Re: [Qemu-ppc] [v2 08/13] Add VSX Vector Move Instructions

[Paolo Bonzini]

Il 11/10/2013 14:03, Tom Musta ha scritto:
> This patch adds the vector move instructions:
> 
>   - xvabsdp - Vector Absolute Value Double-Precision
>   - xvnabsdp - Vector Negative Absolute Value Double-Precision
>   - xvnegdp - Vector Negate Double-Precision
>   - xvcpsgndp - Vector Copy Sign Double-Precision
>   - xvabssp - Vector Absolute Value Single-Precision
>   - xvnabssp - Vector Negative Absolute Value Single-Precision
>   - xvnegsp - Vector Negate Single-Precision
>   - xvcpsgnsp - Vector Copy Sign Single-Precision
> 
> Signed-off-by: Tom Musta <address@hidden>
> ---
>  target-ppc/translate.c |   68
> ++++++++++++++++++++++++++++++++++++++++++++++++
>  1 files changed, 68 insertions(+), 0 deletions(-)
> 
> diff --git a/target-ppc/translate.c b/target-ppc/translate.c
> index 36e04b0..03a352d 100644
> --- a/target-ppc/translate.c
> +++ b/target-ppc/translate.c
> @@ -7216,6 +7216,66 @@ VSX_SCALAR_MOVE(xsnabsdp, OP_NABS, SGN_MASK_DP)
>  VSX_SCALAR_MOVE(xsnegdp, OP_NEG, SGN_MASK_DP)
>  VSX_SCALAR_MOVE(xscpsgndp, OP_CPSGN, SGN_MASK_DP)
> 
> +#define VSX_VECTOR_MOVE(name, op, sgn_mask)                      \
> +static void glue(gen_, name)(DisasContext * ctx)                 \
> +    {                                                            \
> +        TCGv_i64 xbh, xbl;                                       \
> +        if (unlikely(!ctx->vsx_enabled)) {                       \
> +            gen_exception(ctx, POWERPC_EXCP_VSXU);               \
> +            return;                                              \
> +        }                                                        \
> +        xbh = tcg_temp_new();                                    \
> +        xbl = tcg_temp_new();                                    \
> +        tcg_gen_mov_i64(xbh, cpu_vsrh(xB(ctx->opcode)));         \
> +        tcg_gen_mov_i64(xbl, cpu_vsrl(xB(ctx->opcode)));         \
> +        switch (op) {                                            \
> +            case OP_ABS: {                                       \
> +                tcg_gen_andi_i64(xbh, xbh, ~(sgn_mask));         \
> +                tcg_gen_andi_i64(xbl, xbl, ~(sgn_mask));         \
> +                break;                                           \
> +            }                                                    \
> +            case OP_NABS: {                                      \
> +                tcg_gen_ori_i64(xbh, xbh, (sgn_mask));           \
> +                tcg_gen_ori_i64(xbl, xbl, (sgn_mask));           \
> +                break;                                           \
> +            }                                                    \
> +            case OP_NEG: {                                       \
> +                tcg_gen_xori_i64(xbh, xbh, (sgn_mask));          \
> +                tcg_gen_xori_i64(xbl, xbl, (sgn_mask));          \
> +                break;                                           \
> +            }                                                    \
> +            case OP_CPSGN: {                                     \
> +                TCGv_i64 xah = tcg_temp_new();                   \
> +                TCGv_i64 xal = tcg_temp_new();                   \
> +                tcg_gen_mov_i64(xah, cpu_vsrh(xA(ctx->opcode))); \
> +                tcg_gen_mov_i64(xal, cpu_vsrl(xA(ctx->opcode))); \
> +                tcg_gen_andi_i64(xah, xah, (sgn_mask));          \
> +                tcg_gen_andi_i64(xal, xal, (sgn_mask));          \
> +                tcg_gen_andi_i64(xbh, xbh, ~(sgn_mask));         \
> +                tcg_gen_andi_i64(xbl, xbl, ~(sgn_mask));         \
> +                tcg_gen_or_i64(xbh, xbh, xah);                   \
> +                tcg_gen_or_i64(xbl, xbl, xal);                   \
> +                tcg_temp_free(xah);                              \
> +                tcg_temp_free(xal);                              \
> +                break;                                           \

Same as before.  Also, you may want to use a temporary for the other
cases as well, so that the constant is reused.  Using andc for OP_ABS
also makes sense, since on some RISC machines 0x80000... is cheaper than
0x7FFFF...  On target that lack andc, you'll just get the same code
you're generating now.

Paolo

> +            }                                                    \
> +        }                                                        \
> +        tcg_gen_mov_i64(cpu_vsrh(xT(ctx->opcode)), xbh);         \
> +        tcg_gen_mov_i64(cpu_vsrl(xT(ctx->opcode)), xbl);         \
> +        tcg_temp_free(xbh);                                      \
> +        tcg_temp_free(xbl);                                      \
> +    }
> +
> +VSX_VECTOR_MOVE(xvabsdp, OP_ABS, SGN_MASK_DP)
> +VSX_VECTOR_MOVE(xvnabsdp, OP_NABS, SGN_MASK_DP)
> +VSX_VECTOR_MOVE(xvnegdp, OP_NEG, SGN_MASK_DP)
> +VSX_VECTOR_MOVE(xvcpsgndp, OP_CPSGN, SGN_MASK_DP)
> +VSX_VECTOR_MOVE(xvabssp, OP_ABS, SGN_MASK_SP)
> +VSX_VECTOR_MOVE(xvnabssp, OP_NABS, SGN_MASK_SP)
> +VSX_VECTOR_MOVE(xvnegsp, OP_NEG, SGN_MASK_SP)
> +VSX_VECTOR_MOVE(xvcpsgnsp, OP_CPSGN, SGN_MASK_SP)
> +
> +
> 
>  /***                           SPE
> extension                               ***/
>  /* Register moves */
> @@ -9711,6 +9771,14 @@ GEN_XX2FORM(xsnabsdp, 0x12, 0x16, PPC2_VSX),
>  GEN_XX2FORM(xsnegdp, 0x12, 0x17, PPC2_VSX),
>  GEN_XX3FORM(xscpsgndp, 0x00, 0x16, PPC2_VSX),
> 
> +GEN_XX2FORM(xvabsdp, 0x12, 0x1D, PPC2_VSX),
> +GEN_XX2FORM(xvnabsdp, 0x12, 0x1E, PPC2_VSX),
> +GEN_XX2FORM(xvnegdp, 0x12, 0x1F, PPC2_VSX),
> +GEN_XX3FORM(xvcpsgndp, 0x00, 0x1E, PPC2_VSX),
> +GEN_XX2FORM(xvabssp, 0x12, 0x19, PPC2_VSX),
> +GEN_XX2FORM(xvnabssp, 0x12, 0x1A, PPC2_VSX),
> +GEN_XX2FORM(xvnegsp, 0x12, 0x1B, PPC2_VSX),
> +GEN_XX3FORM(xvcpsgnsp, 0x00, 0x1A, PPC2_VSX),
>  GEN_XX3FORM_DM(xxpermdi, 0x08, 0x01),
> 
>  #undef GEN_SPE