PR other/8062
Jan Hubicka
jh@suse.cz
Thu Oct 17 08:59:00 GMT 2002
> On Thu, Oct 17, 2002 at 11:34:23AM +0200, Jan Hubicka wrote:
> > Remember, however mainline don't have my xor changes. It does xorss in
> > the other changes, so this is just consistent.
> > So should we
> > 1) consider the move patch for mainline
> > 2) fix only those two cases
> > 3) do dfmode by pxor and sfmode by xorps?
>
> I think we can fix all the xorps cases without pulling
> in the entire move patch.
OK. Can I do that as incremental patch?
I must say that after some tought, I would preffer the move patch
instead - we have reformating penalties on some loads that I think
renders SSE2 integer support to be unusable and breaks Intel
specification that states what instructions should be used to load
integer data:
(define_insn "movv4si_internal"
[(set (match_operand:V4SI 0 "nonimmediate_operand" "=x,m")
(match_operand:V4SI 1 "nonimmediate_operand" "xm,x"))]
"TARGET_SSE"
;; @@@ let's try to use movaps here.
"movaps\t{%1, %0|%0, %1}"
[(set_attr "type" "ssemov")
(set_attr "mode" "V4SF")])
And similar patterns. Of course I can submit patch to fix these
separately too.
Honza
>
>
> r~
More information about the Gcc-patches
mailing list