| Issue |
241462
|
| Summary |
[flang][OpenMP] User-defined default mapper ignored or replaced, depending on DECLARE MAPPER order
|
| Labels |
flang
|
| Assignees |
chichunchen
|
| Reporter |
chichunchen
|
### Summary
When a `DECLARE MAPPER` maps a component whose type has a user-defined default mapper, Flang attaches that mapper only if its `omp.declare_mapper` op already exists in the module. Mappers are lowered one at a time, so the result depends on declaration order:
1. If the component type's mapper is declared later, the component is mapped without a mapper.
2. If the component type also has an allocatable component, lowering generates an implicit default mapper under the user mapper's symbol name. Lowering then skips the user's `DECLARE MAPPER`, so the user mapper is discarded for **every** use of that type.
3. Imported mappers are materialized in symbol-name order, so renaming a type can change which mappers apply.
None of these cases emits a diagnostic.
### Reproducer 1: component mapper not attached
```fortran
subroutine later_inner_mapper()
type :: inner
integer :: a(4)
end type
type :: outer
type(inner) :: in
end type
type(outer) :: x
!$omp declare mapper(outer :: w) map(tofrom: w%in)
!$omp declare mapper(inner :: v) map(to: v%a(1:2))
!$omp target map(tofrom: x)
x%in%a(1) = 1
!$omp end target
end subroutine
```
```
flang -fc1 -emit-hlfir -fopenmp -fopenmp-version=52 -fopenmp-targets=amdgcn-amd-amdhsa repro1.f90 -o -
```
The `outer` mapper maps `w%in` without a mapper:
```
omp.declare_mapper @_QQFlater_inner_mapperouter_omp_default_mapper : ... {
...
omp.map.info var_ptr(... !fir.type<_QFlater_inner_mapperTinner{...}>) map_clauses(tofrom) capture(ByRef) name("w%in")
```
If the two `DECLARE MAPPER` lines are swapped, the map has `mapper(@..._inner_omp_default_mapper) name("w%in")`. `-emit-llvm` succeeds in both cases, so the `inner` mapper is silently not applied to `x%in`.
### Reproducer 2: user mapper replaced by a compiler-generated mapper
```fortran
subroutine later_inner_mapper_alloc()
type :: inner
integer, allocatable :: p(:)
integer :: n
end type
type :: outer
type(inner) :: in
end type
type(inner) :: y
!$omp declare mapper(outer :: w) map(tofrom: w%in)
!$omp declare mapper(inner :: v) map(to: v%n)
!$omp target map(tofrom: y)
y%n = 1
!$omp end target
end subroutine
```
Here `@..._inner_omp_default_mapper` contains the compiler-generated body:
- It maps `p` with `map_clauses(implicit, tofrom, ref_ptee)`.
- The user's `map(to: v%n)` does not appear anywhere in the module.
- `y`, which is mapped directly and not through `outer`, also uses this replacement mapper.
With `-fno-implicit-default-mapper-allocatable-members`, the user mapper is kept, but `w%in` again gets no mapper, as in reproducer 1. An older trunk build (`e38e87dbb0f3`) keeps the user mapper here, so this case may be newer.
### Reproducer 3: imported mappers depend on symbol-name order
```fortran
! m.f90
module m
type :: b_inner
integer :: a(4)
end type
type :: a_outer
type(b_inner) :: in
end type
!$omp declare mapper(b_inner :: v) map(to: v%a(1:2))
!$omp declare mapper(a_outer :: w) map(tofrom: w%in)
end module
```
```fortran
! use.f90
subroutine use_m()
use m
type(a_outer) :: x
!$omp target map(tofrom: x)
x%in%a(1) = 1
!$omp end target
end subroutine
```
```
flang -fc1 -emit-hlfir -fopenmp -fopenmp-version=52 -module-dir . m.f90 -o /dev/null
flang -fc1 -emit-hlfir -fopenmp -fopenmp-version=52 -fopenmp-targets=amdgcn-amd-amdhsa -J . use.f90 -o -
```
When `m.f90` is compiled, its source order is correct, so `w%in` gets `mapper(@_QQMmb_inner_omp_default_mapper)`. In `use.f90`, the materialized `a_outer` mapper maps `w%in` without a mapper. After renaming the types so the inner type's name sorts first (`a_inner` and `b_outer`), `use.f90` attaches the mapper.
### Expected behavior
The `inner` default mapper is visible to both `DECLARE MAPPER` directives. Therefore, `w%in` should use it regardless of declaration order or module symbol names. A compiler-generated implicit mapper should never replace a user-declared default mapper.
Assisted with Copilot and Claude Opus 5.5.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs