New:
- mold is now up to 10% faster when linking very large, debug
info-enabled executables such as Blender (~1.8 GiB) or Clang (~3.8
GiB), thanks to several improvements we've made to the string merging
algorithm. (53ebcd8, d714301, 40f6b17, c9faf3d)
- -z start-stop-visibility=hidden is now supported so that
linker-synthesized __start_<section-name> and __stop_<section-name>
symbols can be completely hidden from other ELF modules. Previously,
only -z start-stop-visibility=protected was supported. (99a5b15)
- -Bsymbolic-non-weak and -Bsymbolic-non-weak-functions options are now
supported for compatibility with LLVM lld. Just like lld, these options
control which symbols are exported as dynamic symbols.
-Bsymbolic-non-weak makes the linker to export only weak symbols,
whereas -Bsymbolic-non-weak-functions makes it to export only weak
function symbols. (7d17aa8)
Bug fixes and compatibility improvements:
- Previously, if a linker script contains a newline character in the
beginning four bytes of a file, it was not recognized as a linker
script by mold. Now, mold allows newlines at the beginning of a file.
(ea054cc)
- Under rare circumstances, the INPUT linker script command may have
found a different file than GNU ld would. Now, mold's behavior aligns
with GNU ld's. (163975d)
- Previously, the --repro option produced corrupted tar files. Now the
bug has been fixed. (32c4a09)
- mold generally guarantees that its output is reproducible, meaning that
if you run the linker with the exact same command line options and
input files, the output is guaranteed to be bit-for-bit identical to
the previous outputs. However, under rare circumstances, it might
produce different output due to a bug. It's reported that this
nondeterminism caused random crashes for some programs (#1247). This
bug has been fixed. (6463a7c)
- mold no longer sets the address of the .text section as the entry point
address if --entry option is not given, just like LLVM lld. (020b1a7)
- [RISC-V] __global_pointer$ symbol is now exported from executables as
required by the processor-specific ABI. (3df7c8e)
- [ARM32] --long-plt option is now recognized as known option by mold.
mold ignores the option, though, because the PLTs generated by our
linker is always long. (d432e98)
Release Notes:
https://github.com/rui314/mold/releases/tag/v2.31.0
Signed-off-by: Sean Khan <datapronix@protonmail.com>
Link: https://github.com/openwrt/openwrt/pull/15403
Signed-off-by: Robert Marko <robimarko@gmail.com>
Some package might require to fix their pkg-config file to point to host
or hostpkg file. This is the case for glib2 library that provides with
pkg-config variables, tools to generates files from xml. Those tools
should use the host binary instead of the targets one to correctly build
packages that makes use of such tools.
Link: https://github.com/openwrt/openwrt/pull/15134
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
In the 0.191 update dirname was used instead of <libgen.h> to fix the
poisoned basename error:
/usr/include/libgen.h:35:9: error: attempt to use poisoned "basename"
35 | #define basename __xpg_basename
However, doing this has lead to libelf.a pulling in xmalloc, xstrdup and
friends and statically linking them thus leading to a symbol name conflict
with FRR host build and anything else that links against libelf and uses
xmalloc and friends.
Well, it turns out that upstream has added a helper[1] for basename so it
can compile with musl 1.2.5 which dropped the basename declaration, but it
also means that we must NOT include <libgen.h> and that poisoned error is
intentional and added to prevent duplicate basename definitions.
This also means that for macOS we dont need to do any additional header
inclusions as the new helper takes care of basename.
So, to fix the symbol conflict we can simply drop the <dirname.h> inclusion
and build from elfutils.
Tested on Fedora 40 as well as macOS 14.4.1.
[1] https://sourceware.org/git/?p=elfutils.git;a=commit;h=a2194f6b305bf0d0b9dd49dccd0a5c21994c8eeaFixes: #24030
Fixes: b6f025b424 ("tools/elfutils: update to 1.91")
Link: https://github.com/openwrt/openwrt/pull/15337
Signed-off-by: Robert Marko <robimarko@gmail.com>
Quilt refresh combined two sets of changes to the same file.
