From: Ingo Molnar <mingo@elte.hu>
Date: Thu, 28 Apr 2011 20:32:05 GMT
Subject: Re: [PATCH] git gc: Speed it up by 18% via faster hash comparisons
Message-ID: <20110428203205.GD24755@elte.hu>
In-Reply-To: <BANLkTim90XOLRBPtVCXFb1ptkmUvHGRqeg@mail.gmail.com>


* Erik Faye-Lund <kusmabite@gmail.com> wrote:

> Thanks. I also timed on my end (on Windows), and I came to the same 
> conclusion (but the improvements of your original was somewhat smaller in my 
> end; could be due to the test-case). It seems like the early-out wasn't the 
> only reason your original patch performed faster. It could be that memcmp 
> (probably) didn't get inlined, and the extra function call outweighs the 
> complexity. [...]

Function calls arent that heavy really. My measurements identified the 
following effects:

 - profiling of stalled cycles clearly pinpointed the REP MOV string 
   instruction.

 - the patched code had less branch-misses - the clearer and inlined open-coded 
   loop is probably easier for the CPU to speculate along - while REP MOV 
   string ops are 'opaque' and the result might be harder to speculate.

So i think the main benefit of my patch is that it avoids the REP MOV 
instruction.

Thanks,

	Ingo