Re: [PATCH v2 2/4] sched/fair: add util_est on top of PELT

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Dec 20, 2017 at 09:57:47AM +0100, Peter Zijlstra wrote:
> On Fri, Dec 15, 2017 at 03:41:40PM +0000, Patrick Bellasi wrote:
> > Close enough, the actual code is:
> > 
> >         util_est = p->util_est.ewma;
> >     5218:       f9403ba3        ldr     x3, [x29,#112]
> >     521c:       f9418462        ldr     x2, [x3,#776]
> >         if (abs(util_est - util_last) <= (SCHED_CAPACITY_SCALE / 100))
> >     5220:       eb010040        subs    x0, x2, x1
> >     5224:       da805400        cneg    x0, x0, mi
> >     5228:       f100281f        cmp     x0, #0xa
> >     522c:       54fff9cd        b.le    5164 <dequeue_task_fair+0xa04>
> 
> Ah, that cneg instruction is cute; on x86 we end up with something like:
> 
> bool abs_test(long s)
> {
>         return abs(s) < 32;
> }
> 
>         cmpl    $-31, %eax
>         jl      .L107
>         movq    -8(%rbp), %rax
>         cmpl    $31, %eax
>         jg      .L107
>         movl    $1, %eax
>         jmp     .L108
> .L107:
>         movl    $0, %eax
> .L108:
> 
> 
> But I figured you can actually do:
> 
> 	abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
> 
> Which, if y is a constant, should result in nicer code, and it does for
> x86:
> 
>         addq    $31, %rax
>         cmpq    $62, %rax
>         setbe   %al
>         movzbl  %al, %eax
> 
> Just not measurably faster, I suppose because of all the dependencies :/

Ah no, it actually is, I'm an idiot and used 'long' for return value. If
I use bool we loose that last movzbl and we go from around 4.0 cycles
down to 3.4 cycles.