verify if cblas.pc and lapack.pc should be replaced by links to openblas.pc

[dimpase]

currently openblas's spkg-configure.pc unconditionally makes cblas.pc and lapack.pc copies of openblas.pc, which in some cases if incorrect, e.g. Arch Linux has libcblas linked to libopenblas
and containing stuff missing in libopenblas, resulting in errors.

this has been reported on sage-devel
https://groups.google.com/d/msg/sage-devel/pIOnFyFJMtM/_FbzM2OxCQAJ

So we should make these installations conditional.

Depends on #29051

CC: @isuruf @mkoeppe @antonio-rojas

Component: build: configure

Author: Dima Pasechnik

Branch/Commit: 3a4524e

Reviewer: Isuru Fernando, Matthias Koeppe

Issue created by migration from https://trac.sagemath.org/ticket/29071

[dimpase]

comment:1

It depends on the fortran name mangling. For gfortran, one symbol would be dgeqrf_.

we can do an AC_LINK_IFELSE() test with a subroutine dgeqef call,
and try to link it against libopenblas. this would be independent of Fortran compiler.

http://www.netlib.org/lapack/explore-html/df/dc5/group__variants_g_ecomputational_ga3766ea903391b5cf9008132f7440ec7b.html

[dimpase]

comment:2

And for cblas we can do AC_CHECK_LIB() on cblas_dgemm, as Antonio proposed.

[isuruf]

comment:3

You can also use AC_FC_FUNC to get the mangled name.

[dimpase]

comment:4

Replying to @isuruf:

You can also use AC_FC_FUNC to get the mangled name.

OK, that's better, thanks.

[dimpase]

Author: Dima Pasechnik

[dimpase]

Branch: u/dimpase/packages/openblaspcfix

[dimpase]

comment:5

needs testing on Arch in particular.

---

New commits:

ccc5d82

install cblas.pc and lapack.pc links conditionally

[dimpase]

Commit: ccc5d82

[videlec]

comment:6

trying now on arch

[sagetrac-git]

Branch pushed to git repo; I updated commit sha1. New commits:

f6714fb

use AC_SEARCH_LIBS

[sagetrac-git]

Changed commit from ccc5d82 to f6714fb

[videlec]

comment:8

With the branch applied, I end up with the same permission issues as with ./configure --without-system-openblas. That is to say

Copying package files from temporary location /opt/sage/local/var/tmp/sage/build/openblas-0.3.6.p0/inst to /opt/sage/local
cp: cannot create regular file '/opt/sage/local/./lib/pkgconfig/blas.pc': Permission denied
cp: cannot create regular file '/opt/sage/local/./lib/pkgconfig/cblas.pc': Permission denied
cp: cannot create regular file '/opt/sage/local/./lib/pkgconfig/lapack.pc': Permission denied

[dimpase]

comment:9

Is /opt/sage/local your $SAGE_LOCAL?
If so, could you try rm -rf /opt/sage/local/pkgconfig/ ?

[dimpase]

comment:10

sorry, I meant rm -rf /opt/sage/local/lib/pkgconfig/

[dimpase]

comment:11

even more precisely, rm -f /opt/sage/local/lib/pkgconfig/*blas.pc && rm -f /opt/sage/local/lib/pkgconfig/lapack.pc

[dimpase]

comment:12

I presume you get permission problems while running make. This is due to #29003 - where the installation of these is done somewhat carelessly.

[sagetrac-git]

Changed commit from f6714fb to 7ac1293

[sagetrac-git]

Branch pushed to git repo; I updated commit sha1. New commits:

7ac1293

added missing module names - cblas and lapack

[videlec]

comment:14

When I do make distclean it does remove local entirely. At which step do you want me to try rm XXX?

[dimpase]

comment:15

the logic of this branch is still not 100% correct - if we end up installing OpenBLAS we should not install any *.pc files, i.e. should not run any AC_CONFIG_LINKS.

[dimpase]

comment:16

Replying to @videlec:

When I do make distclean it does remove local entirely. At which step do you want me to try rm XXX?

these rm from comment:11 should be run after ./configure and before make, assuming I understand what's going on.

[dimpase]

comment:17

By the way, please pull the branch again - comment:13 is rather important for Arch.

[videlec]

comment:18

After ./configure there is no directory local/lib/pkgconfig/

$ ls -R local/
local/:
bin  etc  include  lib  lib64  share  var

local/bin:

local/etc:

local/include:

local/lib:

local/share:

local/var:
lib

local/var/lib:
sage

local/var/lib/sage:
installed

local/var/lib/sage/installed:

[dimpase]

comment:19

local/lib/pkgconfig/ may or may not be present, it depends upon the state of the system.
(e.g. if you did make distclean before ./configure you won't have
local/ at all)

Now, what happens if you run make? Do you still get the same error?

