Re: [Qemu-devel] [PATCH 2/6] target-ppc: add vextu[bhw]lx instructions

[Richard Henderson]

On 09/27/2016 10:45 PM, Rajalakshmi Srinivasaraghavan wrote:
> +#if defined(HOST_WORDS_BIGENDIAN)
> +#define VEXTULX_DO(name, elem)                                  \
> +target_ulong glue(helper_, name)(target_ulong a, ppc_avr_t *b)  \
> +{                                                               \
> +    target_ulong r = 0;                                         \
> +    int i;                                                      \
> +    int index = a & 0xf;                                        \
> +    for (i = 0; i < elem; i++) {                                \
> +        r = r << 8;                                             \
> +        if (index + i <= 15) {                                  \
> +            r = r | b->u8[index + i];                           \
> +        }                                                       \
> +    }                                                           \
> +    return r;                                                   \
> +}
> +#else
> +#define VEXTULX_DO(name, elem)                                  \
> +target_ulong glue(helper_, name)(target_ulong a, ppc_avr_t *b)  \
> +{                                                               \
> +    target_ulong r = 0;                                         \
> +    int i;                                                      \
> +    int index = 15 - (a & 0xf);                                 \
> +    for (i = 0; i < elem; i++) {                                \
> +        r = r << 8;                                             \
> +        if (index - i >= 0) {                                   \
> +            r = r | b->u8[index - i];                           \
> +        }                                                       \
> +    }                                                           \
> +    return r;                                                   \
> +}
> +#endif
> +
> +VEXTULX_DO(vextublx, 1)
> +VEXTULX_DO(vextuhlx, 2)
> +VEXTULX_DO(vextuwlx, 4)
> +#undef VEXTULX_DO

Ew.

This should be one 128-bit shift and one and.

Since the shift amount is a multiple of 8, the 128-bit shift for vextub[lr]x
does not need to cross a double-word boundary, and so can be decomposed into
one 64-bit shift of (count & 64 ? hi : lo).

For vextu[hw]lr]x, you'd need to do the whole left-shift, right-shift, or thing.

But still, fantastically better than a loop.


r~