[flang] Unlimited pointer assignment within FORALL caused FIR verification failure

[DanielCChen]

Consider the following code:

program fpAssgn025a

    type anyContainer
        class (*), pointer :: data => null()
        !character*1, pointer :: data => null()    ! workaround 1
    end type

    type (anyContainer) :: co_a1(10)

    character*1, target :: ch1(10)

!    allocate (c1(10))

    forall (i=1:10)
!    do i = 1, 10                                   ! workaround 2
        co_a1(i)%data => ch1(i)
!    end do
    end forall

end

Flang has

error: loc("t.f":16:9): 'fir.embox' op LEN parameters require CHARACTER or derived type
error: verification of lowering to FIR failed

If I apply either workaround marked in the code (1. don't use unlimited poly; 2. use do loop instead of forall), it compiles successfully.

Update (Oct 22, 2025): After the fix of PR #164279 , the code in this issue now fails at:

fir.convert' op invalid type conversion'!fir.box<!fir.ptr<!fir.char<1>>>' / '!fir.class<!fir.ptr<none>>'

[DanielCChen]

@jeanPerier
The error for fir.convert at the above seems incorrect as it should be able to convert a PointerType to a polymorphic type. Could you please confirm if it should be allowed?

If so, I am planning to post a PR to add

(fir::isPointerType(inType) && fir::isPolymorphicType(outType))

in canBeConverted to allow it.

[jeanPerier]

There is a rational behind that restriction which is that !fir.class are implemented with descriptors with an addendum while !fir.box<!fir.ptr<!fir.char<1>>> are not.

So with the current implementation of fir.box/class and fir.convert, any fir.store using the the convert result would end-up doing bad memory read of the input descriptor when copying it.

So it is not possible to just lift the restriction, these type of conversions should be done with fir.rebox

I think the solution is to insert a rebox to the lhs box type when fir::ConvertOp::canBeConverted(rhsBoxType, lhsBoxTtype) is false after the code you modified in Forall pointer lowering.

[DanielCChen]

There is a rational behind that restriction which is that !fir.class are implemented with descriptors with an addendum while !fir.box<!fir.ptr<!fir.char<1>>> are not.

So with the current implementation of fir.box/class and fir.convert, any fir.store using the the convert result would end-up doing bad memory read of the input descriptor when copying it.

So it is not possible to just lift the restriction, these type of conversions should be done with fir.rebox

I think the solution is to insert a rebox to the lhs box type when fir::ConvertOp::canBeConverted(rhsBoxType, lhsBoxTtype) is false after the code you modified in Forall pointer lowering.

@jeanPerier
Thanks for the explanation! The restriction makes sense now. I will try your suggestion.
Just to clarify, for the LIT test I added for PR #164279 , we generate code like the following for pointer assignment inside FORALL when LHS is regular polymorphic (not unlimited) and the RHS is not polymorphic.

! CHECK: %[[V_46:[0-9]+]] = fir.convert %[[V_45]] : (!fir.box<!fir.ptr<!fir.array<10x!fir.type<_QFforallpolymorphicTdt1{dt:!fir.type<_QFforallpolymorphicTdt{ptr:!fir.class<!fir.ptr<!fir.array<?x!fir.type<_QFforallpolymorphicTdt>>>>}>}>>>>) -> !fir.class<!fir.ptr<!fir.array<?x!fir.type<_QFforallpolymorphicTdt{ptr:!fir.class<!fir.ptr<!fir.array<?x!fir.type<_QFforallpolymorphicTdt>>>>}>>>>
! CHECK: fir.store %[[V_46]] to %[[V_44]] : !fir.ref<!fir.class<!fir.ptr<!fir.array<?x!fir.type<_QFforallpolymorphicTdt{ptr:!fir.class<!fir.ptr<!fir.array<?x!fir.type<_QFforallpolymorphicTdt>>>>}>>>>>

This seems a fix.box<fir.ptr>... to fix.class<!fix.ptr>... conversion as well. Would this also need a rebox?

Update: this passed verify() as the inType Is deemed isBoxedRecordType and the outType is isPolymorphicType.

[jeanPerier]

Update: this passed verify() as the inType Is deemed isBoxedRecordType and the outType is isPolymorphicType.

Yes because all derived type descriptor always have an addendum, there is currently no difference between a class<type> and a box<type>.

So you only need the rebox for the case where the LHS is unlimited polymorphic and the RHS descriptor layout may not be compatible.

[DanielCChen]

I see. That makes sense.
Thanks again for the explanation!