The switch from using libgen.h to dirname.h because of function poisoning
from gnulib's import of basename() was added as a new patch hunk instead
of an edit to the original one.
The original patch hunk was to fix build errors on an earlier version of
elfutils before the "dirname" module was being imported to fix further
build errors with the 0.191 version.
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
A false tdestroy() function was added in order to make elfutils build on
macOS again. A previous commit added declarations for a real version of
tdestroy() into gnulib, which is already imported, as well as the
preprocessor flags and the triggers for the Makefile.am conditional in
order to include the source to be built.
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
On macOS, stdlib.h in the standard include paths does not provide
reallocarray() while both elfutils and gnulib do, however they are
declared differently, leading to an error:
./system.h:101:1: error: static declaration of 'reallocarray' follows non-static declaration
reallocarray (void *ptr, size_t nmemb, size_t size)
A normal "configure && make" build cycle results in both declarations
being enabled as a result of both elfutils and gnulib having completely
separate configure checks where gnulib uses an internal placeholder symbol
HAVE_REALLOCARRAY, and elfutils uses a standard autoconf macro
HAVE_DECL_REALLOCARRAY.
Fix this by excluding the import of the reallocarray module which causes
gnulib checks in the configure stage to not even consider whether to
declare reallocarray later on, so the decision is only between the
standard include stdlib.h and the elfutils header.
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
Organize the Makefile lines involved in gnulib importing and its
workarounds. It improves readability and keeps git history organized.
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
Instead of editing the SUBDIRS variable with a patch, it can be overriden
at the end of the command line when invoking Make.
This tool has a series of recursive Makefiles in each subdirectory,
therefore SUBDIRS is set to a pattern of Make functions so that the result
is variable depending on the current subdirectory that Make is being
invoked in.
It's not necessary to have gnulib-cache.m4 in EXTRA_DIST since we don't
need to re-import after packaging this in the SDK, so get rid of the
entire patch hunk for ./Makefile.am
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
Release Notes:
https://sourceware.org/pipermail/elfutils-devel/2024q1/006876.html
Manually refresh:
- 100-portability.patch
Change:
- replace libgen.h with gnulib "dirname" module for compilation errors:
In file included from ./../libdw/libdwP.h:38,
from eblobjnote.c:42:
/usr/include/libgen.h:35:9: error: attempt to use poisoned "basename"
35 | #define basename __xpg_basename
| ^
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Co-Developed-by: Nick Hainke <vincent@systemli.org>
Signed-off-by: Nick Hainke <vincent@systemli.org>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
Update to latest stable release.
The following commits in gnulib caused a conflict in locally bootstrapped
coreutils with stable gnulib:
8f4b4e52c991de2233b471f8e35a068866b31f01
2749234203959df8d72cd8638d4e00a9fff450db
A module (strftime) was marked deprecated and replaced by another module
(nstrftime) in the version of gnulib that coreutils was released with
compared to the stable branch that we use for importing. Conflicts from
the previous version of coreutils are now gone, so other imported headers
are now good.
Refresh patch:
- 000-bootstrap.patch
Remove upstreamed patch:
- 001-bootstrap-sync.patch
Link: https://lists.gnu.org/archive/html/coreutils/2024-03/msg00132.html
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
Update to latest stable release.
Add configure option to disable support for the Year 2038 problem.
(for now, as some versions of GCC do not yet support it)
Syncing bootstrap script fails, backport an upstream patch which can be
removed at next coreutils update.
Several headers from the stable gnulib branch cause build failure because
the changes in the imported versions are incompatible with the Makefile
that gets generated for coreutils. This version of coreutils was released
after being bootstrapped and autoreconf'ed with a significantly different
version of gnulib compared to our local gnulib, so skip importing them
(and restore the backup).
While at it, organize restoring the originally shipped version of files
into a Make foreach function.
