[Fortran] Handle variable names starting with underscore

[Thirumalai-Shaktivel]

No description provided.

[certik]

Perfect. Is there a test that this now enables to compile that fails in master?

We need to figure out how to test this. One option is to enable this pass in the "fortran" backend by default, which is the one used by the cmake test suite. And then add a test that would otherwise fail, but now works.

[Thirumalai-Shaktivel]

Is there a test that this now enables to compile that fails in master?

Yup, for that we have to apply passes. I came across this issue while testing the array_op pass.

All the variables staring with _ are introduced after applying some passes like array_op, intrinsic_function, ...

We need to apply passes for the Fortran backend, right? Currently we don't.
If we have to apply passes, default passes for LLVM would be correct one, right? Or should I skip some passes there?

[certik]

We should apply some passes. I think we should apply this pass in the Fortran backend (in --backend=fortran, not in --show-fortran).

Do you want to try enabling it and see if some new test passes? That would be a good test for this PR.

[Thirumalai-Shaktivel]

Consider the following example:

program expr2
implicit none
integer :: x(2)
x = (2+3)*5
end program

I get, --show-fortran

program expr2
implicit none
integer(4), dimension(2) :: x
x = (2 + 3)*5
end program expr2

--backend=fortran --apply-fortran-mangling

program expr2
implicit none
integer(4) :: v__1_t
integer(4) :: v__libasr_created_scalar_auxiliary_variable
integer(4), dimension(2) :: x
v__libasr_created_scalar_auxiliary_variable = (2 + 3)*5
v__1_t = lbound(x, 1) - 1
do while (v__1_t + 1 <= ubound(x, 1))
    v__1_t = v__1_t + 1
    x(v__1_t) = v__libasr_created_scalar_auxiliary_variable
end do
end program expr2

The following also works
--show-fortran --pass=array_op

program expr2
implicit none
integer(4) :: __1_t
integer(4) :: __libasr_created_scalar_auxiliary_variable
integer(4), dimension(2) :: x
__libasr_created_scalar_auxiliary_variable = (2 + 3)*5
do __1_t = lbound(x, 1), ubound(x, 1)
    x(__1_t) = __libasr_created_scalar_auxiliary_variable
end do
end program expr2

[Thirumalai-Shaktivel]

Enabling the passes requires most of the changes from this PR: #2747, so let's merge that and then work on this.

[certik]

Ok, go ahead and rebase this.

Important: the --backend=fortran should not be applying the array_op pass. The only passes it should be applying is to rewrite LFortran extensions to standard Fortran, such as templates or unsigned integers.

I think the application of this PR is when we are printing the ASR using --dump-all-passes-fortran, and we want to print it in such a way to be able to compile. I think it should not be on by default (so that we can see exactly what the ASR has), but as an option (if we want to compile with gfortran, let's say).

[Thirumalai-Shaktivel]

I see. In that case, I think we won't be able to test this PR changes using integration_tests CMakeLists. As I said earlier, the _ symbols are introduced only after applying the passes.

Should I test it using tests.toml?

[Thirumalai-Shaktivel]

Ready for review.

[certik]

Does this mean the unique_symbols pass is always applied even without the --apply-fortran-mangling option? Is this only for the --backend-fortran, or also --dump-all-passes-fortran?

[Thirumalai-Shaktivel]

The unique_symbols passes are applied only when we pass some options like: https://github.com/lfortran/lfortran/pull/2806/files#diff-0b1b8d6c43c5a0e83ce23cda4c34fde853efebc4baa0cb1676e9f4298f82c532R505-R516

It is used by both --show-fortran and --backend=fortran.
--dump-all-passes-fortran dumps all the passes output.

[certik]

Ok, I think this is fine then.

[certik]

I think this is fine.

[Thirumalai-Shaktivel]

Thanks for the review!