Re: [Qemu-devel] [PATCH v2 5/6] Do constant folding for shift operations.

[Richard Henderson <rth@twiddle.net>]

On 06/09/2011 03:45 AM, Kirill Batuzov wrote:
> +    case INDEX_op_shl_i32:
> +#if TCG_TARGET_REG_BITS == 64
> +        y &= 0xffffffff;
> +    case INDEX_op_shl_i64:
> +#endif
> +        return x << y;
> +
> +    case INDEX_op_shr_i32:
> +#if TCG_TARGET_REG_BITS == 64
> +        x &= 0xffffffff;
> +        y &= 0xffffffff;
> +    case INDEX_op_shr_i64:
> +#endif
> +        /* Assuming TCGArg to be unsigned */
> +        return x >> y;

Don't assume when you've got a uint64_t type readily available.

> +    case INDEX_op_sar_i32:
> +#if TCG_TARGET_REG_BITS == 64
> +        x &= 0xffffffff;
> +        y &= 0xffffffff;
> +#endif
> +        return (int32_t)x >> (int32_t)y;

Masks are redundant with the casts.

> +    case INDEX_op_rotr_i32:
> +#if TCG_TARGET_REG_BITS == 64
> +        x &= 0xffffffff;
> +        y &= 0xffffffff;
> +#endif
> +        x = (x << (32 - y)) | (x >> y);

Have you looked to see if this gets recognized as a rotate
by the compiler?  I suspect that it will if you use a cast
to uint32_t here, but not if it is left as a 64-bit TCGArg.

> +#if TCG_TARGET_REG_BITS == 64
> +    case INDEX_op_rotl_i64:
> +        x = (x << y) | (x >> (64 - y));
> +        return x;
> +#endif

Likewise it's probably best to cast to uint64_t here.


r~