Refresh patch:
- 000-bootstrap.patch
New patch:
- 001-bootstrap-sync.patch
Link: https://lists.gnu.org/archive/html/coreutils/2023-08/msg00099.html
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
Force bison to ignore the M4 environment variable and hardcode it to the
locally built m4 during build operations using the relocatable path
variable STAGING_DIR_HOST.
This allows bison to continue to function while we are forcefully avoiding
autoreconf and other autoconf and automake-like operations by giving a
fake path to m4 with the M4 environment variable.
The specific path can still be overridden independently from the
environment within the line of invocation that runs bison by setting
STAGING_DIR_HOST within the command, so document this in the help printout.
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
The tdestroy() function, which is a GNU extension to the standard C
library, is defined in gnulib in tsearch.c but is missing it's
corresponding declaration in search.in.h by being completely missing...
This patch is large but upstreamable, including all of the macros and
conditionals and configure checks that upstream GNU would expect for
portable support, like using the @@ placeholder/substitution method to
determine whether or not to have declarations based on whether or not
tdestroy() is already declared within the standard headers of the default
include paths.
There were also some typedefs and aliases missing, along with the warnings
and preprocessor exceptions that need to be added for consistency with the
usage of the rest of the functions in the files.
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
For modules that depend on the reallocarray module, like ialloc, xalloc,
and safe-alloc, it was not possible to skip importing the reallocarray
module as they all contained at least one function that called
reallocarray() and would cause build failure if the host system didn't
declare it.
This upstreamable patch adds macros that toggle whether to define
functions that depend on reallocarray() based on whether the reallocarray
module is being imported.
Tested-by: Georgi Valkov <gvalkov@gmail.com> # MacOS
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
There are other wrapper scripts released with makeinfo like texi2pdf which
are required by the build prerequisites of some tools, and have a similar
purpose and usage.
Let the makeinfo perl script handle all of these cases.
It's worth mentioning that "texi2any" is the actual program and "makeinfo"
is one of it's aliases. From upstream GNU:
makeinfo: texi2any
rm -f $@
-$(LN_S) texi2any $@
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
Currently, trying to compile LLVM-BPF will fail with:
[2225/3517] Linking CXX shared library lib/libLLVM-15.so
FAILED: lib/libLLVM-15.so
/usr/bin/ld: staging_dir/host/lib/libzstd.a(zstd_common.o): relocation R_X86_64_32 against `.rodata.str1.1' can not be used when making a shared object; recompile with -fPIC
/usr/bin/ld: failed to set dynamic section sizes: bad value
collect2: error: ld returned 1 exit status
So, to fix it enable PIC for the host ZSTD.
Fixes: #15247
Signed-off-by: Bryan Roessler <bryanroessler@gmail.com>
Proposed fixup has been replaced and merged with an advanced version.
install-pc-mt has been dropped and replaced for intall-pc MT=1 to
generate a .pc file with multithread libs.
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
Trying to compile with new new enough GCC but older binutils that dont
support AVX-VNNI will error out on the assembler, so backport an upstream
fix for it.
Fixes: 338b463e1e ("tools: libdeflate: update to 1.20")
Signed-off-by: Robert Marko <robimarko@gmail.com>
Changes:
* Improved CRC-32 performance on recent x86 CPUs by adding
VPCLMULQDQ-accelerated implementations using 256-bit and 512-bit vectors.
* Improved Adler-32 performance on recent x86 CPUs by adding
VNNI-accelerated implementations using 256-bit and 512-bit vectors.
* Improved CRC-32 and Adler-32 performance on short inputs.
* Optimized the portable implementation of Adler-32.
* Added some basic optimizations for RISC-V.
* Dropped support for gcc versions older than v4.9 (released in 2014) and
clang versions older than v3.9 (released in 2016).
* Dropped support for CRC-32 acceleration on 32-bit ARM using the ARMv8 pmull or crc32 instructions.
