Re: [ANN] Lua 5.2.1 (work1) now available

[KHMan <keinhong@...>]

On 3/26/2012 1:09 AM, Roberto Ierusalimschy wrote:
> > Intern on first compare is not easy to implement, there are a lot of
> > complications. A quick-and-dirty attempt at using long string hash
> > values to avoid extra memcmp() work didn't really work (a few %
> > slower than default lua-5.2.1wk1) because in the examples I am
> > using, there are a lot of equal compares (76%) which means memcmp()
> > is still mostly needed. Short string compares is still much faster.
>
> A possible small optimization is to check whether "a == b" in
> 'luaS_eqlngstr', before doing the memcmp.

No improvement for this particular script. Since the script loads 
from 2 separate data files, I think the longstr compares are 
always a!=b. OTOH, there seems to be little penalty for the extra 
test.

same files (timing in sec, lower of 2 runs)
=========================================
dataset ->      md5     sha1    sha256
 hex str len    32      40      64
--------------------------------------
lua-5.1.5       2.621   2.646   2.709
lua-5.2.0       2.356   2.379   2.447
lua-5.2.1wk1
 shortlen=16    3.107   3.336   3.526
 shortlen=32    2.547   2.976+  3.223+
 shortlen=48    2.355   2.361   3.042
 shortlen=64    2.365   2.344   2.413
 shortlen=128   2.382   2.376   2.429
 try2*          2.551   3.006+  3.190+

* try2 has (a==b) added in luaS_eqlngstr()
  with default shortlen=32
+ times that compare effect of adding (a==b)

--
Cheers,
Kein-Hong Man (esq.)
Kuala Lumpur, Malaysia