RFC: x86-64: Use fxsave/xsave/xsavec in _dl_runtime_resolve [BZ #21265]

[Florian Weimer]

On 10/20/2017 03:21 PM, Florian Weimer wrote:
> LD_DEBUG=statistics shows this:
> 
>  Â Â Â Â Â  9506:
>  Â Â Â Â Â  9506:Â Â Â Â  runtime linker statistics:
>  Â Â Â Â Â  9506:Â Â Â Â Â Â  total startup time in dynamic loader: 19960074 cycles
>  Â Â Â Â Â  9506:Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  time needed for relocation: 19105814 cycles 
> (95.7%)
>  Â Â Â Â Â  9506:Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  number of relocations: 87
>  Â Â Â Â Â  9506:Â Â Â Â Â Â Â Â Â Â  number of relocations from cache: 3
>  Â Â Â Â Â  9506:Â Â Â Â Â Â Â Â Â Â Â Â  number of relative relocations: 1226
>  Â Â Â Â Â  9506:Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  time needed to load objects: 701382 cycles 
> (3.5%)
>  Â Â Â Â Â  9506:
>  Â Â Â Â Â  9506:Â Â Â Â  runtime linker statistics:
>  Â Â Â Â Â  9506:Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  final number of relocations: 781589
>  Â Â Â Â Â  9506:Â Â Â Â  final number of relocations from cache: 3
> 
> This is a main program which contains 1,500 function calls.  The 
> functions are defined in a single DSO, and each function calls 520 other 
> functions, giving a total number of 781,500 relocations from the test.
> 
> On my laptop (Intel(R) Core(TM) i7-4810MQ CPU @ 2.80GHz), I get this 
> (ten runs, real time measured in seconds):
> 
>  > t.test(prev_laptop, after_laptop)
> 
>  Â Â Â Â Welch Two Sample t-test
> 
> data:Â  prev_laptop and after_laptop
> t = -14.932, df = 18, p-value = 1.392e-11
> alternative hypothesis: true difference in means is not equal to 0
> 95 percent confidence interval:
>  Â -0.05749145 -0.04330855
> sample estimates:
> mean of x mean of y
>  Â Â  0.2345Â Â Â  0.2849
> 
> So it's definitely not in the noise.  The penalty appears to be around 
> 65ns per relocation.

I did some more benchmarks.  Ryzen takes a bit of a hit (100ns or ~30%). 
  Purley looks very good (35ns or ~5%).  The outlier is KNL, with 750ns 
added per relocation (at lower clock rates admittedly) and ~50% longer 
relocation times overall.  (The caveat is that this is lab hardware, 
which may or may not match production silicon.)

I still think these numbers are okay.  To put this into perspective, the 
total number of relocations that are processed when running yum, a 
non-trivial Python application, is less than 22,000.  So there is an 
expected additional startup overhead of about 16.5ms for KNL, but that 
is still in the noise for a simple yum command such as “yum repolist”.

Any other ideas what else we could benchmark?

Thanks,
Florian