------- Comment #10 from hjl at lucon dot org 2007-04-05 20:05 -------
(In reply to comment #9)
> (In reply to comment #8)
>
> > > Please also change operands 1 and 2 of sdot_prodv8hi expander to
> > > register_operand to avoid further suprises.
> > >
> >
> > I am not sure if there is an issue since op0 of sse2_pmaddwd must be
> > register and it must be the same as op1 or op2.
>
> But only after reload. Nothing prevents gcc to generate two MEM input
> operands.
> Although reload will fix this by putting one of input operands into register
> (and match it with output operand), but this will be too late for any
> pre-reload pass to make any optimization on inputs.
>
Will something like
--- sse.md.pmaddwd 2007-04-04 17:44:43.000000000 -0700
+++ sse.md 2007-04-05 12:02:28.000000000 -0700
@@ -3044,7 +3044,7 @@
(match_operand:V8HI 1 "nonimmediate_operand" "")
(match_operand:V8HI 2 "nonimmediate_operand" "")
(match_operand:V4SI 3 "register_operand" "")]
- "TARGET_SSE2"
+ "TARGET_SSE && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
{
rtx t = gen_reg_rtx (V4SImode);
emit_insn (gen_sse2_pmaddwd (t, operands[1], operands[2]));
work?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31478
