Re: [musl] Build error with -flto option

[Rich Felker <dalias@libc.org>]

On Tue, Mar 03, 2020 at 03:25:47PM -0500, Yuseok wrote:
> Hi all,
> 
> I got a below error message during building the musl with clang and -flto
> option.
> - my build option was ./configure CC=clang CFLAGS="-fPIC -flto -O2".
> - I also tried building with "CFLAGS="-flto -O2". However, the result was
> the same.
> - I used clang 9.
> 
> This was my error message.
> 
> > ...................................
> > j/src/unistd/setuid.o obj/src/unistd/setxid.o obj/src/unistd/sleep.o
> > obj/src/unistd/symlink.o obj/src/unistd/symlinkat.o obj/src/unistd/sync.o
> > obj/src/unistd/tcgetpgrp.o obj/src/unistd/tcsetpgrp.o
> > obj/src/unistd/truncate.o obj/src/unistd/ttyname.o
> > obj/src/unistd/ttyname_r.o obj/src/unistd/ualarm.o obj/src/unistd/unlink.o
> > obj/src/unistd/unlinkat.o obj/src/unistd/usleep.o obj/src/unistd/write.o
> > obj/src/unistd/writev.o
> > ranlib lib/libc.a
> >
> > */usr/bin/ld: error: /tmp/lto-llvm-200919.o: requires dynamic
> > R_X86_64_PC32 reloc against '__dls2' which may overflow at runtime;
> > recompile with -fPIC*/tmp/lto-llvm-200919.o:ld-temp.o:function
> > _dlstart_c: error: undefined reference to '__dls2'
> > clang: error: linker command failed with exit code 1 (use -v to see
> > invocation)
> > make: *** [Makefile:161: lib/libc.so] Error 1
> 
> 
> Could you please advise how I should handle this error?

While in theory musl's code aims to be correct so that LTO can't break
it (i.e. not wrongly relying on translation units as compiler
barriers), there are a lot of invalid transformations LLVM does that
are fairly harmless until they start happening across translation unit
boundaries. There are also things like this where LTO misses
identifying cross-asm link dependencies and drops code it needs, which
is what seems to have happened here.

A few ppl have hacked through making LTO work, and found,
unsurprisingly, that it just makes libc.so larger and slower. So I
would recommend just not using it unless you want to track down and
fix the issues on the compiler side. Until LTO is at least
theoretically okay (i.e. not risking broken floating point code,
etc.), I'm probably not going to accept anything upstream on the musl
side in the form of hacks just to make it link.

Rich