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

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

Hi Richard,

On 9/24/21 23:53, Richard Henderson wrote:
> On 9/24/21 11:08 AM, WANG Xuerui wrote:
>> Oops, for some reason I only received this at about 8 pm...
>
> That was my fault.  I wrote a bunch of stuff off-line yesterday while 
> traveling, and the mail queue only flushed this morning.
>
> I'll note there's a bug in my example code wrt initializing rd with 
> addi, then overwriting with cu32i.d.
>
> I like your v4 version of movi, with the high-bit-set predicate. The 
> only case I can think of that you miss is e.g. 0x7fffffffffffffff, 
> which can be
>
>     addi.w    rd, zero, -1
>     cu52i.d    rd, rd, 0x7ff
>
> One possibility is to extract a subroutine:
>
> static void tcg_out_movi_i32(TCGContext *s, TCGReg rd, int32_t val)
> {
>     /* Single instruction cases */
>     /* else lu12i.w + ori */
> }
>
> static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd,
>                          tcg_target_long val)
> {
>     if (type == TCG_TYPE_I32 || val == (int32_t)val) {
>         tcg_out_movi_i32(s, rd, val);
>         return;
>     }
>
>     /* PC-relative cases */
>
>     if (ctz64(val) >= 52) {
>         tcg_out_opc_cu52i_d(s, rd, TCG_REG_ZERO, val >> 52);
>         return;
>     }
>
>     /* Slow path.  Initialize the low 32-bits, then concat high bits. */
>     tcg_out_movi_i32(s, rd, val);
>
>     rd_high_bits_are_ones = (int32_t)val < 0);
>
>     /* Your imm_part_needs_loading checks; rd is always written. */
> }
>
This is so impressive understanding of the LoongArch assembly, working 
around cu32i.d's single operand limitation so nicely. I'll just 
(shamelessly) take this for v5, thanks again!
>
> r~