On Tue, Sep 22, 2015 at 5:14 PM, Kirill Yukhin <kirill.yuk...@gmail.com> wrote:
> Hello,
> Patch in the bottom fixes iterator for k<logic> insns
> since QI mode is only available for AVX-512DQ.
>
> It also adds support for kshift[rl][bwdq]. This patterns
> will be used for mask load/store autogeneration on which
> Ilya Enkovich is working on.
>
> gcc/
>         * config/i386/i386.md (define_code_attr mshift): New.
>         (define_mode_iterator SWI1248_AVX512BW): Rename ...
>         (SWI1248_AVX512BW): ... to this. Make QI enabled for TARGET_AVX512DQ
>         only.
>         (define_insn "*k<logic><mode>"): Use new iterator name.
>         (define_insn "*<mshift><mode>3"): New.
>
> Bootstrapped and regtest in progress
>
> Is it ok for trunk (if regtest pass)?

This is AVX512 stuff, so - your call ;) .

Uros.

> --
> Thanks, K
>
> commit 3a521c3cfd7044008635d1c813320d3667fc1e90
> Author: Kirill Yukhin <kirill.yuk...@intel.com>
> Date:   Tue Sep 22 16:48:50 2015 +0300
>
>     AVX-512. Add kshift[lr][bwdq]. Fix iterator.
>
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> index 2f8cdb2..c0911d4 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -893,6 +893,9 @@
>  (define_code_attr shift [(ashift "sll") (lshiftrt "shr") (ashiftrt "sar")])
>  (define_code_attr vshift [(ashift "sll") (lshiftrt "srl") (ashiftrt "sra")])
>
> +;; Mask variant left right mnemonics
> +(define_code_attr mshift [(ashift "shiftl") (lshiftrt "shiftr")])
> +
>  ;; Mapping of rotate operators
>  (define_code_iterator any_rotate [rotate rotatert])
>
> @@ -935,9 +938,9 @@
>  ;; All integer modes.
>  (define_mode_iterator SWI1248x [QI HI SI DI])
>
> -;; All integer modes with AVX512BW.
> -(define_mode_iterator SWI1248_AVX512BW
> -  [QI HI (SI "TARGET_AVX512BW") (DI "TARGET_AVX512BW")])
> +;; All integer modes with AVX512BW/DQ.
> +(define_mode_iterator SWI1248_AVX512BWDQ
> +  [(QI "TARGET_AVX512DQ") HI (SI "TARGET_AVX512BW") (DI "TARGET_AVX512BW")])
>
>  ;; All integer modes without QImode.
>  (define_mode_iterator SWI248x [HI SI DI])
> @@ -7841,9 +7844,9 @@
>                             (match_dup 2)))])
>
>  (define_insn "*k<logic><mode>"
> -  [(set (match_operand:SWI1248_AVX512BW 0 "mask_reg_operand" "=k")
> -       (any_logic:SWI1248_AVX512BW (match_operand:SWI1248_AVX512BW 1 
> "mask_reg_operand" "k")
> -                         (match_operand:SWI1248_AVX512BW 2 
> "mask_reg_operand" "k")))]
> +  [(set (match_operand:SWI1248_AVX512BWDQ 0 "mask_reg_operand" "=k")
> +       (any_logic:SWI1248_AVX512BWDQ (match_operand:SWI1248_AVX512BWDQ 1 
> "mask_reg_operand" "k")
> +                         (match_operand:SWI1248_AVX512BWDQ 2 
> "mask_reg_operand" "k")))]
>    "TARGET_AVX512F"
>    {
>      if (!TARGET_AVX512DQ && <MODE>mode == QImode)
> @@ -9358,6 +9361,15 @@
>  ;; shift pair, instead using moves and sign extension for counts greater
>  ;; than 31.
>
> +(define_insn "*<mshift><mode>3"
> +  [(set (match_operand:SWI1248_AVX512BWDQ 0 "register_operand" "=k")
> +       (any_lshift:SWI1248_AVX512BWDQ (match_operand:SWI1248_AVX512BWDQ 1 
> "register_operand" "k")
> +                                      (match_operand:QI 2 
> "immediate_operand" "i")))]
> +  "TARGET_AVX512F"
> +  "k<mshift><mskmodesuffix> %2, %1, %0|%0, %1, %2"
> +  [(set_attr "type" "msklog")
> +   (set_attr "prefix" "vex")])
> +
>  (define_expand "ashl<mode>3"
>    [(set (match_operand:SDWIM 0 "<shift_operand>")
>         (ashift:SDWIM (match_operand:SDWIM 1 "<ashl_input_operand>")

Reply via email to