On 4 December 2013 21:59, Richard Henderson <r...@twiddle.net> wrote:
> On 12/05/2013 08:33 AM, Peter Maydell wrote:
>> +static TCGv_i64 read_cpu_reg(DisasContext *s, int reg, int sf)
>> +{
>> + TCGv_i64 v = new_tmp_a64_zero(s);
>> + if (reg != 31) {
>
> This is going to issue a garbage "movi tmp, 0" opcode for every non-XZR read.
> While the optimizer will remove that, it's only going to slow things down, and
> easily avoidable.
>
> Better to allocate the temporary and only perform the movi for reg == 31.
Yeah, I guess we should have a new_tmp_a64() that gives you an
uninitialized auto-freeing tmp, and then have new_tmp_a64_zero()
do { v = new_tmp_a64(); tcg_gen_movi_i64(v, 0); return v; }
(or not have the _zero variant at all, depending on whether it's
particularly useful.)
thanks
-- PMM
