autoPatchelfHook actually doesn't depend on stdenv and only needs
bintools (with its wrapper). This change uses $NIX_BINTOOLS instead of
$NIX_CC and makes the dependency on bintools explicit.
Fully enabling crossSystem support for autoPatchelfHook came with some
perhaps unintended consequences of being a bit more aggressive about
patching ELF files from architectures/ABIs that differ from the target
(previously, those files would be ignored because ldd usually couldn't
handle them).
This change adds architecture and rough OS ABI detection to the script
so that it doesn't try to blindly replace the interpreter of files that
can't possibly use that interpreter, and also makes sure it doesn't
accidentally use libraries of other architectures/ABIs.
`--enable-deterministic-archives` is a GNU specific strip flag and
causes other strip implementations (for example LLVM's, see #138013)
to fail. Since strip failures are ignored, this means that stripping
doesn't work at all in certain situation (causing unnecessary
dependencies etc.).
To fix this, no longer pass `--enable-deterministic-archives`
unconditionally, but instead add it in a GNU binutils specific strip
wrapper only.
`commonStripFlags` was only used for this flag, so we can remove
it altogether.
Future work could be to make a generic strip wrapper, with support for
nix-support/strip-flags-{before,after} and NIX_STRIP_FLAGS_{BEFORE,AFTER}.
This possibly overkill and unnecessary though -- also with the
additional challenge of incorporating the darwin strip wrapper somehow.
In #84415, autoPatchelfHook was taught to use the correct path to the
readelf binary when a crossSystem is specified. Unfortunately, the
remainder of the functionality in the script depended on ldd, which only
reads ELF files of its own architecture. It has the further unfortunate
quality of not reporting any useful error, but rather that the file is
not a dynamic executable.
This change uses patchelf to directly analyze the DT_NEEDED tags in the
target files instead, which correctly works across architectures. It
also updates the use of objdump to be prefix-aware $OBJDUMP (which would
have been required in the PR mentioned above, but we never made it that
far into the script execution).
This caused shebangs that were already store paths to be rewritten.
Introduced by ab4c359822 in #94642
Example difference:
$ echo "hello world" | tail -c+3
llo world
$ str="hello world"; echo ${str:3}
lo world
Using the full store hash as the random seed occasionally caused
reference cycles when the invocation was stored in output artifacts.
For example, cross-compiled gcc was failing due to this:
https://hydra.nixos.org/eval/1631713#tabs-now-fail
Simply truncating the hash is sufficient to avoid this.
Use "find -exec" to strip rather than "find … | xargs …". The former
ensures that stripping is attempted for each file, whereas the latter
will stop stripping at the first failure. Unstripped files can fool
runtime dependency detection and bloat closure sizes.
For example graphviz has chained symlinked manpages: dot2gxl.1 is
a symlink to gv2gxl.1 which is a symlink to gxl2gv.1
The second loop replaces each non-compressed symlink to a compressed
symlink. The target is determined with 'readlink -f', which follows
links recursively until the first name that is not a link (so either
the 'target name' or the first 'dangling' symlink).
This means that if the loop converted dot2gxl.1 before converting
gv2gxl.1 it would add a symlink `dot2gxl.1.gz->gxl2gv.1.gz`. When
it converted gv2gxl.1 first, it would then add a
`dot2gxl.1.gz->gv2gxl.1.gz` symlink.
Both are 'correct', but it's weird the result depends on the order
in which 'find' returns the files. This PR makes the behaviour
deterministic.
fixes#104708
