make libs/readline know where headers/libs are

[dimpase]

To allow non-standard locations for readline (e.g. with MacOS Homebrew) one should use # distutils: include_dirs = etc in sage/libs/readline.pxd to make it look in non-standard locations.

readline/spkg-configure.m4 should set up appropriate variables to be filled in and exported in src/bin/sage-env-config.in; in turn they should make it into cython_aliases in sage/env.py.

This came up in this ​thread on sage-devel.

See also (readline-related):

• #29573 qepcad build fails on slackware-14.2-maximal
• #29572 Fix perl_term_readline_gnu build error on conda-forge-macos-maximal

See also (similar failure mode):

• #29607 comment 45
• #29448

[gh-mwageringel]

Also note that several packages seem to link against the macOS readline instead of the Homebrew one. In fact, the only one I am sure links against the Homebrew readline is rpy2. What would be needed to solve this? The Homebrew version provides a pkg-config file, in case that is useful.

I am hesitant to permanently install Homebrew's readline to /usr/local as this is against what is intended by Homebrew. In the case of readline though, I am not sure how much of a problem this really is.

[dimpase]

GAP needs a "real" readline, not a MacOS's libedit.

IIRC, sagelib's readline also needs "real" readline.

I did try installing Homebrew's readline into /usr/local, it seems to cause no harm, and allowed builds to work on MacOS 10.13.

[gh-mwageringel]

Replying to dimpase:

GAP needs a "real" readline, not a MacOS's libedit.

This might explain why I currently see several persistent doctest timeouts with this setup. Giac seems to be affected, as well. I am a bit surprised the build does not fail with this incompatible readline (other than for sage.libs.readline).

[mkoeppe]

Thanks for adding the link to the problem description.

What is happening here is that you are relying on an automake feature that our build system (unfortunately) does not have.

It is true that one can set variables by passing them to configure, and they will be respected during the configure run and remembered in config.status. However, in our build system they are not saved in any configure-generated output file:

$ ./configure LDFLAGS="-L/xyzzy1" CPPFLAGS="-I/xyzzy2" PKG_CONFIG_PATH="/xyzzy3"
$ grep xyzzy build/make/Makefile-auto build/make/Makefile src/Makefile src/bin/sage-env-config build/bin/sage-build-env-config build/pkgs/sage_conf/src/sage_conf.py build/pkgs/sage_conf/src/setup.cfg

... except in the automake-generated file build/make/Makefile-auto, which is not used at all in our build process.

build/make/Makefile-auto:CPPFLAGS =  -I/xyzzy2
build/make/Makefile-auto:LDFLAGS = -L/xyzzy1
build/make/Makefile-auto:PKG_CONFIG_PATH = /xyzzy3

[gh-mwageringel]

Thank you for explaining this. This clears up a lot. I think I will avoid the keg-only packages for now.

[mkoeppe]

Also PARI's spkg-install.in hardcodes readline to come from SAGE_LOCAL

bash ./Configure --prefix="$SAGE_LOCAL" \
    --with-readline="$SAGE_LOCAL" $SAGE_CONFIGURE_GMP \
    --with-runtime-perl="/usr/bin/env perl" \
    --kernel=gmp $PARI_CONFIGURE 

which is not correct when readline/spkg-configure.m4 found a system readline.

On local-homebrew-macos-standard (#29417), this leads to:

Hi-Res Graphics: none
###
### Editline wrapper detected, building without readline support
###
### Building without GNU readline support

[dimpase]

I believe these are missing for packages for which readline is not a dependency (but a feature). On the other hand, perhaps it can potentially pollute the compiling/linking calls.

Let's fix it...

[mkoeppe]

Another case:

• #29572 Fix perl_term_readline_gnu build error on conda-forge-macos-maximal

[dimpase]

It appears this will need AX_ABSOLUTE_HEADER, as we use for GMP/MPIR - I don't see another way around. :-(

[mkoeppe]

Then it has to be.

[dimpase]

underlinking of libreadline needs to be tested in readline's spkg-configure. And the name of the library has to be found from ac_cv_search_wresize set by AC_SEARCH_LIBS in ncurses' spkg-configure, or the result of pkg-config --libs ncurses (more precisely, the autoconf equivalent of the latter).

[isuruf]

Or we should make sure CFLAGS and LDFLAGS are respected by all the spkgs

[mkoeppe]

How's this coming along?

[dimpase]

hope to fix today (SG time)

[mkoeppe]

Anything I can help?

[dimpase]

feel free to take over - I won't have much time for it in the coming 48 hours. The branch I pushed is only a start.

---

New commits:

​af7cec6

compute NCURSESLIBNAMES, swicth to dependence macro

[mkoeppe]

Hm, not tonight

[dimpase]

I think slackware problem (#29573) is a different one - it ships an underlinked readline (and does not provide .pc file for readline, to make everything harder than needed).

[mkoeppe]

Moving to 9.4, as 9.3 has been released.

[dimpase]

should we proceed with this - in view of recent readline changes?

[mkoeppe]

The issue in the ticket description is indeed no longer relevant after #32073, so let's close this.

There are still issues with readline in other packages, as mentioned in ​https://trac.sagemath.org/ticket/29000#comment:7 above, but it does not help to keep the ticket open.

[dimpase]

OK