The change reverts commit 7df4387ebd
A few reasons to revert the commit:
1. The change was not enough to restore `-idirafter` override semantic
to match unwrapped compiler.
2. The change broke override semantics for cross-compilers
3. The change made override semantics different between cross-compiler
and native compiler
All of three have some overlap between, but I think it's important
to call all of them out.
The main fallout is the uboot builds, reported by cynerd.
Used the following test to check the override recovery:
$ nix shell github:NixOS/nixpkgs/release-22.05#pkgsCross.aarch64-multiplatform.stdenv.cc
$$ cat stdio.h
# empty
$$ printf "#include <stdio.h>" | aarch64-unknown-linux-gnu-gcc -E - -o - -idirafter . >/dev/null; echo $?
0
It failed before the change and succeded after.
Now that we use the standard builder, the commands produced by
pre-configure.nix are wrapped in a bash function. Inside of a bash
function, `export foo=` will still add `foo` to the environment of
any child processes forked after that point, but those variables
will *not* be visible to bash code which is outside of the
function-scope in which the `export` occurs.
Weird crap like this is yet another reason why we need to move away
from using bash for logic. Let's switch.
The bash `declare` builtin works differently when it occurs inside
function, as it now does due to breaking up the monolithic
`builder.sh` into separate phases. We have to add `-g` in order to
get the outside-of-a-bash-function behavior when using `declare`
within a bash function.
This commit adds arguments to `builder.nix`, making it a callable
function, and splits up the single massive shell script into
separate attributes for each phase. It also drops the first two
lines and the last line because these are part of the default
builder.
All script lines which were not part of a phase function have been
moved into `preUnpack` since this is the first phase that runs.
Subsequent commits will move parts of this to more sensible
locations.
This commit performs two search-and-replace operations:
- replace all `${` with `''${`
- replace `''` with `""` in shell scripts
This commit is left unsquashed to make review of the subsequent
commits easier.
It broke build: https://hydra.nixos.org/build/227264335
Motivation: coreboot-toolchain.* depend on gnat11.
It's just a hack-fix. gnat12 seems OK, just as other languages on 11,
at least those built on Hydra. /cc PR #224822 which enabled the flag.
When reporting 'gcc' bugs upstream I'm occasionally asked for the actual
configure flags I used to build gcc.
Before the change `gcc -v` always reported empty string:
$ gcc -v
...
Target: x86_64-unknown-linux-gnu
Configured with:
After the change `gcc -v` contains original options with slightly
mangled nix store paths to avoid build-only dependency retention:
$ gcc -v
...
Target: x86_64-unknown-linux-gnu
Configured with: ../source/configure --prefix=/nix/store/eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee-gcc-14.0.0 --with-gmp-include=/nix/store/eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee-gmp-6.2.1-dev/include --with-gmp-lib=/nix/store/eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee-gmp-6.2.1/lib --with-mpfr-include=/nix/store/eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee-mpfr-4.2.0-dev/include --with-mpfr-lib=/nix/store/eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee-mpfr-4.2.0/lib --with-mpc=/nix/store/eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee-libmpc-1.3.1 --with-native-system-header-dir=/nix/store/eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee-glibc-2.37-8-dev/include --with-build-sysroot=/ --program-prefix= --enable-lto --disable-libstdcxx-pch --without-included-gettext --with-system-zlib --enable-checking=release --enable-static --enable-languages=c,c++ --disable-multilib --enable-plugin --disable-libcc1 --with-isl=/nix/store/eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee-isl-0.20 --disable-bootstrap --build=x86_64-unknown-linux-gnu --host=x86_64-unknown-linux-gnu --target=x86_64-unknown-linux-gnu
While a bit verbose I think it's useful to see how we build gcc.
We want a `libgcc_s.so` to be built by the first stage
cross-compiler (withoutTargetLibc), since that is the compiler which
will compile the target libc.
This commit accomplishes that, by making three changes:
1. Replacing the `targetPlatform.libc == "msvcrt" &&` conditional
with `enableShared`, so that the code which cross-build
`libgcc_s.so` is used for all cross compilers capable of emitting
shared libraries.
2. Removing the `targetPlatform == hostPlatform` guard from the code
which produces the `libgcc` output.
3. Looking for build products in in "lib/${targetPlatform.config}/"
rather than "lib/", so we will find them when cross compiling.
This commit allows `gccCrossStageStatic` to build dynamically-linked
libraries. Since is no longer restricted to building static
libraries its name is no longer appropriate, and this commit also
renames it to the more-accurate `gccWithoutTargetLibc`.
By default, you can't build a gcc that knows how to create dynamic
libraries unless you have already built the targetPlatform libc.
Because of this, our gcc cross-compiler is built in two stages:
1. Build a cross-compiler (gccCrossStageStatic) that can build
only static libraries.
2. Use gccCrossStageStatic to compile the targetPlatform libc.
3. Use the targetPlatform libc to build a fully-capable cross
compiler.
