[Flang] Preprocessor duplicates backslashes

[cjosey]

I was testing Flang 20 (commit: 7cbb365) in a CMake project. When using the Ninja build system, it will always run the preprocessor over the files and then compile the preprocessed result.

The preprocessor will escape the backslashes but the compiler will not re-combine them, resulting in a duplication of backslashes.

MWE (named test.f90):

program main
    implicit none

    write(*,*) "\"
end program main

Run:

flang -cpp -E test.f90 > test-pp.f90
flang test-pp.f90
./a.out

Expected result:

 \

Result shown:

 \\

If just the compile stages are provided the flag -fbackslash the result becomes correct. If both are provided -fbackslash it fails as the original string is unescaped.

[llvmbot]

@llvm/issue-subscribers-flang-frontend

Author: Colin Josey (cjosey)

I was testing Flang 20 (commit: 7cbb365) in a CMake project. When using the Ninja build system, it will always run the preprocessor over the files and then compile the preprocessed result.

The preprocessor will escape the backslashes but the compiler will not re-combine them, resulting in a duplication of backslashes.

MWE (named test.f90):

program main
    implicit none

    write(*,*) "\"
end program main

Run:

flang -cpp -E test.f90 > test-pp.f90
flang test-pp.f90
./a.out

Expected result:

 \

Result shown:

 \\

If just the compile stages are provided the flag -fbackslash the result becomes correct. If both are provided -fbackslash it fails as the original string is unescaped.

[cjosey]

Possibly related, the MWE:

program main
    implicit none

1   format('\')
    write(*, 1)
end program main

Also prints:

\\

without preprocessing. It expands even further:

\\\\

with preprocessing.

[olee7]

I created a PR to attempt at this issue:
#115822

Excuse me @klausler, can I please get a review for the above PR?

[olee7]

Never mind, one of the tests is failing.

[olee7]

Excuse me @cjosey , by chance, are you looking for this option: -fbackslash

According to the below doc:
https://flang.llvm.org/docs/FlangCommandLineReference.html

-fbackslash is required to specify the backslash in string introduces an escape character,

flang -cpp -E -fbackslash test.f90 > test-pp.f90
flang test-pp.f90
./a.out

produced the result you are looking for?

[cjosey]

The correct output is achieved only if the second command is given the flag and not the first:

flang -cpp -E test.f90 > test-pp.f90
flang -fbackslash test-pp.f90

If it is put on the first call, it will fail to compile as \ is un-escaped.

If this is expected or preferred behavior I could let the CMake folks know that the build stage must have -fbackslash regardless of the flags on the preprocessing phase. I only thought to mention it here first as it diverges from gfortran and ifort behavior.

The second example with the \ in the format statement, however, cannot be fixed with -fbackslash. I cannot find a way to make it emit one \ to stdout that way.