Re: [Caml-list] ackermann microbenchmark strange results

[Alain Frisch <alain@frisch.fr>]

+1

I've already seen 20% speedup obtained by adding some dead code (in an 
OCaml program, but this is irrelevant).

-- Alain

On 04/24/2013 05:57 PM, John Carr wrote:
> Try changing loop alignment by editing assembly code.  The address of
> ack is different in the different versions.  Modern Intel processors are
> sensitive to code alignment.  There is a limit on the number of branch
> prediction table entries per cache line.  An instruction that crosses a
> cache line boundary may be slower than an instruction within a cache
> line.  I am not surprised to see a 20% difference caused by an apparently
> irrelevant code change.
>
>> Moreover, the generated assembly code for the main loop is the same, afaics. The only
>> difference is the initialization of structure fields and the initial call to ack. Please can anybody
>> explain the performance difference? I understand that microbenchmarks are no way the basis to draw
>> performance conclusions upon, but I cannot explain these results to myself in any meaninful way.
>> Please help! :)
>