fabic974 random (binary) wanderings in the philament empires…

LLVM Clang: Building it from scratch - take 3 (libunwind vs libgcc_s)

Investigating the possibility of replacing usage of libgcc_s with libunwind. Edit/sept. 19th: even though this “worked”, this is not that much a correct way of doing things.


Importing libunwind from its Git repo.

[fabi@wall] ~/dev/llvm (master ✗)  git submodule add -b master git://git.sv.gnu.org/libunwind.git libunwind

Cloning into 'libunwind'...
remote: Counting objects: 14396, done.
remote: Compressing objects: 100% (3806/3806), done.
remote: Total 14396 (delta 10522), reused 14388 (delta 10518)
Receiving objects: 100% (14396/14396), 3.32 MiB | 630.00 KiB/s, done.
Resolving deltas: 100% (10522/10522), done.
Checking connectivity... done.

Building it

NOTE: There’s no ./configure script when retrieving project from its repo. and autoreconf won’t complete (errors).

Hence I’m just getting around this problem by checking out v1.1 and extracting the released tarball over it :

[fabi@wall] ~/dev/llvm/libunwind (master)  gco -b v1.1 v1.1

[fabi@wall] ~/dev/llvm/libunwind (v1.1)  wget -O - 'http://download.savannah.gnu.org/releases/libunwind/libunwind-1.1.tar.gz' |
	tar -zxf - --strip-components=1

^ Ok, empty git diff ;-

# Using my local build of Clang :
[fabi@wall] ~/dev/llvm/libunwind (v1.1 ✗)  source ../environment.sh

[fabi@wall] ~/dev/llvm/libunwind (v1.1 ✗)  ./configure --prefix=/home/fabi/dev/llvm/local | tee -a clog.txt

[fabi@wall] ~/dev/llvm/libunwind (v1.1 ✗)  time make -j2

[fabi@wall] ~/dev/llvm/libunwind (v1.1 ✗)  time make install

Testing if libgcc_s gets out of the way

Tried the symlinks quickfix below (also rebuilt test.cpp), but it didn’t work (ldd a.out lists the host’s libgcc_s) :

[fabi@wall] ~/dev/llvm/local/lib (master ✗)  ln -siv libunwind.so libgcc_s.so
‘libgcc_s.so’ -> ‘libunwind.so’
[fabi@wall] ~/dev/llvm/local/lib (master ✗)  ln -siv libunwind.a libgcc.a
‘libgcc.a’ -> ‘libunwind.a’
[fabi@wall] ~/dev/llvm/local/lib (master ✗)  ln -siv libunwind.a libgcc_eh.a

Also tried to pass -lunwind hoping resolution of symbols through libunwind would have made libgcc_s disappear, it did not :

[fabi@wall] ~/dev/llvm (master ✗)  clang++ -std=c++11 -stdlib=libc++          -nodefaultlibs -lc++ -lc++abi -lm -lc -lunwind -Wl,-rpath=$PWD/local/lib test.cpp

^ a.out did end up linked against luunwind however, but libgcc_s remained.

Patching Clang

Mainly found out what to do from “Building LLVM @ musl-libc wiki” (thanks to Jeremy Huntwork probably) – with miscellaneous investigations and trial’n’errors along the way.


[fabi@wall] ~/dev/llvm (master ✗)  time find -L llvm/ -type f \( -name '*.cpp' -o -name '*.c' \) \
    -exec grep -EHn '\-lgcc' {} \;
... output is mainly some c/o lines ...


[fabi@wall] ~/dev/llvm (master ✗)  sed -i -e 's@"-lgcc[^"]*"@"-lunwind"@g' \

Replacing the hard-coded dynamic loader :

While at it I’m replacing the hard-coded dynamic-loader “name” – I’ll be dropping a symlink there later on that points do musl-libc ;- or not :

[fabi@wall] ~/dev/llvm (master ✗)  time find -L llvm/ -type f \( -name '*.cpp' -o -name '*.c' \) \
    -exec grep -EHn 'ld-linux-x86' {} \;
... ok, one meaningful file only ...
llvm/tools/clang/lib/Driver/Tools.cpp:7231:    return "/lib64/ld-linux-x86-64.so.2";

Ok, do it :

[fabi@wall] ~/dev/llvm (master ✗)  sed -i -e 's@/lib64/ld-linux-x86-64.so.2@/lib/ld-hack-x86_64.so@' \

Symlinking under /lib :

[fabi@wall] ~/dev/llvm (master ✗)  su -
[root@wall] ~ # cd /lib

[root@wall] lib # ls ld* -l
-rwxr-xr-x 1 root root 144920 May 29 23:29 ld-2.17.so
lrwxrwxrwx 1 root root     22 Aug 24 19:33 ld-linux.so.2 -> ../lib32/ld-linux.so.2
lrwxrwxrwx 1 root root     10 Aug 24 19:33 ld-linux-x86-64.so.2 -> ld-2.17.so

[root@wall] lib # ln -s ld-2.17.so ld-hack-x86_64.so
[root@wall] lib # ln -s ../lib32/ld-linux.so.2 ld-hack.so

Setting libcxx as the default standard lib. :

[fabi@wall] ~/dev/llvm (master ✗)  sed -i '/^  return ToolChain::CST_Libstdcxx/s@stdcxx@cxx@' \

Re-bootstraping LLVM Clang!

NOTE: You did try to have it built with -stdlib=libc++ but cmake conf. tests revealed a few unexpected failed tests, like the one complaining about -fPIC being not supported by the CXX ABI.

[fabi@wall] ~/dev/llvm(master ✗)  screen
[fabi@wall] ~  cd dev/llvm

[fabi@wall] ~/dev/llvm(master ✗)  mkdir build
[fabi@wall] ~/dev/llvm(master ✗)  cd build/

[fabi@wall] ~/dev/llvm/build (master ✗)  cmake -DCMAKE_INSTALL_PREFIX=/home/fabi/dev/llvm/local ../llvm/

[fabi@wall] ~/dev/llvm/build (master ✗)  time make -j3 -l3

It’s late – tired – spent too much time doin’ these kind of “computer” things today.

End of transmission – Sept. 14th