Re: [PATCH 08/20] tcg/i386: Implement avx512 variable shifts

From: "Alex Bennée" <alex.bennee@linaro.org>
To: Richard Henderson <richard.henderson@linaro.org>
Cc: qemu-devel@nongnu.org
Subject: Re: [PATCH 08/20] tcg/i386: Implement avx512 variable shifts
Date: Tue, 01 Feb 2022 20:33:29 +0000	[thread overview]
Message-ID: <87h79iwcmb.fsf@linaro.org> (raw)
In-Reply-To: <20211218194250.247633-9-richard.henderson@linaro.org>

Richard Henderson <richard.henderson@linaro.org> writes:

> AVX512VL has VPSRAVQ, and
> AVX512BW has VPSLLVW, VPSRAVW, VPSRLVW.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

Reviewed-by: Alex Bennée <alex.bennee@linaro.org>

> ---
>  tcg/i386/tcg-target.c.inc | 32 ++++++++++++++++++++++++--------
>  1 file changed, 24 insertions(+), 8 deletions(-)
>
> diff --git a/tcg/i386/tcg-target.c.inc b/tcg/i386/tcg-target.c.inc
> index 316e550b38..7b9302fcc2 100644
> --- a/tcg/i386/tcg-target.c.inc
> +++ b/tcg/i386/tcg-target.c.inc
> @@ -418,9 +418,13 @@ static bool tcg_target_const_match(int64_t val, TCGType type, int ct)
>  #define OPC_VPBROADCASTQ (0x59 | P_EXT38 | P_DATA16)
>  #define OPC_VPERMQ      (0x00 | P_EXT3A | P_DATA16 | P_VEXW)
>  #define OPC_VPERM2I128  (0x46 | P_EXT3A | P_DATA16 | P_VEXL)
> +#define OPC_VPSLLVW     (0x12 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
>  #define OPC_VPSLLVD     (0x47 | P_EXT38 | P_DATA16)
>  #define OPC_VPSLLVQ     (0x47 | P_EXT38 | P_DATA16 | P_VEXW)
> +#define OPC_VPSRAVW     (0x11 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
>  #define OPC_VPSRAVD     (0x46 | P_EXT38 | P_DATA16)
> +#define OPC_VPSRAVQ     (0x46 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
> +#define OPC_VPSRLVW     (0x10 | P_EXT38 | P_DATA16 | P_VEXW | P_EVEX)
>  #define OPC_VPSRLVD     (0x45 | P_EXT38 | P_DATA16)
>  #define OPC_VPSRLVQ     (0x45 | P_EXT38 | P_DATA16 | P_VEXW)
>  #define OPC_VZEROUPPER  (0x77 | P_EXT)
> @@ -2742,16 +2746,13 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
>          OPC_PMAXUB, OPC_PMAXUW, OPC_PMAXUD, OPC_UD2
>      };
>      static int const shlv_insn[4] = {
> -        /* TODO: AVX512 adds support for MO_16.  */
> -        OPC_UD2, OPC_UD2, OPC_VPSLLVD, OPC_VPSLLVQ
> +        OPC_UD2, OPC_VPSLLVW, OPC_VPSLLVD, OPC_VPSLLVQ
>      };
>      static int const shrv_insn[4] = {
> -        /* TODO: AVX512 adds support for MO_16.  */
> -        OPC_UD2, OPC_UD2, OPC_VPSRLVD, OPC_VPSRLVQ
> +        OPC_UD2, OPC_VPSRLVW, OPC_VPSRLVD, OPC_VPSRLVQ
>      };
>      static int const sarv_insn[4] = {
> -        /* TODO: AVX512 adds support for MO_16, MO_64.  */
> -        OPC_UD2, OPC_UD2, OPC_VPSRAVD, OPC_UD2
> +        OPC_UD2, OPC_VPSRAVW, OPC_VPSRAVD, OPC_VPSRAVQ
>      };
>      static int const shls_insn[4] = {
>          OPC_UD2, OPC_PSLLW, OPC_PSLLD, OPC_PSLLQ
> @@ -3242,9 +3243,24 @@ int tcg_can_emit_vec_op(TCGOpcode opc, TCGType type, unsigned vece)
>  
>      case INDEX_op_shlv_vec:
>      case INDEX_op_shrv_vec:
> -        return have_avx2 && vece >= MO_32;
> +        switch (vece) {
> +        case MO_16:
> +            return have_avx512bw;
> +        case MO_32:
> +        case MO_64:
> +            return have_avx2;
> +        }
> +        return 0;
>      case INDEX_op_sarv_vec:
> -        return have_avx2 && vece == MO_32;
> +        switch (vece) {
> +        case MO_16:
> +            return have_avx512bw;
> +        case MO_32:
> +            return have_avx2;
> +        case MO_64:
> +            return have_avx512vl;
> +        }
> +        return 0;
>      case INDEX_op_rotlv_vec:
>      case INDEX_op_rotrv_vec:
>          return have_avx2 && vece >= MO_32 ? -1 : 0;

-- 
Alex Bennée