Re: [PATCH v3 09/30] tcg/loongarch64: Implement tcg_out_mov and tcg_out_movi

[WANG Xuerui <i.qemu@xen0n.name>]

Hi Richard,

On 9/23/21 02:51, Richard Henderson wrote:
> On 9/22/21 11:09 AM, WANG Xuerui wrote:
>> +    if (sextreg(val, 0, 52) == val) {
>> +        /*
>> +         * Fits in 52-bits, upper bits are already properly 
>> sign-extended by
>> +         * cu32i.d.
>> +         */
>> +        return;
>> +    }
>> +    tcg_out_opc_cu52i_d(s, rd, rd, top);
>> +}
>
> Oh, future improvement: constants with 52 low zero bits can be loaded 
> with cu52i(rd, zero, val >> 52).
>
> Which means there's a set of interesting constants:
>
>   abc0_0000_0000_0def
>
>     ori    rd, zero, 0xdef
>     cu52i    rd, rd, 0xabc
>
>   abcf_ffff_ffff_fdef
>
>     cu52i    rd, zero, 0xabc - 1
>     addi.d    rd, rd, 0xdef
I think I'll try to implement this in some kind of follow-up patch, yeah.
>
> Also,
>
>> +    tcg_out_opc_lu12i_w(s, rd, upper);
>> +    if (low != 0) {
>> +        tcg_out_opc_ori(s, rd, rd, low & 0xfff);
>> +    }
>
> when upper == 0 and low != 0, we can omit the lu12i.
>
This optimization can't be blindly implemented though; because cu32i.d 
only takes one input register, in case upper == low == 0 but higher != 
0, we would have no proper input. So in that case something like "move 
rd, zero" is necessary, and logic is a little bit complicated. I'll 
include this in v4 though.
>
> r~