You might notice that this pattern looks very similar to what we do
with `xgcc` in the stdenv bootstrap. Indeed it is! I would like to
work towards getting the existing stdenv bootstrap to handle cross
compilers as well. However we don't want to cripple `stdenv.xgcc`
by taking away its ability to build dynamic libraries.
It turns out that the only thing gcc needs the targetPlatform libc
for is to emit a DT_NEEDED for `-lc` into `libgcc.so`. That's it!
And since we don't use `gccCrossStageStatic` to build anything other
than libc, it's safe to omit the `DT_NEEDED` because that `libgcc`
will never be loaded by anything other than `libc`. So `libc` will
already be in the process's address space.
Other people have noticed this; crosstool-ng has been using this
approach for a very long time:
36ad0b17a7/scripts/build/cc/gcc.sh (L638-L640)
This commit deduplicates libgcc-related boilerplate which appears in
every version of our gcc expression, by moving it into libgcc.nix.
I will be submitting a separate PR which changes this boilerplate,
but that PR will be much easier to review if I can make the change
in just one place.
Meanwhile, *this* commit has no effect on eval:
$ for A in 10 11 12 13 4.8 4.9 6 7 8 9; do nix-instantiate . -A gcc$(echo $A | tr -d .); done 2>/dev/null | sort | tee before
/nix/store/1a37lnzpnz0dhm3lphiy2gcdrxgqa7ma-gcc-wrapper-4.8.5.drv
/nix/store/5szdivc8il0c3g94dq4wqnq5j77a9h6p-gcc-wrapper-11.4.0.drv
/nix/store/bmmc717wmnp1j2xkd3if5dfxicnflvn5-gcc-wrapper-7.5.0.drv
/nix/store/fc1ggpixv3wqcazchhl2hnn5zl5ds30l-gcc-wrapper-13.1.0.drv
/nix/store/j9c2b20w35r3ag5nxmklhagbwsgjhds2-gcc-wrapper-4.9.4.drv
/nix/store/nq7q57bxmsk2g457wr4b9449as3f216w-gcc-wrapper-12.3.0.drv
/nix/store/sqmkkfapzykapcs4azvxm83n786ga7q1-gcc-wrapper-10.4.0.drv
/nix/store/vxnz30i23mkl4ldsq485kxn7q0p2y4nf-gcc-wrapper-8.5.0.drv
/nix/store/yfhv0bv15cg5kj2xsb9fcgb6pdlw42v0-gcc-wrapper-6.5.0.drv
/nix/store/yi5gr75pb6kddnll10jg25hhndhkba7s-gcc-wrapper-9.5.0.drv
$ for A in 10 11 12 13 4.8 4.9 6 7 8 9; do nix-instantiate . -A gcc$(echo $A | tr -d .); done | sort | tee after
/nix/store/1a37lnzpnz0dhm3lphiy2gcdrxgqa7ma-gcc-wrapper-4.8.5.drv
/nix/store/5szdivc8il0c3g94dq4wqnq5j77a9h6p-gcc-wrapper-11.4.0.drv
/nix/store/bmmc717wmnp1j2xkd3if5dfxicnflvn5-gcc-wrapper-7.5.0.drv
/nix/store/fc1ggpixv3wqcazchhl2hnn5zl5ds30l-gcc-wrapper-13.1.0.drv
/nix/store/j9c2b20w35r3ag5nxmklhagbwsgjhds2-gcc-wrapper-4.9.4.drv
/nix/store/nq7q57bxmsk2g457wr4b9449as3f216w-gcc-wrapper-12.3.0.drv
/nix/store/sqmkkfapzykapcs4azvxm83n786ga7q1-gcc-wrapper-10.4.0.drv
/nix/store/vxnz30i23mkl4ldsq485kxn7q0p2y4nf-gcc-wrapper-8.5.0.drv
/nix/store/yfhv0bv15cg5kj2xsb9fcgb6pdlw42v0-gcc-wrapper-6.5.0.drv
/nix/store/yi5gr75pb6kddnll10jg25hhndhkba7s-gcc-wrapper-9.5.0.drv
$ diff -u after before
$
Because the build still uses a pre-4.4 version of GNU Make,
this option makes the whole build serial rather than only
the installation.
See also https://gcc.gnu.org/PR109898
This reverts commit f3995cee01.
Before this change a Darwin gcc would output binaries that avoid
aligned_alloc, regardless of target platform. That's fine for Darwin
targets, but not for non-darwin targets such as pkgsCross.raspberryPi.
This change replaces the single configure flag with a flag for
each of build, host, target.
Idea by @trofi.
It's a copy of `gcc12.cc` implementation. Nothing special added here.
Note that gccgo13 does not build yet (similar to existing gccgo12).
The failure is related to libgcc_s.so lookup problems. Not specific
to gcc-13 release.
Co-authored-by: Weijia Wang <9713184+wegank@users.noreply.github.com>
Co-authored-by: Sandro <sandro.jaeckel@gmail.com>
