Re: [Qemu-devel] [PATCH v2 05/11] target-arm: Implement ccmp branchless

[Peter Maydell]

On 8 September 2015 at 06:18, Richard Henderson <address@hidden> wrote:
> On 09/07/2015 10:31 AM, Peter Maydell wrote:
>>>
>>> -    if (cond < 0x0e) { /* continue */
>>> -        gen_set_label(label_continue);
>>> +    /* If COND was false, force the flags to #nzcv.
>>> +       Note that T1 = (COND ? 0 : -1), T2 = (COND ? -1 : 0).  */
>>> +    tcg_t1 = tcg_temp_new_i32();
>>> +    tcg_t2 = tcg_temp_new_i32();
>>> +    tcg_gen_neg_i32(tcg_t1, tcg_t0);
>>> +    tcg_gen_subi_i32(tcg_t2, tcg_t0, 1);
>>
>>
>> t2 is ~t1, right? Do we get better/worse code if we use
>> tcg_gen_andc_i32(..., tcg_t1) rather than creating t2 and
>> using gen_and_i32 ?
>>
>>> +
>>> +    if (nzcv & 8) { /* N */
>>> +        tcg_gen_or_i32(cpu_NF, cpu_NF, tcg_t1);
>>> +    } else {
>>> +        tcg_gen_and_i32(cpu_NF, cpu_NF, tcg_t2);
>>> +    }
>>> +    if (nzcv & 4) { /* Z */
>>> +        tcg_gen_and_i32(cpu_ZF, cpu_ZF, tcg_t2);
>>> +    } else {
>>> +        tcg_gen_or_i32(cpu_ZF, cpu_ZF, tcg_t0);
>>> +    }
>>> +    if (nzcv & 2) { /* C */
>>> +        tcg_gen_or_i32(cpu_CF, cpu_CF, tcg_t0);
>>> +    } else {
>>> +        tcg_gen_and_i32(cpu_CF, cpu_CF, tcg_t2);
>>> +    }
>>> +    if (nzcv & 1) { /* V */
>>> +        tcg_gen_or_i32(cpu_VF, cpu_VF, tcg_t1);
>>> +    } else {
>>> +        tcg_gen_and_i32(cpu_VF, cpu_VF, tcg_t2);
>
>
> If the host supports andc, it's probably better to use only the one temp.
> But otherwise we may save 4 not insns.

The tcg common code isn't smart enough to notice it only
needs to calculate not(t1) once ?

In the overwhelmingly common case (x86 tcg backend)
we would save an insn every time, right?

>  Is it worth complicating the code
> for that?

I wouldn't bother to make the front-end generate different
code for the backend does/doesn't have andc situations,
certainly.

Anyway, I'm just guessing here, you probably have a better
feel than me for what codegen choices work better, so I'll
leave the choice up to you.

thanks
-- PMM