If this can be described without a fir.box in FIR, this must
     // be contiguous.
-    if (!hlfir::isBoxAddressOrValueType(getFirBase().getType()))
+    if (!hlfir::isBoxAddressOrValueType(getFirBase().getType()) || isScalar())
       return true;
     // Otherwise, if this entity has a visible declaration in FIR,
     // or is the dereference of an allocatable or contiguous pointer
diff --git a/flang/test/HLFIR/assign-codegen.fir b/flang/test/HLFIR/assign-codegen.fir
index 581d1ab0e7739c..5e48784284a8b4 100644
--- a/flang/test/HLFIR/assign-codegen.fir
+++ b/flang/test/HLFIR/assign-codegen.fir
@@ -427,3 +427,13 @@ func.func @test_upoly_expr_assignment(%arg0: !fir.class<!fir.array<?xnone>> {fir
 // CHECK:           }
 // CHECK:           return
 // CHECK:         }
+
+func.func @test_scalar_box(%arg0: f32, %arg1: !fir.box<!fir.ptr<f32>>) {
+  hlfir.assign %arg0 to %arg1 : f32, !fir.box<!fir.ptr<f32>>
+  return
+}
+// CHECK-LABEL:   func.func @test_scalar_box(
+// CHECK-SAME:                               %[[VAL_0:.*]]: f32,
+// CHECK-SAME:                               %[[VAL_1:.*]]: !fir.box<!fir.ptr<f32>>) {
+// CHECK:           %[[VAL_2:.*]] = fir.box_addr %[[VAL_1]] : (!fir.box<!fir.ptr<f32>>) -> !fir.ptr<f32>
+// CHECK:           fir.store %[[VAL_0]] to %[[VAL_2]] : !fir.ptr<f32>

[razvanlupusoru]

Thank you!

[tblah]

I getting two gfortran tests segfaulting at runtime. The bisect pointed me to this patch:

• https://github.com/llvm/llvm-test-suite/blob/main/Fortran/gfortran/regression/optional_absent_4.f90
• https://github.com/llvm/llvm-test-suite/blob/main/Fortran/gfortran/regression/optional_absent_5.f90

Reproduce with flang -o /tmp/out optional_absent_4.f90 && /tmp/out

Please could you take a look?

[DavidSpickett]

Just got the same result myself. This is failing on our bots but because they had existing failures, no notification went out.

https://lab.llvm.org/buildbot/#/builders/198/builds/1633

[DavidSpickett]

Smaller version:

module z
  implicit none
contains
  subroutine sum_2 (input, res, mask)
    logical, intent(in), optional :: mask
    integer, intent(in) :: input(:,:)
    integer, dimension(:), intent(out) :: res
    res = sum (input, dim=1, mask=mask)
  end subroutine sum_2
end module z

program main
  use z
  implicit none
  integer :: i2(2,3) = reshape([1,2,4,8,16,32], [2,3])
  integer, dimension(3) :: res3
  res3 = -2
  call sum_2 (i2, res3)
  if (any (res3 /= [3, 12, 48])) stop 2
end program main

[jeanPerier]

I getting two gfortran tests segfaulting at runtime. The bisect pointed me to this patch:

Thanks Tom and David, I saw that too this morning, I have a fix here: #125215