Re: [PATCH v2 0/3] add strnncmp() function

["Ondřej Bílka" <neleai@seznam.cz>]

On Tue, Jun 17, 2014 at 01:08:05PM +0200, Torsten Bögershausen wrote:
> On 2014-06-17 09.34, Jeremiah Mahler wrote:
> > Add a strnncmp() function which behaves like strncmp() except it takes
> > the length of both strings instead of just one.
> > 
> > Then simplify tree-walk.c and unpack-trees.c using this new function.
> > Replace all occurrences of name_compare() with strnncmp().  Remove
> > name_compare(), which they both had identical copies of.
> > 
> > Version 2 includes suggestions from Jonathan Neider [1]:
> > 
> >   - Fix the logic which caused the new strnncmp() to behave differently
> > 	from the old version.  Now it is identical to strncmp().
> > 
> >   - Improve description of strnncmp().
> > 
> > Also, strnncmp() was switched from using memcmp() to strncmp()
> > internally to make it clear that this is meant for strings, not
> > general buffers.
> I don't think this is a good change, for 2 reasons:
> - It changes the semantics of existing code, which should be carefully
>   reviewed, documented and may be put into a seperate commit.
> - Looking into the code for memcmp() and strncmp() in libc,
>   I can see that memcmp() is written in 13 lines of assembler,
>   (on a 386 system) with a fast
>     repz cmpsb %es:(%edi),%ds:(%esi)
>   working as the core engine.
>   
>   strncmp() uses 83 lines of assembler, because after each comparison
>   the code needs to check of the '\0' in both strings.
> - I can't see a reason to replace efficient code with less efficient code,
>   so moving the old function "as is" into a include file, and declare
>   it "static inline" could be the first step.
> 
That is not true, a rep cmpsb was fast for 486 but is relatively slow
for newer processors. For performance a correct answer is to measure it than do 
blind guess. Are these strings null terminated or is giving a size just
a hint? If it is a hint then a plain strcmp could be faster (this
depends on implementation). A reason is that for implementations that
check more bytes at once it is easier to combine a terminating null mask with 
difference than trying to first find which of first 16 bytes are different and 
then compare if it is within size.