[PATCH i386 AVX512] [19/n] Extends AVX-512 broadcasts.
Uros Bizjak
ubizjak@gmail.com
Wed Aug 20 13:16:00 GMT 2014
On Wed, Aug 20, 2014 at 2:51 PM, Kirill Yukhin <kirill.yukhin@gmail.com> wrote:
> Hello Uroš,
> On 15 Aug 20:29, Uros Bizjak wrote:
>> Can you avoid insn constraints like:
>>
>> > + "TARGET_AVX512DQ && (<MODE_SIZE> == 64 || TARGET_AVX512VL)"
>>
>> This should be split to two insn patterns, each with different
>> baseline insn constraint.
>
> I've splitted pattern into two similar w/ different mode iterators.
>
> Bootstrapped and avx512-regtested.
>
> Is it ok for trunk?
>
> gcc/
> * config/i386/sse.md
> (define_mode_iterator VI4F_BRCST32x2): New.
> (define_mode_attr 64x2mode): Ditto.
> (define_mode_attr 32x2mode): Ditto.
> (define_insn "<mask_codefor>avx512dq_broadcast<mode><mask_name>"
> with VI4F_BRCST32x2 mode iterator): Ditto.
> (define_insn "<mask_codefor>avx512vl_broadcast<mode><mask_name>_1"): Ditto.
> (define_insn "<mask_codefor>avx512dq_broadcast<mode><mask_name>_1"
> with V16FI mode iterator): Ditto.
> (define_insn "<mask_codefor>avx512dq_broadcast<mode><mask_name>_1"
> with VI8F_512 mode iterator): Ditto.
> (define_insn "<mask_codefor>avx512dq_broadcast<mode><mask_name>_1"
> with VI8F_256 mode iterator): Ditto.
> --
> Thanks, K
>
> diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> index 4632b3a..6a5faee 100644
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -482,6 +482,7 @@
> (define_mode_iterator VI8F_128 [V2DI V2DF])
> (define_mode_iterator VI4F_256 [V8SI V8SF])
> (define_mode_iterator VI8F_256 [V4DI V4DF])
> +(define_mode_iterator VI8F_512 [V8DI V8DF])
> (define_mode_iterator VI8F_256_512
> [V4DI V4DF (V8DI "TARGET_AVX512F") (V8DF "TARGET_AVX512F")])
> (define_mode_iterator VI48F_256_512
> @@ -14514,6 +14515,83 @@
> (set_attr "prefix" "vex")
> (set_attr "mode" "<sseinsnmode>")])
>
> +;; For broadcast[i|f]32x2. Yes there is no v4sf version, only v4si.
> +(define_mode_iterator VI4F_BRCST32x2
> + [V16SI (V8SI "TARGET_AVX512VL") (V4SI "TARGET_AVX512VL")
> + V16SF (V8SF "TARGET_AVX512VL")])
> +
> +(define_mode_attr 64x2mode
> + [(V8DF "V2DF") (V8DI "V2DI") (V4DI "V2DI") (V4DF "V2DF")])
> +
> +(define_mode_attr 32x2mode
> + [(V16SF "V2SF") (V16SI "V2SI") (V8SI "V2SI")
> + (V8SF "V2SF") (V4SI "V2SI")])
> +
> +(define_insn "<mask_codefor>avx512dq_broadcast<mode><mask_name>"
> + [(set (match_operand:VI4F_BRCST32x2 0 "register_operand" "=v")
> + (vec_duplicate:VI4F_BRCST32x2
> + (vec_select:<32x2mode>
> + (match_operand:<ssexmmmode> 1 "nonimmediate_operand" "vm")
> + (parallel [(const_int 0) (const_int 1)]))))]
> + "TARGET_AVX512DQ"
> + "vbroadcast<shuffletype>32x2\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}"
> + [(set_attr "type" "ssemov")
> + (set_attr "prefix_extra" "1")
> + (set_attr "prefix" "evex")
> + (set_attr "mode" "<sseinsnmode>")])
> +
> +(define_insn "<mask_codefor>avx512vl_broadcast<mode><mask_name>_1"
> + [(set (match_operand:VI4F_256 0 "register_operand" "=v,v")
> + (vec_duplicate:VI4F_256
> + (match_operand:<ssexmmmode> 1 "nonimmediate_operand" "v,m")))]
> + "TARGET_AVX512VL"
> + "@
> + vshuf<shuffletype>32x4\t{$0x0, %t1, %t1, %0<mask_operand2>|%0<mask_operand2>, %t1, %t1, 0x0}
> + vbroadcast<shuffletype>32x4\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}"
> + [(set_attr "type" "ssemov")
> + (set_attr "prefix_extra" "1")
> + (set_attr "prefix" "evex")
> + (set_attr "mode" "<sseinsnmode>")])
> +
> +(define_insn "<mask_codefor>avx512dq_broadcast<mode><mask_name>_1"
> + [(set (match_operand:V16FI 0 "register_operand" "=v,v")
> + (vec_duplicate:V16FI
> + (match_operand:<ssehalfvecmode> 1 "nonimmediate_operand" "v,m")))]
> + "TARGET_AVX512DQ"
> + "@
> + vshuf<shuffletype>32x4\t{$0x44, %g1, %g1, %0<mask_operand2>|%0<mask_operand2>, %g1, %g1, 0x44}
> + vbroadcast<shuffletype>32x8\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}"
> + [(set_attr "type" "ssemov")
> + (set_attr "prefix_extra" "1")
> + (set_attr "prefix" "evex")
> + (set_attr "mode" "<sseinsnmode>")])
> +
> +(define_insn "<mask_codefor>avx512dq_broadcast<mode><mask_name>_1"
> + [(set (match_operand:VI8F_512 0 "register_operand" "=v,v")
> + (vec_duplicate:VI8F_512
> + (match_operand:<64x2mode> 1 "nonimmediate_operand" "v,m")))]
> + "TARGET_AVX512DQ"
> + "@
> + vshuf<shuffletype>64x2\t{$0x0, %g1, %g1, %0<mask_operand2>|%0<mask_operand2>, %g1, %g1, 0x0}
> + vbroadcast<shuffletype>64x2\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}"
> + [(set_attr "type" "ssemov")
> + (set_attr "prefix_extra" "1")
> + (set_attr "prefix" "evex")
> + (set_attr "mode" "<sseinsnmode>")])
> +
> +(define_insn "<mask_codefor>avx512dq_broadcast<mode><mask_name>_1"
> + [(set (match_operand:VI8F_256 0 "register_operand" "=v,v")
> + (vec_duplicate:VI8F_256
> + (match_operand:<64x2mode> 1 "nonimmediate_operand" "v,m")))]
> + "TARGET_AVX512DQ && TARGET_AVX512VL"
> + "@
> + vshuf<shuffletype>64x2\t{$0x0, %t1, %t1, %0<mask_operand2>|%0<mask_operand2>, %t1, %t1, 0x0}
> + vbroadcast<shuffletype>64x2\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}"
> + [(set_attr "type" "ssemov")
> + (set_attr "prefix_extra" "1")
> + (set_attr "prefix" "evex")
> + (set_attr "mode" "<sseinsnmode>")])
Please merge the above two patterns. Please introduce VI8F_BRCST64x2
in the same way as for broadcast32x2. Using TARGET_AVX512VL condition
in the mode constraint and TARGET_AVX512DQ as the baseline insn
constraint, you will also simplify insn constraint.
OK with this change.
Thanks,
Uros.
More information about the Gcc-patches
mailing list