glibc, gcc: closure-size fixes for ARM targets

[elitak]

Cross compiling to arm and most of every other target was broken by the closure-size merge.

Here're fixes for ARMv5 and 7, tested on a some packages from the small set that are actually cross-compilable to begin with.

I also disposed of builder.sh for glibc, as its only contents were a couple of hooks that I just moved into nix. I did this so that the cross/native differentiation was a bit cleaner.

[bjornfor]

History is a bit unclean (for instance, commit "works but rebasethese changes"), but +1 for fixing cross-compilation.

[elitak]

Yes, sorry, I will recapitulate before it's merged, reorganizing the fixes to different gcc versions into parallel rather than serial commits. Since this includes changes to the core of nixpkgs, I wanted to leave lots of time for review.

I would like approval especially of removing glibc's builder.sh; it may have consequences I haven't been able to imagine.

[elitak]

The CI build is failing because of logsize/timeout. Is there any way to up these limits or at least run the same tests locally?

[bjornfor]

"nox-review wip --against origin/master"

[dezgeg]

cc @edolstra for the glibc parts, I suppose.

Generally looks good to me; now test-building which will take some time...

Also what's our policy for dropping old GCC versions? 4.5 has no in-tree users, 4.6 has only one that could be replaced with a patch.

[edolstra]

Removing the glibc builder is fine.

Removing GCC 4.5 would be fine, except that it does have a in-tree user: gnat45, which is the most recent version of gnat we have. Would be nice if gnat could be upgraded to a more recent version (@viric?). Or perhaps we should just drop gnat...

[dvc94ch]

After checking out this branch and rebasing it on master I'm having an issue with:
nix-build pkgs/stdenv/linux/make-bootstrap-tools-cross.nix -A armv7l

/nix/store/g8wapfiw6y5237q3r1xwi38yqsmrvm50-binutils-2.26-armv7l-unknown-linux-gnueabi/armv7l-unknown-linux-gnueabi/bin/ld: cannot find -lcrypt
/nix/store/g8wapfiw6y5237q3r1xwi38yqsmrvm50-binutils-2.26-armv7l-unknown-linux-gnueabi/armv7l-unknown-linux-gnueabi/bin/ld: cannot find -lm
collect2: error: ld returned 1 exit status
Makefile:716: recipe for target 'busybox_unstripped' failed
make: *** [busybox_unstripped] Error 1
builder for ‘/nix/store/11kp4ysi6nxwjlvs1k0jdmg4djyczd98-busybox-1.23.2-armv7l-unknown-linux-gnueabi.drv’ failed with exit code 2
error: build of ‘/nix/store/11kp4ysi6nxwjlvs1k0jdmg4djyczd98-busybox-1.23.2-armv7l-unknown-linux-gnueabi.drv’ failed

[dezgeg]

Yes, that is still left broken. I have a branch (based on this) to make it work and will apply it at the same time as this (as both touch mass-rebuild packages) once I've cleaned it up a little.

[cleverca22]

i think this PR has caused glibc to depend on the bootstrap tools at runtime, testing further

[vcunat]

In any case, the dependency is there now, which is rather bad. I'm sorry I haven't checked better.

[vcunat]

Thanks for finding and reporting. I reverted the last staging merge to quickly avoid it for now; fixing can be done on staging later.

[viric]

Sorry, I missed the question on gnat45. If we drop gnat, we would not be able to build any ghdl, which I do use a lot. I have not tried to update gnat45 because it has been working fine until now and I have very limited Ada knowledge.

[vcunat]

e8ca9dc is the culprit. I don't yet understand why exactly.

[viric]

fwiw, you need to call the strip specific for the objects architecture. Applying x86_64 strip to arm objects breaks them.

[vcunat]

I meant I can't see why the auto-stripping doesn't strip those files; it's possible some fixes inside #15339 would make it apply. Anyway, I suppose for now I'd simply go for:

diff --git a/pkgs/development/libraries/glibc/default.nix b/pkgs/development/libraries/glibc/default.nix
index 0ea6b4b..2a1652a 100644
--- a/pkgs/development/libraries/glibc/default.nix
+++ b/pkgs/development/libraries/glibc/default.nix
@@ -72,6 +72,14 @@ in
       # Get rid of more unnecessary stuff.
       rm -rf $out/var $out/sbin/sln

+      # For some reason these aren't stripped otherwise and retain reference
+      # to bootstrap-tools; on cross-arm this stripping would break objects.
+      if [ -z "$crossConfig" ]; then
+        for i in "$out"/lib/*.a; do
+            strip -S "$i"
+        done
+      fi
+
       # Put libraries for static linking in a separate output.  Note
       # that libc_nonshared.a and libpthread_nonshared.a are required
       # for dynamically-linked applications.

[cleverca22]

i was also surprised this was even possible, the stdenv had code to prevent this

https://github.com/NixOS/nixpkgs/blob/master/pkgs/stdenv/linux/default.nix#L290

but now i see it was disabled

[vcunat]

The problem there is that allow* doesn't play well with multiple outputs, as typically some of the outputs do have a large closure and you can't specify these per-output.

[cleverca22]

i see, and the comment explains why that safety is no longer doing its job