This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: -mfpmath=sse -mno-80387 troubles
On Fri, 28 Jan 2005 13:08:21 -0800, Richard Henderson <rth@redhat.com> wrote:
> 64-bit value into floating point. We *could* expand that sort of
> thing inline, with SSE2 using cvtpi2pd, but no one's written the
> pattern for it.
That was what i was expecting by lying about the signedness.
I suppose warning the user about that predictable mess would be too
much of a hassle.
I was using -mno-80387 as a stopgap anyway so i'm more than happy to
get rid of it.
There's still a little performance difference with ICC no matter how
much kludging i put in mainly because, AFAIK, gcc reg. allocator isn't
as smart (not a surprise) and gcc reluctancy to use fancy addressing
modes.
About that last point i rarely see gcc using a *8 scale factor (unlike
ICC) and it would obviously help with those fat assed vectors.
Is there any work going in that area (or things i should know)?