This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH][RFC] Enable -flto during bootstrap


On Wed, 19 Aug 2009, Richard Guenther wrote:

> 
> This patch automatically enables -flto for all host binaries and libraries
> via BOOT_CFLAGS and BOOT_LDFLAGS when lto is enabled.  It also adds
> lto to the set of stage1 languages.
> 
> Before applying this patch we should fix PR41079, the linker aggregating
> all lto sections from static libraries into the final executable.
> 
> I will followup with a bootstrap time comparison.

On a slightly loaded system I configured with
--enable-lto --enable-languages=c,lto --enable-stage1-languages=c,lto
and built the stage1.  Then the bootstrap time for finishing the
bootstrap (thus builting stage2 and stage3 and the target libs)
on an x86_64 system (with multilibs enabled) is

make -j2 BOOT_CFLAGS="-O2 -g -flto" BOOT_LDFLAGS="-Wl,-T/tmp/ltoscript 
-flto"
6461.40user 366.25system 1:05:38elapsed 173%CPU (0avgtext+0avgdata 
0maxresident)k
20976inputs+4765256outputs (19major+69723922minor)pagefaults 0swaps

vs.

make -j2
5416.07user 310.33system 53:15.59elapsed 179%CPU (0avgtext+0avgdata 
0maxresident)k
624inputs+3758600outputs (0major+59758977minor)pagefaults 0swaps

the ltoscript linker script strips the lto sections from the final
executables (otherwise they are currently 100MB vs 10MB and likely
that would affect the boostrap time).

that's a 16% overhead during stage2 and stage3 where the difference
is most prominent - you have to add the overhead of bootstrapping lto1
itself.  Not too bad IMHO and acceptable on the branch for the extra
lto testing coverage.

Richard.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]