package/Makefile.in: work around buggy buildsystems wrt LDFLAGS

Some buildsystems (or their use of it by packages) will cause flags in
LDFLAGS to be re-orderded, or even dropped, causing some
incomprehensible mayhem... For example, gpsd [0] has this in its
SConscript [1](typoes not mines, for once):

    591 # scons uses gcc, or clang, to link. Thus LDFLAGS does not serve its
    592 # traditional function of providing arguments to ln. LDFLAGS set in the
    593 # environment before running scons get moved into CCFLAGS by scons.
    594 # LDFLAGS set while running scons get ignored.
    [--SNIP--]
    611 for i in ["ARFLAGS",
    612           "CCFLAGS",
    613           "CFLAGS",
    614           "CPPFLAGS",
    615           "CXXFLAGS",
    616           "LDFLAGS",
    617           "LINKFLAGS",
    618           "SHLINKFLAGS",
    619           ]:
    620     if i in os.environ:
    621         # MergeFlags() puts the options where scons wants them, not
    622         # where you asked them to go.
    623         env.MergeFlags(Split(os.getenv(i)))

So, when LDFLAGS (our TARGET_LDFLAGS) contains "-z text"  (without the
quotes), that gets turned into a command line like (line-splitted for
readability):

    [...]/buildroot/output/host/bin/aarch64_be-buildroot-linux-musl-gcc \
        -o gpsd-3.25/drivers/driver_rtcm2.os \
        -c \
        --sysroot=[...]/buildroot/output/host/aarch64_be-buildroot-linux-musl/sysroot \
        -O3 -g0 \
        -z \
        -fPIC -D_LARGEFILE_SOURCE -D_LARGEFILE64_SOURCE -D_FILE_OFFSET_BITS=64 \
        gpsd-3.25/drivers/driver_rtcm2.c

Notice how there is a lone "-z" without any following keyword.

This then causes a build failure that looks totally unrelated [2]:

    In file included from gpsd-3.25/drivers/../include/gpsd.h:36,
                     from gpsd-3.25/drivers/driver_rtcm2.c:65:
    gpsd-3.25/drivers/../include/os_compat.h:40:8: error: redefinition of ‘struct timespec’
       40 | struct timespec {
          |        ^~~~~~~~
    In file included from [...]/buildroot/output/host/aarch64_be-buildroot-linux-musl/sysroot/usr/include/sys/select.h:16,
                     from gpsd-3.25/drivers/../include/gpsd.h:31:
    [...]/buildroot/output/host/aarch64_be-buildroot-linux-musl/sysroot/usr/include/bits/alltypes.h:237:8: note: originally defined here
      237 | struct timespec { time_t tv_sec; int :8*(sizeof(time_t)-sizeof(long))*(__BYTE_ORDER==4321); long tv_nsec; int :8*(sizeof(time_t)-sizeof(long))*(__BYTE_ORDER!=4321); };
          |        ^~~~~~~~
    gpsd-3.25/drivers/../include/os_compat.h:48:5: error: conflicting types for ‘clock_gettime’; have ‘int(clockid_t,  struct timespec *)’ {aka ‘int(int,  struct timespec *)’}
       48 | int clock_gettime(clockid_t, struct timespec *);
          |     ^~~~~~~~~~~~~
    In file included from gpsd-3.25/drivers/../include/gpsd.h:33:
    [...]/buildroot/output/host/aarch64_be-buildroot-linux-musl/sysroot/usr/include/time.h:104:5: note: previous declaration of ‘clock_gettime’ with type ‘int(clockid_t,  struct timespec *)’ {aka ‘int(int,  struct timespec *)’}
      104 | int clock_gettime (clockid_t, struct timespec *);
          |     ^~~~~~~~~~~~~
    scons: *** [gpsd-3.25/drivers/driver_rtcm2.os] Error 1
    scons: building terminated because of errors.
    make[1]: *** [package/pkg-generic.mk:289: [...]/buildroot/output/build/gpsd-3.25/.stamp_built] Error 2
    make: *** [Makefile:83: _all] Error 2

Although undocumented, neither in gcc not ld (clang unchecked by lack of
a clang toolchain here, and by lack of clang knowledge), -z accepts the
keyword to be snatch-glued onto it, like -zkeyword, rather than be
spearated with a space. So, use that to pass -ztext.

Fixes:
    http://autobuild.buildroot.org/results/c03/c039989947b960ac6af17c87090366abc26dcb6d/
    http://autobuild.buildroot.net/results/bc35d3e7b0e8c59c776652070650af3c749250ee/

[0] Our other scons-based package, mongodb, does not build for another
reason that probably hides the same issue as seen with gpsd.

[1] As explained in gpsd's SConscript (se above), Scons does play tricks
with variables:
    https://scons.org/doc/production/HTML/scons-man.html#f-MergeFlags
    https://scons.org/doc/production/HTML/scons-man.html#f-ParseFlags

Quoting:
    Flag values are translated according to the prefix found, and added
    to the following construction variables:
    [...]
    -Wl,                    LINKFLAGS
    [...]
    -                       CCFLAGS
    [...]
    Any other strings not associated with options are assumed to be the
    names of libraries and added to the $LIBS construction variable.

So in our case, it finds that -z is an unknown option that matches the
'-' prefix, so it is added to CFLAGS, while 'text' is a string on its
own, so added to LIBS, and thus it would try to link with -ltext
(supposedly, because we do not even go that far). Funnily enough, we can
se that "-Wl," is a known option prefix, that is added to LINKFLAGS (to
properly be used with gcc or clang, not ld).

As a consequence, gpsd's buildsystem does drop -ztext from the link
flags, and only passes it to the compile flags, which brings us back to
before we banned textrels in a1a2f498d7 (package/Makefile.in: ban
textrels on musl toolchains). Fixing gpsd is a task for another,
separate patch...

[2] I spent quite some time to look at recent, time-related changes in
Buildroot, especially due to the infamous time64_t issues... Alas, that
was not related, and only a git-bisect pinpointed the actual issue. Poor
polar bears...

Signed-off-by: Yann E. MORIN <yann.morin@orange.com>
Cc: J. Neuschäfer <j.neuschaefer@gmx.net>
Reviewed-by: J. Neuschäfer <j.neuschaefer@gmx.net>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
This commit is contained in:
Yann E. MORIN 2024-08-09 15:08:22 +02:00 committed by Thomas Petazzoni
parent 33557477c0
commit 3a39e706a9

View File

@ -155,12 +155,17 @@ TARGET_LDFLAGS = $(call qstrip,$(BR2_TARGET_LDFLAGS))
#
# See also: https://www.openwall.com/lists/musl/2020/09/25/4
#
# NOTE: We're using "-z text" instead of "-Wl,-z,text" here, because some
# NOTE: We're using "-ztext" instead of "-Wl,-z,text" here, because some
# packages pass TARGET_LDFLAGS directly to ld rather than gcc, and ld doesn't
# support -Wl,[...]. -z is supported by both gcc and clang, so it probably
# won't cause us problems.
#
# We're using "-ztext" instead of "-z text" here, because some buildsystems
# (like scons, for gpsd) will reorder and/or drop LDFLAGS, causing a lone
# "-z" to be passed and the "text" keyword to be dropped otherwise. Both
# gcc and ld supports that, so it probably won't cause us problems.
ifeq ($(BR2_TOOLCHAIN_USES_MUSL):$(BR2_STATIC_LIBS),y:)
TARGET_LDFLAGS += -z text
TARGET_LDFLAGS += -ztext
endif
# By design, _FORTIFY_SOURCE requires gcc optimization to be enabled.