Re: Long linking time

[Alan Modra <amodra at bigpond dot net dot au>]

On Thu, Aug 03, 2006 at 11:06:18AM +0200, Nicolas Savoire wrote:
> I don't know if this is a gcc or ld issue, but since I updated gcc to 4.1.1 
> version, the linking times of the project I work on have skyrocketted. One 
> particular shared library, which used to link in less than a second now takes  
> 6 minutes to link!!!! This library instantiates a lot of templates (use of 
> itk image library and boost).
> After further investigation, it appears that disabling debug info generation 
> during compilation or asking ld to strip debug info (-S flag) leads to normal 
> linking time (~1s). 
> I use binutils 2.17, but versions 2.15 and 2.16 exhibit the same behaviour.
> A quick profiling show that almost all the execution time is spent in qsort, 
> the call stack at that point looks like:
> #0  elf_sort_elf_symbol (arg1=0x99e49d0, arg2=0x99eb618) at elf.c:8524
> #1  0xb7eb5ddb in bsearch () from /lib/libc.so.6
> #2  0xb7eb5f3a in qsort () from /lib/libc.so.6
> #3  0x0809fbfc in bfd_elf_match_symbols_in_sections (sec1=0x81889e4, 
> sec2=0x83aa974) at elf.c:8628
> #4  0x080ac0a1 in match_group_member (sec=0x83aa974, group=0x817ad2c) at 
> elflink.c:6778
> #5  0x080ac113 in _bfd_elf_check_kept_section (sec=0x83aa974) at 
> elflink.c:6802
> #6  0x080acca0 in elf_link_input_bfd (finfo=0xbffe0684, input_bfd=0x81528c0) 
> at elflink.c:7155
> #7  0x080af013 in bfd_elf_final_link (abfd=0x8133128, info=0x8126920) at 
> elflink.c:8266
> #8  0x0805f39f in ldwrite () at ldwrite.c:557
> #9  0x0805c6ec in main (argc=82, argv=0xbffe0914) at .././ld/ldmain.c:468
> 
> I had already posted a message about this issue, but it remains without 
> answer. I hope this more descriptive version will help.

This is a ld problem.  We are paying a high price for some checks that
HJ added to validate linkonce/comdat sections.  I approved at least some
of the patches so I share some blame too.  HJ has posted some patches to
improve time spent in _bfd_elf_check_kept_section, but they haven't been
applied yet to FSF binutils.  I'm not too keen on approving them because
I think they are just bandaids, and the real fix is to revert the symbol
matching code.

You might like to try compiling HJ's latest release from
ftp.kernel.org/pub/linux/devel/binutils/ which I assume has HJ's
bandaids applied.  Please report the link times vs mainline binutils
back here, with and without -g.

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre