Re: [PATCH, i386]: Committed: Fix PR target/32708

["Uros Bizjak" <ubizjak at gmail dot com>]

On 7/11/07, Jan Hubicka <hubicka@ucw.cz> wrote:

> > k8:
> >
> > long2vector:
> > .LFB3:
> >        movq    %rdi, -8(%rsp)
> >        movq    -8(%rsp), %xmm0
> >        ret
> >
> > core2:
> >
> > long2vector:
> > .LFB3:
> >        movq    %rdi, %xmm0
> >        ret
>
> Nice. However
> > +(define_insn "*vec_concatv2di_rex"
> > +  [(set (match_operand:V2DI 0 "register_operand"     "=Yt,Yi,!Yt,Yt,x,x,x")
> > +     (vec_concat:V2DI
> > +       (match_operand:DI 1 "nonimmediate_operand" "  m,r ,*y ,0 ,0,0,m")
> > +       (match_operand:DI 2 "vector_move_operand"  "  C,C ,C  ,Yt,x,m,0")))]
>
> It seems to me that for NO_INTER_UNIT_MOVES you need the variant of this
> pattern with second alternative skipped. Otherwise we will end up
> preferencing the register in some cases.  I wonder how the
> INTER_UNIT_MOVES affected your testcase above?

Jan,

please note that Yi depends on

(define_register_constraint "Yi"
"TARGET_SSE2 && TARGET_INTER_UNIT_MOVES ? SSE_REGS : NO_REGS"
"@internal Any SSE register, when SSE2 and inter-unit moves are enabled.")

So, in case of !TARGET_INTER_UNIT_MOVES, second alternative is
effectively disabled, and different code is genereated, as confirmed
by above asm dumps (k8 vs core2 dumps).

Uros.