Re: [PATCH 1/1] string: Add stpecpy(3)

[Alejandro Colomar <alx.manpages@gmail.com>]

[-- Attachment #1.1: Type: text/plain, Size: 1803 bytes --]



On 12/27/22 00:24, Alejandro Colomar wrote:
> 
> 
> On 12/26/22 23:25, Alejandro Colomar wrote:
>> Hi Noah,
>>
>> On 12/26/22 03:43, Noah Goldstein wrote:
>>
>>> I see. The code you commented earlier was NULL for both.
>>
>> I don't remember; maybe there was a typo...
>>
>>>
>>> Either way you can just make it:
>>>
>>> ```
>>> if((dst - 1UL) >= (end - 1UL)) {
>>
>> I didn't expect that would be significantly better than `(dst == end || dst == 
>> NULL)`.  However, I compiled both just to see, and the assembly output for 
>> your code is shorter.  I'll benchmark both to see if there are performance 
>> differences.
>>
>> I wonder why the compiler doesn't generate this code if it's better; it has 
>> all the information to decide that it can be transformed into that.
>>
>> Both clang and GCC produce better code with your suggestion (although in the 
>> case of GCC, the difference is bigger.
> 
> Self-correction:
> 
> They produce smaller code for your suggestion, but it seems to be slower code.

However, and this is interesting, calling strnlen(3) results in faster code even 
when no truncation occurs; at least for the short string I tested.

I suspect it might be because it is already heavily optimized in glibc, and it 
implicitly simplifies surrounding code:

-       slen = strlen(src);
         dsize = end - dst;
-       trunc = (slen >= dsize);
+       slen = strnlen(src, dsize);
+       trunc = (slen == dsize);

The generated assembly code is one line smaller (on my system, and phase of the 
moon), and some small percent faster.  :)

Cheers,

Alex


> 
>>
>> Cheers,
>>
>> Alex
>>
>>>   return dst; // either dst == NULL or dst == end.
>>> }
>>> ```
>>
>>
> 

-- 
<http://www.alejandro-colomar.es/>

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]