On 12/22/2011 12:42 PM, Richard Sandiford wrote:
> Richard Henderson <r...@redhat.com> writes:
>> + /* Use PUL/PLU to produce { L, H } op { H, L }.
>> + By reversing the pair order, rather a pure interleave high,
>> + we don't produce erroneous exceptional conditions. */
>
> "rather than". We don't produce erroneous exceptional conditions
> that would result from H op H? It probably should be obvious,
> but it took me a few seconds to work out...
Will fix.
>> + case V4HImode:
>> + /* Perform the first reduction with interleave,
>> + and subsequent reductions with shifts. */
>> + emit_insn (gen_loongson_punpckhwd_hi (tmp, in, in));
>> + emit_insn (gen (in, in, tmp));
>> + emit_insn (gen_vec_shr_v4hi (tmp, in, force_reg (SImode, GEN_INT
>> (16))));
>> + break;
>> +
>> + case V8QImode:
>> + emit_insn (gen_loongson_punpckhwd_qi (tmp, in, in));
>> + emit_insn (gen (in, in, tmp));
>> + emit_insn (gen_vec_shr_v8qi (tmp, in, force_reg (SImode, GEN_INT
>> (16))));
>> + emit_insn (gen (in, in, tmp));
>> + emit_insn (gen_vec_shr_v8qi (tmp, in, force_reg (SImode, GEN_INT
>> (8))));
>> + break;
>
> Seems safer to have a second temporary rather than assign to the input here.
Ouch. That's a real bug.
r~