[dimpase]

comment:20

It seems I got Arch running in an lxc container, so I guess I can polish it on my own.

[dimpase]

comment:72

I also don't get

+      # We already checked (in the initial portion of configure.ac)
+      # whether the Fortran compiler accepts free-format source code (as

IMHO there is nothing either in configure.ac or build/pkgs/gcc/spkg-configure.m4 that
does any Fortran checks.

[dimpase]

comment:73

Replying to @dimpase:

+       if test "$SAGE_HAVE_FC_FREEFORM" = yes ; then dnl Have a suitable Fortran compiler

could we please stick to AS_IF ?

actually, I don't see the point of this if at all. If we get there then it's already true, as it's inside the check for gfortran as a pre-req, in the scope of SAGE_SPKG_DEPCHECK([gfortran], ...).
`

[dimpase]

comment:74

The trouble is in AC_FC_FUNC(), which throws a fit if there is no Fortran compiler around, instead of just suffering in silence.

[sagetrac-git]

Branch pushed to git repo; I updated commit sha1. This was a forced push. New commits:

8cf2a3b	configure.ac: Check some FC properties earlier
c98c6b2	configure.ac, build/pkgs/{openblas,gfortran}: Do not fail if there is no FC

[sagetrac-git]

Changed commit from d2b5e13 to c98c6b2

[mkoeppe]

comment:76

Replying to @dimpase:

I also don't get

+      # We already checked (in the initial portion of configure.ac)
+      # whether the Fortran compiler accepts free-format source code (as

IMHO there is nothing either in configure.ac or build/pkgs/gcc/spkg-configure.m4 that
does any Fortran checks.

Sorry, I forgot to include one commit on this branch.
Fixed.

[mkoeppe]

comment:77

Replying to @dimpase:

The trouble is in AC_FC_FUNC(), which throws a fit if there is no Fortran compiler around, instead of just suffering in silence.

Yes, exactly, and more specifically: Because of AC_REQUIRE and AC_DEFUN_ONCE one cannot easily make all things conditional that one would liike.

[mkoeppe]

comment:78

Replying to @dimpase:

Replying to @dimpase:

+       if test "$SAGE_HAVE_FC_FREEFORM" = yes ; then dnl Have a suitable Fortran compiler

could we please stick to AS_IF ?

No, they are unfortunately not equivalent because of autoconf's AC_PREREQ/AC_DEFUN_ONCE magic.

[dimpase]

New commits:

39a317d	install cblas.pc and lapack.pc links conditionally
7adc712	use AC_SEARCH_LIBS
abda83e	added missing module names - cblas and lapack
683a926	make installing *.pc files conditional
a0acbcb	correct quoting of m4 index variable
5ee6844	ensure gfortran is available
56541ae	wrap AC_FC_FUNC so that it does not throw an error without Fortran
3a4524e	Merge remote-tracking branch 'trac/public/packages/numpy/no_DEFAULT_and_ALL_numpy_site_cfg' into openblaspcfix

[dimpase]

Changed branch from u/mkoeppe/packages/openblaspcfix to u/dimpase/packages/openblaspcfix

[dimpase]

Changed commit from c98c6b2 to 3a4524e

[dimpase]

comment:80

Replying to @mkoeppe:

I don't like the idea of requiring a Fortran compiler to even run ./configure successfully.

Fixed by wrap AC_FC_FUNC so that it does not throw an error without Fortran: 56541ae

[dimpase]

comment:81

IMHO it's an autoconf bug/feature that AC_FC_FUNC errors out without Fortran, while e.g. AC_FC_FREEFORM does not.

[mkoeppe]

comment:82

Replying to @dimpase:

IMHO it's an autoconf bug/feature that AC_FC_FUNC errors out without Fortran, while e.g. AC_FC_FREEFORM does not.

Yes, I agree.

[mkoeppe]

comment:83

Replying to @dimpase:

Replying to @mkoeppe:

I don't like the idea of requiring a Fortran compiler to even run ./configure successfully.

Fixed by wrap AC_FC_FUNC so that it does not throw an error without Fortran: 56541ae

Great, a much easier fix than what I put together in ​8cf2a3b/​c98c6b2.
Testing at https://github.com/mkoeppe/sage/actions/runs/32186055

[mkoeppe]

comment:84

Tests finished at https://github.com/mkoeppe/sage/actions/runs/32186055

[mkoeppe]

comment:85

Looking good, and certainly fixes what was original reported on this ticket (fflas_ffpack on archlinux). The build error of fflas_ffpack on debian-jessie-standard (https://github.com/mkoeppe/sage/runs/412891766) is unrelated. It is good on all other configurations.

[vbraun]

Changed branch from u/dimpase/packages/openblaspcfix to 3a4524e