This code only worked on CPUs that also have a 64-bit mode, and it was
already disabled on many compiler versions due to compiler limitations.
CRC-32 acceleration remains fully supported on 64-bit ARM.
Signed-off-by: Robert Marko <robimarko@gmail.com>
Now that instead of relying on env variables for the GH download script
invoking ZSTD tarball compression it passes the full arguments via tar -I
we can drop the CLI max compression level override.
Signed-off-by: Robert Marko <robimarko@gmail.com>
The current .pc generated by make build system for zstd is wrong and
always assume a single-threaded library is used.
This wasn't the case for the meson build system that used his own logic
to generate the pkg-config file.
Add a patch to fix this by introducing install-mt-pc make target to
generate the correct pkg-config gile.
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
This reverts commit 8d934c1196.
The update seems to be causing issues that need to be further
explored [0]. Let's revert it.
[0] - https://github.com/openwrt/openwrt/pull/15078
Signed-off-by: Nick Hainke <vincent@systemli.org>
We forgot to make sure install-includes is called for the libzstd in
order for it to install the required headers.
Fixes: 4b920e799f ("tools: zstd: convert to make and drop meson dependency")
Signed-off-by: Robert Marko <robimarko@gmail.com>
When using zst instead of xz, the hash changes. This commit fixes the
hash for packages and tools in core.
Signed-off-by: Paul Spooren <mail@aparcar.org>
In the light of recent XZ events, and fundamental XZ issues lets work on
moving away from using XZ.
So, use gz compressed tarballs as sources whenever possible.
dwarves only offers bz2 compressed tarballs, so use those as size
difference is minor compared to XZ.
Signed-off-by: Robert Marko <robimarko@gmail.com>
dwarves
libdeflate is currently intentionally being fetched via GIT.
However, with the move to using ZSTD to compress the cloned GIT repo
tarballs it means that we would first need to compile ZSTD.
But that means that we need to be able to unpack gzipped tarballs first
which we currently do by using libdeflate-gzip.
So, in order to do so lets fetch libdeflate as a tarball, use gzip to
extract it and then use libdeflate as regular for all other tools.
Signed-off-by: Robert Marko <robimarko@gmail.com>
ZSTD and libdeflate do not depend on SED nor flock, so instead of the whole
for loop that will filter 2 out of 4 core packages just specify that patch
and tar depend on sed explicitly.
ZSTD now depends on libdeflate since libdeflate-gzip will then be used to
unpack ZSTD as well as most tool archives.
Signed-off-by: Rosen Penev <rosenp@gmail.com>
Signed-off-by: Robert Marko <robimarko@gmail.com>
We dont really have a reason to deviate from the upstream default for
ZSTD_LEGACY_SUPPORT value of 5, as it will save a bit of space but
prevent decompressing data compressed with legacy ZSTD versions.
Signed-off-by: Robert Marko <robimarko@gmail.com>
ZTSD limits the safe compression level to a max of 19 as 20 to 22 cause
increased RAM usage. Higher levels require --ultra arg passed.
There isn't currently a way to set --ultra using ENV options similar to
ZSTD_CLEVEL and ZSTD_CLEVEL is limited to 19.
To fix this, we can increase the max safe compression level by providing
a custom ZSTDCLI_CLEVEL_MAX value with CFLAGS.
The max safe level is increased to 20.
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
Convert to make and drop meson dependency since it's not a core tools
and can't depend on advanced build system like cmake or meson.
On top of this make is the official build support and cmake/meson are
supported by 3rd parties.
Signed-off-by: Christian Marangi <ansuelsmth@gmail.com>
GCC14 no longer treats integer types and pointer types as equivalent in
assignments (including implied assignments of function arguments and return
values), and instead fails the compilation with a type error.
So, as a workaround lets disable the newly introduced error
-Werror=int-conversion and just make it print a warning to enable compiling
with GCC14 as Fedora 40 now defaults to it.
Signed-off-by: Robert Marko <robimarko@gmail.com>
