Re: [RFC PATCH 09/43] target/loongarch: Implement vhaddw/vhsubw

[Richard Henderson]

On 12/24/22 00:15, Song Gao wrote:
> +#define S_EVEN(a, bit) \
> +        ((((int64_t)(a)) << (64 - bit / 2)) >> (64 - bit / 2))
> +
> +#define U_EVEN(a, bit) \
> +        ((((uint64_t)(a)) << (64 - bit / 2)) >> (64 - bit / 2))
> +
> +#define S_ODD(a, bit) \
> +        ((((int64_t)(a)) << (64 - bit)) >> (64 - bit/ 2))
> +
> +#define U_ODD(a, bit) \
> +        ((((uint64_t)(a)) << (64 - bit)) >> (64 - bit / 2))
> +
> +#define S_EVEN_Q(a, bit) \
> +        ((((__int128)(a)) << (128 - bit / 2)) >> (128 - bit / 2))
> +
> +#define U_EVEN_Q(a, bit) \
> +        ((((unsigned __int128)(a)) << (128 - bit / 2)) >> (128 - bit / 2))
> +
> +#define S_ODD_Q(a, bit) \
> +        ((((__int128)(a)) << (128 - bit)) >> (128 - bit/ 2))
> +
> +#define U_ODD_Q(a, bit) \
> +        ((((unsigned __int128)(a)) << (128 - bit)) >> (128 - bit / 2))

I suspect all of these are wrong.  I believe bit is in [0-127], which means both (64 - 
bit) and (128 - bit) generate out-of range shifts.

Also, you can't use __int128 directly.

I'm somewhat surprised that you're shifting at all, rather than indexing the correct 
element from of the vec_t arrays.


r~