Fix -fwhole-program on LTO -- Another SPEC2000 benchmakring round

[Richard Guenther]

On Thu, 8 Oct 2009, Steven Bosscher wrote:

> On Thu, Oct 8, 2009 at 4:05 PM, Vladimir Makarov <vmakarov@redhat.com> wrote:
> (selected bits of Vlad's message:)
> > x86:
> > Â o Int2000:
> > Â  - Compiler is 1.85 times slower in pcu time with LTO (compiler
> > Â  Â  was built with --enable-checking=release).
> > Â o FP2000
> > Â  - Compiler is 2.16 times slower with LTO
> > x86_64:
> > Â o Int2000:
> > Â  - Compiler is 1.8 times slower with LTO
> > Â o FP2000
> > Â  - Compiler is 2.2 times slower with LTO
> 
> Has anyone already done the necessary profiling to see what causes
> this extra compile time?
> 
> Is this kind of slow-down expected (i.e. similar to LTO cost of other
> compilers)?

Basically at the moment each functions is built twice as the intermediate
object files have both LTO bytecode and object code (so you can link
them regularly as well, without doing link-time optimizations).

Richard.