Re: PING: PATCH: Add --with-math=sse for x86

[Paolo Bonzini <bonzini at gnu dot org>]

> > 1. Improve 32bit FP performance significantly on current
> > x86 processors without changing Makefiles. SPEC CPU
> > FP can be improved by as much as 20%.
> > 2. A new psABI designed for SSE math won't happen
> > in the foreseeable future.  -with-math=sse provides a
> > a way to improve FP performance with backward binary
> > compatibility.
> > 3. The performance gain is automatic since it requires
> > very few changes in software packages if any changes
> > are needed at all.
> > 4. MacOS chose Linux psABI with SSE math enabled by
> > default and increased alignment when they defined a new
> > psABI. It is an indication that SSE math can be used on
> > entire OS with improved FP performance.
> >
> > OK for trunk?

As a build maintainer, no, I don't think this is okay.  It's okay if 
other people override my thoughts, but an option is forever and you are 
just adding one to paper over RA deficiencies.  What you actually want 
is to use 387 on function-call heavy functions, and SSE on the others. 
The right way would be to make -mfpmath=sse,387 work well enough that we 
can enable it by default.  Maybe after IRA is merged.

In the meanwhile, I don't remember if the appropriate -march options 
enable -msse, -msse2, etc.  If not, that's definitely a good thing to do.

Paolo