From: Linus Torvalds <torvalds@linux-foundation.org>
Date: Sun, 18 Mar 2007 01:29:38 GMT
Subject: Re: [PATCH 2/2] Implement a simple delta_base cache
Message-ID: <Pine.LNX.4.64.0703171822280.4964@woody.linux-foundation.org>
In-Reply-To: <118833cc0703171814n4e56ab9fwfaaea81c903ae235@mail.gmail.com>



On Sat, 17 Mar 2007, Morten Welinder wrote:
>
> >         samples  %        app name                 symbol name
> >         41527    15.6550  git                      strlen
> 
> Almost 16% in strlen?  Ugh!
> 
> That's a lot of strings, or perhaps very long strings.  Or a profiling bug.

It's likely real, and the problem is likely lots of small strings.

Each git tree entry is:

	"<octal mode> name\0" <20-byte sha1>

so you do have a *lot* of strlen() calls when doing any tree parsing. And 
for some inexplicable reason, glibc thinks strings are long on average, so 
it has a fancy algorithm to do 8 bytes at a time and tries to do things 
aligned etc.

The size of strlen() on x86-64 with glibc is 232 bytes. I'm not kidding.

			Linus