The lmbench package (largely unmaintained upstream, last release in
2006) uses a home-grown build system. The scripts/build script sets
LDLIBS to -lm and Buildroot's makefile changes that to:
LDLIBS+=-lm
Except that when Buildroot passes LDLIBS="`$(PKG_CONFIG_HOST_BINARY)
--libs libtirpc`", the LDLIBS variable ends up being equal to:
-ltirpc-lm
Yes, without any space between -ltirpc and -lm.
Due to this, the checks in scripts/build that use ${LDLIBS} all fail,
and in particular the test that checks when socklen_t is a known type
fails, causing lmbench to provide its own definition, which clashes
with the C library headers definition, and therefore causing build
failures such as:
bench.h:81:13: error: conflicting types for 'socklen_t'; have 'int'
This commit fixes that by adjusting scripts/build using a patch to
properly allow passing additional ${LDLIBS} value, with the needed
space to separate the value from -lm.
Fixes:
http://autobuild.buildroot.net/results/f1715de95b46a1d08143e529bd4574bc7dbcfb3e/
We have been unable to determine exactly when this issue was
introduced. The first build failure we could find is the one
referenced above, which dates back from Aug 19, 2024. Since this date,
lmbench has been consistently failing on a very regular
basis. However, prior to Aug 19, 2024, the previous failure was from
December 2022, and was unrelated. It is unclear what changed in Aug
2024 to cause this issue to surface. The one thing that changed right
before commit ce3dedc2 (first failing commit) are changes to
genrandconfig, which ensures all autobuilders now generated fully
random configurations instead of configurations based on a
well-defined list of arch/toolchain configurations. But even with
this, this lmbench issue should have appeared earlier, and we have
been unable to find a scientific explanation.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
(cherry picked from commit 80f25d47060dea19b8f3b5fd84b6238375c54729)
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
17fedf85c5