Test suite fails when using GCC 10

[DimStar77]

Describe the bug
GCC10 is nearing release and various distros are sarting preparation work. Meson's test suite has been identified as a component that is failing when using gcc19

[  873s] Running compile:
[  873s] Working directory:  /home/abuild/rpmbuild/BUILD/meson-0.54.0/b 536484cc32/meson-private/tmpsv_1caih
[  873s] Command line:  gfortran /home/abuild/rpmbuild/BUILD/meson-0.54.0/b 536484cc32/meson-private/tmpsv_1caih/testfile.f90 -pipe -cpp -E -P -D_FILE_OFFSET_BITS=64 -cpp -P -O0 
[  873s] 
[  873s] Code:
[  873s]  
[  873s]         #ifdef __has_include
[  873s]          #if !__has_include("omp.h")
[  873s]           #error "Header 'omp.h' could not be found"
[  873s]          #endif
[  873s]         #else
[  873s]          #include <omp.h>
[  873s]         #endif
[  873s] Compiler stdout:
[  873s]  
[  873s] 
[  873s] 
[  873s]         #ifdef 0
[  873s]          #if !0("omp.h")
[  873s]           #error "Header 'omp.h' could not be found"
[  873s]          #endif
[  873s]         #else
[  873s]          #include <omp.h>
[  873s]         #endif
[  873s] 
[  873s] Compiler stderr:
[  873s]  /home/abuild/rpmbuild/BUILD/meson-0.54.0/b 536484cc32/meson-private/tmpsv_1caih/testfile.f90:2:0:
[  873s] 
[  873s]     2 |         #ifdef __has_include
[  873s]       | 
[  873s] Error: "__has_include" used outside of preprocessing directive
[  873s] /home/abuild/rpmbuild/BUILD/meson-0.54.0/b 536484cc32/meson-private/tmpsv_1caih/testfile.f90:2:0: Error: missing '(' before "__has_include" operand
[  873s] /home/abuild/rpmbuild/BUILD/meson-0.54.0/b 536484cc32/meson-private/tmpsv_1caih/testfile.f90:2:0: Error: operator "__has_include" requires a header-name
[  873s] /home/abuild/rpmbuild/BUILD/meson-0.54.0/b 536484cc32/meson-private/tmpsv_1caih/testfile.f90:3:0:
[  873s] 
[  873s]     3 |          #if !__has_include("omp.h")
[  873s]       | 
[  873s] Error: "__has_include" used outside of preprocessing directive
[  873s] /home/abuild/rpmbuild/BUILD/meson-0.54.0/b 536484cc32/meson-private/tmpsv_1caih/testfile.f90:3:0: Error: missing '(' before "__has_include" operand
[  873s] /home/abuild/rpmbuild/BUILD/meson-0.54.0/b 536484cc32/meson-private/tmpsv_1caih/testfile.f90:3:0: Error: operator "__has_include" requires a header-name
[  873s] 

Expected behavior
A passing test suite - irrespective of the use of gcc100

system parameters

• Tested on openSUSE Tumbleweed Staging (i.e. 'packages to integrate into future openSUSE Tumbleweed snapshots); the only change is the switch from gcc9.3.1 to gcc 10

Always latest build log available at: https://build.opensuse.org/build/openSUSE:Factory:Staging:N/standard/x86_64/meson:test/_log

[lazka]

The MSYS2 jobs in CI show this error now since the update to GCC 10.

[lazka]

since it affects fortran @scivision do you have any idea what's wrong here?

[scivision]

The code put in testfile.f90 isn't a full test if OpenMP is truly linkable. GCC 10 and in particular Gfortran 10 got a lot stricter about invalid code, and this appears ripe for updating anyway.

Old Meson code

doesn't work with gfortran -E -P -cpp for GCC 10

#ifdef __has_include
#if !__has_include("omp.h")
#error "Header \'omp.h\' could not be found"
#endif
#else 
#include <omp.h>
#endif

Fortran-specific stanza for detecting if OpenMP is available and linkable

In general I strongly prefer tests like this, because in general users may have multiple compilers and OpenMP libraries installed, which are not necessarily ABI compatible. This brief test gives higher confidence that an ABI-compatible set is found.

use omp_lib  ! tests that OpenMP is included
implicit none
!$ integer :: N  ! OpenMP reads lines starting like this
N = omp_get_num_threads() ! tests that OpenMP is linked

this shorter test is not as complete, but is still more thorough than the original Meson code:

use omp_lib

[scivision]

If this is suitable, one would slightly modify the Meson code to swap in this stanza in if the language is Fortran. I can do this if others agree.

[lazka]

From what I see this is just the output of the generic has_header, maybe called from

meson/mesonbuild/dependencies/misc.py

Line 98 in 402f7af

if self.clib_compiler.has_header(name, '', self.env, dependencies=[self], disable_cache=True)[0]:

so not easy to replace.

[nirbheek]

The main bug to solve is that compiler.has_header() does not work if compiler is gfortran >= 10.

[nirbheek]

#7136 fixes it.