Re: [Emacs-diffs] master 605f901: Prefer memcpy and memset to doing it by hand

Re: [Emacs-diffs] master 605f901: Prefer memcpy and memset to doing it by hand

[Stefan Monnier]

> +  if (XLI (init) == 0)
> +    memset (p->contents, 0, XFASTINT (length) * sizeof p->contents[0]);
> +  else
> +    for (ptrdiff_t i = 0; i < XFASTINT (length); i++)
> +      p->contents[i] = init;

FWIW, I much prefer keeping just the loop, over using this "if+memset"
which seems like an obvious case or too-early-optimization.


        Stefan

Re: [Emacs-diffs] master 605f901: Prefer memcpy and memset to doing it by hand

[Paul Eggert]

[-- Attachment #1: Type: text/plain, Size: 525 bytes --]

On 02/05/2016 05:25 AM, Stefan Monnier wrote:
> FWIW, I much prefer keeping just the loop, over using this "if+memset"
> which seems like an obvious case or too-early-optimization.
Fair enough, I didn't measure it and you're probably right that it's not 
worth the complexity. I changed it to use just the loop, with the 
attached patch.

This patch's commit message documents what I had *originally* put in 
there in my private copy, something that I fondly hope would have made 
the hair on the back of your neck stand up.

[-- Attachment #2: 0001-Omit-XLI-init-0-optimization-in-make-vector.patch --]
[-- Type: application/x-patch, Size: 1293 bytes --]

Re: [Emacs-diffs] master 605f901: Prefer memcpy and memset to doing it by hand

[John Wiegley]

>>>>> Stefan Monnier <monnier@iro.umontreal.ca> writes:

>> +  if (XLI (init) == 0)
>> +    memset (p->contents, 0, XFASTINT (length) * sizeof p->contents[0]);
>> +  else
>> +    for (ptrdiff_t i = 0; i < XFASTINT (length); i++)
>> +      p->contents[i] = init;

> FWIW, I much prefer keeping just the loop, over using this "if+memset" which
> seems like an obvious case or too-early-optimization.

Is this so performance critical that we can't just always use the memset?
Don't modern compilers inline the memset into the loop these days?

-- 
John Wiegley                  GPG fingerprint = 4710 CF98 AF9B 327B B80F
http://newartisans.com                          60E1 46C4 BD1A 7AC1 4BA2

Re: [Emacs-diffs] master 605f901: Prefer memcpy and memset to doing it by hand

[Paul Eggert]

John Wiegley wrote:
> Is this so performance critical that we can't just always use the memset?

We can't always use the memset. The memset works only if the bit pattern we're 
trying to store can be implemented by storing the same unsigned char value 
repeatedly. The important special case here is storing Qnil, which is 
represented by the all-zeros bit pattern.