Add implicit interfaces for function hooks in Fortran module

[joewallwork]

Closes #2808.
Merges into #2807.

As mentioned in the issue, it turns out that the interfacing to the special Enzyme function hooks can be done using implicit interfaces in Fortran. That is, you can get away without providing an explicit interface block like is done in the examples so long as you declare

external :: __enzyme_autodiff

for example.

Checklist

• Add to Fortran enzyme module
• Use in square example

Hooking up in the allocatableArraySimple example will require fixing #2820.

[joewallwork]

Huh, looks like ifx doesn't like externals with leading underscores. I hadn't tested this with that compiler yet (just flang).

[joewallwork]

Huh, looks like ifx doesn't like externals with leading underscores. I hadn't tested this with that compiler yet (just flang).

Fixed in 237ebc6 by prepending with f for the Fortran versions of the function hooks.

[joewallwork]

Hm ifx still doesn't seem happy. I should have some time to debug later in the week.

[vchuravy]

Currently we test Fortran integration with LLVM 15, we eventually need to update that.

I just wanted to use an Enzyme build with LLVM 21 and ran into

The `opt -passname` syntax for the new pass manager is not supported, please use `opt -passes=<pipeline>` (or the `-p` alias for a more concise version).
See https://llvm.org/docs/NewPassManager.html#invoking-opt for more details on the pass pipeline syntax.

[vchuravy]

To debug the current failure:

/mnt/data/vchuravy/intel/2026.0/bin/ifx -flto -O0 -c -I/home/vchuravy/src/Enzyme/enzyme-core/build/modules /home/vchuravy/src/Enzyme/enzyme-core/enzyme/test/Fortran/ReverseMode/square.f90 -o /dev/stdout | llvm-dis

; Function Attrs: noinline nounwind optnone uwtable
define void @MAIN__() #0 {
alloca_2:
  %"$io_ctx" = alloca [8 x i64], align 8, !llfort.type_idx !5
  %"app_$DX" = alloca float, align 4, !llfort.type_idx !6
  %"app_$X" = alloca float, align 4, !llfort.type_idx !7
  %"var$1" = alloca i32, align 4, !llfort.type_idx !8
  %"(&)val$" = alloca [4 x i8], align 1, !llfort.type_idx !9
  %argblock = alloca <{ i64 }>, align 8, !llfort.type_idx !10
  %"var$2" = alloca i32, align 4, !llfort.type_idx !8
  %"(&)val$5" = alloca [4 x i8], align 1, !llfort.type_idx !11
  %argblock6 = alloca <{ i64 }>, align 8, !llfort.type_idx !10
  %func_result = call i32 @for_set_reentrancy(ptr @anon.f01f2640a9c3f724255ee7144b3336fc.0), !llfort.type_idx !8
  store float 3.000000e+00, ptr %"app_$X", align 4
  %func_result2 = call reassoc ninf nsz arcp contract afn float @math_mp_square_(ptr %"app_$X"), !llfort.type_idx !12
  store i64 0, ptr %"$io_ctx", align 8
  store [4 x i8] c"\1A\01\01\00", ptr %"(&)val$", align 1
  %BLKFIELD_float_ = getelementptr inbounds <{ i64 }>, ptr %argblock, i32 0, i32 0, !llfort.type_idx !13
  store float %func_result2, ptr %BLKFIELD_float_, align 4
  %func_result4 = call i32 (ptr, i32, i64, ptr, ptr, ...) @for_write_seq_lis(ptr %"$io_ctx", i32 -1, i64 2253038970797824, ptr %"(&)val$", ptr %argblock), !llfort.type_idx !8
  store float 0.000000e+00, ptr %"app_$DX", align 4
  call void @f__enzyme_autodiff_.t69p.t70p.t71p(ptr @math_mp_square_, ptr %"app_$X", ptr %"app_$DX"), !llfort.type_idx !14
  %"app_$DX_fetch.3" = load float, ptr %"app_$DX", align 4, !llfort.type_idx !6
  store i64 0, ptr %"$io_ctx", align 8
  store [4 x i8] c"\1A\01\01\00", ptr %"(&)val$5", align 1
  %BLKFIELD_float_7 = getelementptr inbounds <{ i64 }>, ptr %argblock6, i32 0, i32 0, !llfort.type_idx !15
  store float %"app_$DX_fetch.3", ptr %BLKFIELD_float_7, align 4
  %func_result9 = call i32 (ptr, i32, i64, ptr, ptr, ...) @for_write_seq_lis(ptr %"$io_ctx", i32 -1, i64 2253038970797824, ptr %"(&)val$5", ptr %argblock6), !llfort.type_idx !8
  ret void
}

; Function Attrs: noinline nounwind optnone uwtable
define internal void @f__enzyme_autodiff_.t69p.t70p.t71p(ptr %arg0, ptr %arg1, ptr %arg2) #0 {
wrap_start20:
  call void (...) @f__enzyme_autodiff_(ptr %arg0, ptr %arg1, ptr %arg2)
  ret void
}

declare void @f__enzyme_autodiff_(...)

