On Tue, Feb 12, 2019 at 11:44 AM Uros Bizjak <ubiz...@gmail.com> wrote:
>
> On Tue, Feb 12, 2019 at 8:35 PM H.J. Lu <hjl.to...@gmail.com> wrote:
> >
> > On Tue, Feb 12, 2019 at 5:43 AM Uros Bizjak <ubiz...@gmail.com> wrote:
> > >
> > > On Mon, Feb 11, 2019 at 11:55 PM H.J. Lu <hjl.to...@gmail.com> wrote:
> > > >
> > > > PR target/89021
> > > > * config/i386/i386.c (ix86_expand_vector_init_duplicate): Set
> > > > mmx_ok to true if TARGET_MMX_WITH_SSE is true.
> > > > (ix86_expand_vector_init_one_nonzero): Likewise.
> > > > (ix86_expand_vector_init_one_var): Likewise.
> > > > (ix86_expand_vector_init_general): Likewise.
> > > > (ix86_expand_vector_init): Likewise.
> > > > (ix86_expand_vector_set): Likewise.
> > > > (ix86_expand_vector_extract): Likewise.
> > > > * config/i386/mmx.md (*vec_dupv2sf): Changed to
> > > > define_insn_and_split to support SSE emulation.
> > > > (vec_setv2sf): Also allow TARGET_MMX_WITH_SSE.
> > > > (vec_extractv2sf_1 splitter): Likewise.
> > > > (vec_extractv2sfsf): Likewise.
> > > > (vec_setv2si): Likewise.
> > > > (vec_extractv2si_1 splitter): Likewise.
> > > > (vec_extractv2sisi): Likewise.
> > > > (vec_setv4hi): Likewise.
> > > > (vec_extractv4hihi): Likewise.
> > > > (vec_setv8qi): Likewise.
> > > > (vec_extractv8qiqi): Likewise.
> > > > (*vec_extractv2sf_0): Don't allow TARGET_MMX_WITH_SSE.
> > > > (*vec_extractv2sf_1): Likewise.
> > > > (*vec_extractv2si_0): Likewise.
> > > > (*vec_extractv2si_1): Likewise.
> > > > (*vec_extractv2sf_0_sse): New.
> > > > (*vec_extractv2sf_1_sse): Likewise.
> > > > (*vec_extractv2si_0_sse): Likewise.
> > > > (*vec_extractv2si_1_sse): Likewise.
> > >
> > > Please do not introduce new *_sse patterns, use mmx_isa attribute to
> > > disable unwanted alternatives.
> >
> > Will do.
> >
> > > > (define_insn_and_split "*vec_extractv2si_zext_mem"
> > > > - [(set (match_operand:DI 0 "register_operand" "=y,x,r")
> > > > + [(set (match_operand:DI 0 "register_operand" "=x,r")
> > > > (zero_extend:DI
> > > > (vec_select:SI
> > > > - (match_operand:V2SI 1 "memory_operand" "o,o,o")
> > > > + (match_operand:V2SI 1 "memory_operand" "o,o")
> > > > (parallel [(match_operand:SI 2 "const_0_to_1_operand")]))))]
> > > > - "TARGET_64BIT && TARGET_MMX"
> > > > + "TARGET_64BIT"
> > >
> > > Here you need TARGET_64BIT && (TARGET_MMX || TARGET_MMX_WITH_SSE) and
> > > mmx_isa attribute.
> > >
> >
> > Why is && (TARGET_MMX || TARGET_MMX_WITH_SSE) needed? The 3rd
> > alternative doesn't need MMX nor SSE2:
>
> Ah, I didn't notice that. LGTM then.
>
> > (define_insn_and_split "*vec_extractv2si_zext_mem"
> > [(set (match_operand:DI 0 "register_operand" "=y,x,r")
> > (zero_extend:DI
> > (vec_select:SI
> > (match_operand:V2SI 1 "memory_operand" "o,o,o")
> > (parallel [(match_operand:SI 2 "const_0_to_1_operand")]))))]
> > "TARGET_64BIT"
> > "#"
> > "&& reload_completed"
> > [(set (match_dup 0) (zero_extend:DI (match_dup 1)))]
> > {
> > operands[1] = adjust_address (operands[1], SImode, INTVAL (operands[2]) *
> > 4);
> > }
> > [(set_attr "mmx_isa" "native,sse2,base")])
>
> Please write this as "native,*,*".
Did you mean "native,sse2,*"? The second alternative is SSE2 MOVD:
MOVD (when destination operand is XMM register)
DEST[31:0] ← SRC;
DEST[127:32] ← 000000000000000000000000H;
DEST[MAXVL-1:128] (Unmodified)
> This way, it is clear that we enable alternative 0 only for native
> mmx. It looks to me that we need to add similar treatment to a couple
> of other patterns in sse.md, where we allow "y" constraint, e.g.
> *vec_concatv2sf_sse, *vec_concatv2si_sse4_1, etc.
>
I will take a look.
Thanks.
--
H.J.
