Re: [PATCH 00/10] jump label: introduce very_[un]likely + cleanups + docs

[Ingo Molnar <mingo@elte.hu>]

* H. Peter Anvin <hpa@zytor.com> wrote:

> On 02/21/2012 11:25 PM, Ingo Molnar wrote:
> > 
> > There is a fundamental assymetry, and intentionally so. You 
> > *really* have to think what the common case is, and make 
> > sure the build defaults to that. It's not the end of the 
> > world to have it flipped over, but there's costs and those 
> > costs are higher even in the branch path than a regular 
> > likely()/unlikely().
> 
> No, not really -- it's still an unconditional branch, which 
> means you will not tax the branch predictor in any way and 
> which can be followed by the front end without taking a 
> speculation hit. [...]

You are talking about CPU level costs, I am also talking about 
costs introduced at build time.

Fact is, jump-label unlikely branches are moved *out of line*: 
they are often in unlikely portions of the function (near other 
unlikely branches), with instruction cache granularity costs and 
potentially higher instruction-cache miss costs attached, etc.

You are missing three important aspects:

Firstly, instead of:

  ins1
  ins2
  ins3
  ins4
  ins5
  ins-compare
  ins-branch
  ins6
  ins7
  ins8
  ins9
  ins10

We have:

  ins1
  ins2
  ins3
  ins4
  ins5
  ins-jump

  [ hole ]

  ins6
  ins7
  ins8
  ins9
  ins10
  ins-jump back

Where the 'hole' fragments the instruction cache layout. Given 
that most of kernel execution is instruction-cache-cold, the 
'straightness' of kernel code matters quite a bit.

Secondly, there's build time instruction scheduling costs as 
well: GCC will prefer the likely branch over the unlikely one, 
so we might see extra instructions in the out-of-line code:


  ins1
  ins2
  ins3
  ins4
  ins5
  ins-jump

  [ hole ]

  ins-extra-1
  ins-extra-2
  ins6
  ins7
  ins8
  ins9
  ins10
  ins-jump back

In that sense jump labels are unlikely() branches combined with 
a patching mechanism.

Thus *both* aspects are important: if a branch is *truly* 50/50 
then it's quite possibly *NOT* a correct optimization to use 
jump-labels as the 'uncommon' code goes through extra hoops and 
fragments out of the fastpath, which in quite many real life 
cases can outstrip the advantage of the avoidance of a single 
branch ...

Thirdly,

even if it's a correct optimization and both branches happen to 
outperform the pre-jump-label version, regardless of the 
direction of the jump label flag, it's *STILL* fundamentally 
assymetric: due to the hole and due to the possible extra 
instructions the out of line code will be slower by a few 
instruction and the NOP fall-through will be faster.

This is fundamentally so, and any naming that tries to *hide* 
that assymetry and the associated micro-costs is confused.

Thanks,

	Ingo