So the additional indirection is hiding things.

[vchuravy]

So the issue only exists at -O0

at -O1

; Function Attrs: nounwind uwtable
define void @MAIN__() local_unnamed_addr #2 {
alloca_2:
  %"$io_ctx" = alloca [8 x i64], align 8, !llfort.type_idx !7
  %"app_$DX" = alloca float, align 4, !llfort.type_idx !8
  %"app_$X" = alloca float, align 4, !llfort.type_idx !9
  %"(&)val$" = alloca [4 x i8], align 1, !llfort.type_idx !10
  %argblock = alloca <{ i64 }>, align 8, !llfort.type_idx !11
  %"(&)val$7" = alloca [4 x i8], align 1, !llfort.type_idx !12
  %argblock8 = alloca <{ i64 }>, align 8, !llfort.type_idx !11
  %func_result = tail call i32 @for_set_fpe_(ptr nonnull @anon.f01f2640a9c3f724255ee7144b3336fc.0) #4, !llfort.type_idx !13
  %func_result2 = tail call i32 @for_set_reentrancy(ptr nonnull @anon.f01f2640a9c3f724255ee7144b3336fc.1) #4, !llfort.type_idx !13
  store float 3.000000e+00, ptr %"app_$X", align 4, !tbaa !14
  store i64 0, ptr %"$io_ctx", align 8, !tbaa !19
  %.fca.0.gep12 = getelementptr inbounds nuw [4 x i8], ptr %"(&)val$", i64 0, i64 0
  store i8 26, ptr %.fca.0.gep12, align 1, !tbaa !19
  %.fca.1.gep13 = getelementptr inbounds nuw [4 x i8], ptr %"(&)val$", i64 0, i64 1
  store i8 1, ptr %.fca.1.gep13, align 1, !tbaa !19
  %.fca.2.gep14 = getelementptr inbounds nuw [4 x i8], ptr %"(&)val$", i64 0, i64 2
  store i8 1, ptr %.fca.2.gep14, align 1, !tbaa !19
  %.fca.3.gep15 = getelementptr inbounds nuw [4 x i8], ptr %"(&)val$", i64 0, i64 3
  store i8 0, ptr %.fca.3.gep15, align 1, !tbaa !19
  %BLKFIELD_float_ = getelementptr inbounds nuw <{ i64 }>, ptr %argblock, i64 0, i32 0, !llfort.type_idx !20
  store float 9.000000e+00, ptr %BLKFIELD_float_, align 8, !tbaa !21
  %func_result6 = call i32 (ptr, i32, i64, ptr, ptr, ...) @for_write_seq_lis(ptr nonnull %"$io_ctx", i32 -1, i64 2253038970797824, ptr nonnull %"(&)val$", ptr nonnull %argblock) #4, !llfort.type_idx !13
  store float 0.000000e+00, ptr %"app_$DX", align 4, !tbaa !23
  call void (...) @f__enzyme_autodiff_(ptr nonnull @math_mp_square_, ptr nonnull %"app_$X", ptr nonnull %"app_$DX") #4
  %"app_$DX_fetch.3" = load float, ptr %"app_$DX", align 4, !tbaa !23, !llfort.type_idx !8
  store i64 0, ptr %"$io_ctx", align 8, !tbaa !19
  %.fca.0.gep = getelementptr inbounds nuw [4 x i8], ptr %"(&)val$7", i64 0, i64 0
  store i8 26, ptr %.fca.0.gep, align 1, !tbaa !19
  %.fca.1.gep = getelementptr inbounds nuw [4 x i8], ptr %"(&)val$7", i64 0, i64 1
  store i8 1, ptr %.fca.1.gep, align 1, !tbaa !19
  %.fca.2.gep = getelementptr inbounds nuw [4 x i8], ptr %"(&)val$7", i64 0, i64 2
  store i8 1, ptr %.fca.2.gep, align 1, !tbaa !19
  %.fca.3.gep = getelementptr inbounds nuw [4 x i8], ptr %"(&)val$7", i64 0, i64 3
  store i8 0, ptr %.fca.3.gep, align 1, !tbaa !19
  %BLKFIELD_float_9 = getelementptr inbounds nuw <{ i64 }>, ptr %argblock8, i64 0, i32 0, !llfort.type_idx !25
  store float %"app_$DX_fetch.3", ptr %BLKFIELD_float_9, align 8, !tbaa !26
  %func_result11 = call i32 (ptr, i32, i64, ptr, ptr, ...) @for_write_seq_lis(ptr nonnull %"$io_ctx", i32 -1, i64 2253038970797824, ptr nonnull %"(&)val$7", ptr nonnull %argblock8) #4, !llfort.type_idx !13
  ret void
}

