Re: [PATCH v2 04/32] target/mips: Use dup_const() to simplify

[Richard Henderson]

On 10/27/21 11:07 AM, Philippe Mathieu-Daudé wrote:
> +    uint64_t eval_big = dup_const(df, 0x80);
>       TCGv_i64 t0 = tcg_temp_new_i64();
>       TCGv_i64 t1 = tcg_temp_new_i64();
> -    switch (df) {
> -    case DF_BYTE:
> -        eval_zero_or_big = 0x0101010101010101ULL;
> -        eval_big = 0x8080808080808080ULL;
> -        break;
> -    case DF_HALF:
> -        eval_zero_or_big = 0x0001000100010001ULL;
> -        eval_big = 0x8000800080008000ULL;
> -        break;
> -    case DF_WORD:
> -        eval_zero_or_big = 0x0000000100000001ULL;
> -        eval_big = 0x8000000080000000ULL;
> -        break;
> -    case DF_DOUBLE:
> -        eval_zero_or_big = 0x0000000000000001ULL;
> -        eval_big = 0x8000000000000000ULL;

The conversion is incorrect for eval_big.
The conversion creates e.g.

    0x0080 0080 0080 0080
not
    0x8000 8000 8000 8000

You'd have to do something like

    uint64_t eval_one = dup_const(df, 1);
    uint64_t eval_big = eval_one << ((8 << df) - 1);

r~