On 01.03.19 17:26, Richard Henderson wrote:
> On 3/1/19 3:53 AM, David Hildenbrand wrote:
>> + /*
>> + * Check for possible access exceptions by trying to load the last
>> + * element. The first element will be checked first next.
>> + */
>> + t = tcg_temp_new_i64();
>> + gen_addi_and_wrap_i64(s, t, o->addr1, (v3 - v1) * 16 + 8);
>> + tcg_gen_qemu_ld_i64(t, t, get_mem_index(s), MO_TEQ);
>
> qemu_ld expands to enough code that it is a shame to discard this value and
> reload it during this loop. Perhaps load this to t2...
>
>> +
>> + for (;; v1++) {
>> + tcg_gen_qemu_ld_i64(t, o->addr1, get_mem_index(s), MO_TEQ);
>> + write_vec_element_i64(t, v1, 0, ES_64);
>
> Move v1 == v3 break here...
>
>> + gen_addi_and_wrap_i64(s, o->addr1, o->addr1, 8);
>> + tcg_gen_qemu_ld_i64(t, o->addr1, get_mem_index(s), MO_TEQ);
>> + write_vec_element_i64(t, v1, 1, ES_64);
>> + if (v1 == v3) {
>> + break;
>> + }
>> + gen_addi_and_wrap_i64(s, o->addr1, o->addr1, 8);
>> + }
>
> ... and store t2 into v3 element 1 after the loop.
Yes, makes perfect sense, thanks!
--
Thanks,
David / dhildenb
