R 3.6.0 for Debian buster

Johannes Ranke writes:
Johannes, 

It seems that one can avoid the gfortran problems with Fortran
BLAS/LAPACK implementations by compiling with
-fno-optimize-sibling-calls.  

Tomas recently wrote to R Core that

*********************************************************************
- gfortran is leaning to (but no official announcement decision has 
already been reached) making -fno-optimize-sibling-calls the default as 
not to break BLAS/LAPACK, at least temporarily to give BLAS/LAPACK 
authors some time to fix

- gfortran developers started discussing the issue with reference lapack 
maintainers on github, suggesting that the reference should use wrappers 
using portable C bindings using BIND(C) (which I've been suggesting to 
the gfortran developers as I saw it in "Numerical Computing with Modern 
Fortran" but of course they would have known too)
*********************************************************************

See 

https://github.com/Reference-LAPACK/lapack/issues/339
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90329

for more information.

Yesterday I changed R-devel and R-patched to use
-fno-optimize-sibling-calls for gfortran >= 7: it would be great if you
could pull this change into the R 3.6.0 backports for buster.

In principle I think all Fortran BLAS/LAPACK implementations (refblas
and ATLAS) packaged for buster should be recompiled with
-fno-optimize-sibling-calls (they may be fine in case they were compiled
with older version of gfortran-8, but then the next rebuild will cause
trouble): Dirk, any chance you could get the package maintainers to make
these changes?

Best
-k
Am Montag, 29. April 2019, 15:03:54 CEST schrieb Kurt Hornik:
Johannes Ranke writes:
Am Montag, 29. April 2019, 13:44:03 CEST schrieb Kurt Hornik:
Johannes Ranke writes:
Thanks.  You may have seen that with current gfortran in
testing/unstable, there are problems with the R BLAS/LAPACK API entries
using a Fortran interface (and hence in particular when using the BLAS
and LAPACK sources that ship with R).
No, I wasn't aware of this. Is there a bug report where this is
discussed?
Not really, as the issue seems to complicated to condense into a bug
report.  From discussions with Thomas Koenig from the GCC team, it seems
that f2c, g77 and now gfortran have always added additional character
length arguments for each character argument, where the R
F77_NAME/F77_CALL mechanism has always called with the arguments of the
Fortran subroutine but without the additional length arguments.  A
change in gcc trunk also ported to gcc-8-branch apparently changed what
happened in such case, to the effect that we're now seeing about 25
CRAN packages fail their run time checks with segfaults or run time
errors ...

But things are actually hard to pin down for us, and no obvious "fix"
is in sight.  It would be great if at least for the gfortran-8 that
Debian will release we would get the old behavior back ...
I think the likelihood of this would be greater if there was a bug against the 
version of gfortran in unstable/testing... Can you give a small reproducible 
example?
Johannes
Best
-k

Johannes

It seems I can avoid these using
OpenBLAS (but then this really only works find for me provided I setenv
OPENBLAS_NUM_THREADS=1).

-k

Dear all,
Now that the upcoming Debian release "buster" is frozen, I have started
supplying backports for it. Pending mirror synchronisations, R 3.6.0 is
now
available for Debian buster on i386 and amd64 architectures. Please
refer
to>

  https://cran.r-project.org/bin/linux/debian/

for details. At the moment I am not providing binaries for the arm
architecture for buster, as the SD card in my raspberry 3 has died and
I
do
not use these binaries any more anyways. Let me know if this is a
problem.

Kind regards,

Johannes

_______________________________________________
R-SIG-Debian mailing list
R-SIG-Debian at r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-debian

_______________________________________________
R-SIG-Debian mailing list
R-SIG-Debian at r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-debian

-- 
PD Dr. Johannes Ranke
Grenzach-Wyhlen

R 3.6.0 for Debian buster

Thread (10 messages)