Re: [PATCH 36/40] i386: Allow MMX vector expanders with TARGET_MMX_WITH_SSE

[Uros Bizjak <ubizjak@gmail.com>]

On 2/12/19, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Tue, Feb 12, 2019 at 11:44 AM Uros Bizjak <ubizjak@gmail.com> wrote:
>>
>> On Tue, Feb 12, 2019 at 8:35 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>> >
>> > On Tue, Feb 12, 2019 at 5:43 AM Uros Bizjak <ubizjak@gmail.com> wrote:
>> > >
>> > > On Mon, Feb 11, 2019 at 11:55 PM H.J. Lu <hjl.tools@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:

No, my proposed definition is OK, see below.

> MOVD (when destination operand is XMM register)
> DEST[31:0] ← SRC;
> DEST[127:32] ← 000000000000000000000000H;
> DEST[MAXVL-1:128] (Unmodified)

You should also add "isa" attribute with "*,sse2,*", which should be
there from the beginning.

BTW: sse2 is not a member of mmx_isa. attribute.

Uros.

>> 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.
>