[vchuravy]

I was playing around with:

!DIR$ ATTRIBUTES FORCEINLINE :: enzyme_autodiff;
!DIR$ ATTRIBUTES FORCEINLINE :: enzyme_fwddiff;

and variations thereof, but nothing seems to stick.

Adding

    !DIR$ FORCEINLINE
    call enzyme_autodiff(square, x, dx);

does work but is kinda silly.

The options I see is:

1. Don't test -O0 -- meh
2. Add the callsite directive -- meh, and compiler dependent
3. Don't do the clever indirection -- sad

[joewallwork]

I was playing around with:

!DIR$ ATTRIBUTES FORCEINLINE :: enzyme_autodiff;
!DIR$ ATTRIBUTES FORCEINLINE :: enzyme_fwddiff;

and variations thereof, but nothing seems to stick.

Adding

    !DIR$ FORCEINLINE
    call enzyme_autodiff(square, x, dx);

does work but is kinda silly.

The options I see is:

1. Don't test `-O0` -- meh

2. Add the callsite directive -- meh, and compiler dependent

3. Don't do the clever indirection -- sad

Alas. My debugging today has also not really revealed any ways forward with ifx.

I'll work on setting up a flang CI so that we can test the bindings under that compiler (works fine for me so far) and I guess we can return to ifx later.

[vchuravy]

I think the conclusion from Monday was to go with f__enzyme_autodiff or add something like ENZYME_autodiff.

[joewallwork]

I think the conclusion from Monday was to go with f__enzyme_autodiff or add something like ENZYME_autodiff.

Reverted in 57855fa.

[vchuravy]

+ ifx -flto -O0 -c -I/home/runner/work/Enzyme/Enzyme/build/modules /home/runner/work/Enzyme/Enzyme/enzyme/test/Fortran/ReverseMode/square.f90 -o /dev/stdout
+ /usr/lib/llvm-15/bin/opt --enable-new-pm=0 -load=/home/runner/work/Enzyme/Enzyme/build/Enzyme/LLVMEnzyme-15.so --enzyme-attributor=0 -enzyme -o /home/runner/work/Enzyme/Enzyme/build/test/Fortran/ReverseMode/Output/square.f90.tmp
error: <unknown>:0:0: in function f__enzyme_autodiff_.t72p.t73p.t74p void (float (float*)*, float*, float*): Enzyme: failed to find fn to differentiate  call void (...) @f__enzyme_autodiff_(float (float*)* %arg0, float* %arg1, float* %arg2) - found - float (float*)* %arg0

So ifx still doesn't like this at O0

[joewallwork]

So ifx still doesn't like this at O0

Alas, indeed. I just pushed c4fac16 to the other PR #2837 to check that ifx works for all the tests there so long as -O0 is not used.

[joewallwork]

Alas, indeed. I just pushed c4fac16 to the other PR #2837 to check that ifx works for all the tests there so long as -O0 is not used.

Okay that passes. @vchuravy How about we revert back to the neater enzyme_autodiff pattern and just don't test implicit interfaces under ifx with -O0 for now? I can add a documentation page that says you need to use an explicit interface at -O0 with ifx. Would that be an acceptable resolution?

[joewallwork]

Alas, indeed. I just pushed c4fac16 to the other PR #2837 to check that ifx works for all the tests there so long as -O0 is not used.

Okay that passes. @vchuravy How about we revert back to the neater enzyme_autodiff pattern and just don't test implicit interfaces under ifx with -O0 for now? I can add a documentation page that says you need to use an explicit interface at -O0 with ifx. Would that be an acceptable resolution?

Pushed this proposal in recent commits. Eventually we want the docs to live on the webpage but a README should be sufficient at this stage.