Re: libgomp.fortran/async_io_[1,2,3,4,8,9].f90 fail on FreeBSD

[Tobias Burnus <burnus@net-b.de>]

Hi Steve,

On 23.10.21 18:31, Steve Kargl via Fortran wrote:
> Do you know how to run a single libgomp.fortran test? I tried
> % gmake check-fortran RUNTESTSFLAGS="gomp.exp=async_io.f90"
>
> but this runs all the testcases.

First, it should be RUNTESTFLAGS=  (test not tests).

Additionally, gomp.exp (for Fortran) is in 
gcc/testsuite/gfortran.dg/gomp/ but if you want to test 
libgomp/testsuite/libgomp.fortran/ (or libgomp.oacc-fortran/), you have 
to use "fortran.exp" (and not "gomp.exp").

And in order to make RUNTESTFLAGS effective, you cannot run in 
$(BUILD).* You either have to move to $(BUILD)/gcc – and run there "make 
check-fortran". For for libgomp, you need to go to 
$(BUILD)/$target_triplet/libgomp/ and run "make check" there. 
($target_triplet is "x86_64-pc-linux-gnu" on my laptop.)

(* at least it does not seem to work here.)
> commit date of 2020-05-23 for async_io_9.f90, I know these
> tests were passing in the past.  I fear a reason Sandra/Tobias
> commit has broken things.

I sincerely doubt that our Fortran commits have broken this – they are 
not related to anything like TS29113, BIND(C), polymorphism or 
assumed-rank. I am even skeptical that any OpenMP change could have 
caused that, given that those do not really use OpenMP at all – but only 
pthreads (which is the reason they are in libgomp as OpenMP also depends 
on threads). Granted, as they are compiled with -fopenmp, libgomp is 
linked in (but should never be called, except for library initialization).

Nonetheless, if it used to work and now fails, something must have 
changed, causing the breakage.

If you compile and run the program manually without the DejaGNU, does it 
fail when compiled as is, compiled with -pthreads, compiled with -fopenmp?

Tobias