on Arch make+guile is broken

[videlec]

Guile plugin in make under certain not completely clear conditions, likely to do with a version misconfiguration of system libraries,
may fail to build a number of Sage packages. For example, while building Sage 8.2.beta3 on latest archlinux one gets

make: symbol lookup error: /usr/lib/libguile-2.2.so.1: undefined symbol: 
GC_move_disappearing_link

as libguile looks at the LD_LIBRARY_PATH containing a different version of libgc.so than the one it needs.

See also this report on sage-devel.

After deactivating the gc package, the compilation went fine.

However, this cannot be reproduced on other Linux systems.

The workaround in the branch consists in declaring the environment variable LD_PRELOAD so that make uses the system gc. The workaround has to be applied to the 2 standard packages R and rpy2.

The three other packges flint/arb/deformation fail to build for the same reason and we apply a small patch to avoid them redefining LD_LIBRARY_PATH.

Upstream issues

• flint: upstream issue at DLPATH and make flintlib/flint#447
• arb: upstream issue at DLPATH and make flintlib/arb#213
• deformation: issue with recent gcc "/usr/bin/ld: -r and -pie may not be used together" jpflori/pydeformation#5 (for the pie trouble from optional package deformation fails to build #24902)

Depends on #24885

Upstream: Reported upstream. No feedback yet.

CC: @embray @vbraun @EmmanuelCharpentier @defeo @jpflori

Component: packages: standard

Author: Erik Bray, Vincent Delecroix

Branch/Commit: u/vdelecroix/24575 @ 0d8a4a9

Reviewer: Erik Bray, Vincent Delecroix

Issue created by migration from https://trac.sagemath.org/ticket/24575

[videlec]

Description changed:

--- 
+++ 
@@ -6,5 +6,6 @@
 make: symbol lookup error: /usr/lib/libguile-2.2.so.1: undefined symbol: 
 GC_move_disappearing_link
 ```
+After deactivating the `gc` packages, the compilation went fine on this machine.
 
 See also [this report on sage-devel](https://groups.google.com/forum/#!topic/sage-devel/5LqXytiF2nQ).

[videlec]

Description changed:

--- 
+++ 
@@ -6,6 +6,6 @@
 make: symbol lookup error: /usr/lib/libguile-2.2.so.1: undefined symbol: 
 GC_move_disappearing_link
 ```
-After deactivating the `gc` packages, the compilation went fine on this machine.
+After deactivating the `gc` package, the compilation went fine on this machine.
 
 See also [this report on sage-devel](https://groups.google.com/forum/#!topic/sage-devel/5LqXytiF2nQ).

[EmmanuelCharpentier]

comment:3

A (possibly naive) suggestion is to do (almost) what we already do for gcc and related, but simpler :

• Depend on gc
• Test for it (and its version) in the main configure file
  • If found (and sufficient) : use that (possibly symlinking the relevant header/library files)
  • Else : install "our" version.

Drawback : a newer version might break compatibility. A code review of its use is necessary.

BTW : can one use pkg-config in the main configure file ? I think not (alas...).

[dimpase]

comment:4

At what point do you get this error? While building R? (guessing the latter from the subject of sage-devel post)

[dimpase]

comment:5

how come make depends on guile for you? I don't see it.

$ ldd `which gmake`
	linux-vdso.so.1 (0x00007ffe861ce000)
	libdl.so.2 => /lib64/libdl.so.2 (0x00007f0e5b6d3000)
	libc.so.6 => /lib64/libc.so.6 (0x00007f0e5b30f000)
	/lib64/ld-linux-x86-64.so.2 (0x00007f0e5b8d7000)

I've just installed gc-7.6.2 systemwide, and things work for me with Sage, so far.

[dimpase]

comment:6

By the way, there is #23700, which would give you the same major gc version as you apparently need (although I fail to see why).

[dimpase]

comment:7

IMHO, if you can nuke a system utility by installing a library with normal user privileges, then you have a huge security hole. Thus I don't think it's something to fix in Sage.

[vbraun]

comment:8

I think the "symbol lookup error" is just being echoed by make, its not from make not finding a symbol; This happens while R is compiling the MASS package and output is clearly being filtered.

Apparently R sets LD_LIBRARY_PATH while compiling packages so Sage's libraries take precedence over system ones. Which inevitably leads to problems, which is why we removed that from the Sage build system a while ago.

[dimpase]

comment:9

No, R is not setting LD_LIBRARY_PATH, it is merely respecting it. I think we have a case of Sage being built with LD_LIBRARY_PATH set to something, and also guile (indeed, it has nothing to do with Sage AFAIK) involved in the environment somehow; and guile (perhaps invoked from .bashrc?) made to use wrong gc version from Sage.

To reproduce this one needs to have libgc.so.X have the same X in $SAGE_LOCAL/lib and in /usr/lib. On my system X=1 in the former and X=2 in the latter,
and libguile is linked to libgc.so.2. So I went and made

$ ln -sf libgc.so.1 libgc.so.2

in $SAGE_LOCAL/lib. After this I duly get

$ LD_LIBRARY_PATH=./local/lib guile
guile: symbol lookup error: /usr/lib64/libguile-2.2.so.1: undefined symbol: GC_move_disappearing_link

Needless to say, R still builds just fine for me after this hack.

[videlec]

comment:10

For precision, my $LD_LIBRARY_PATH is empty. It should not have anything to do with it. What about the proposition of charpent comment:3?

[dimpase]

comment:11

You have a strange setup on your system, which involves guile into building Sage. Perhaps something in shell configurations, I do not know. Or something wrong with your linker settings or its cache. Guile is a system library which you can only make to fail this way by setting LD_LIBRARY_PATH. But Sage does not do it, something else does.