Re: Profiling LPEG

[Roberto Ierusalimschy <roberto@...>]

>   Okay, one thing I did was recompile LPEG to activate the ptree() function,
> and dump both the 're' compiled version any my LPEG version.  The output was
> very surprising.
> 
> 	're' version [1]:   3,212 lines long
> 	'lpeg' version:   611,931 lines long
> 
>   It appers that 're' does a better job of generating LPEG than I did by
> hand.  I'm surprised that the 'lpeg' version was only twice as slow, given
> the results.

Without seeing your code, one thing popped in my head: In re, the only
way to name a pattern is as a rule in a grammar. In LPeg, we can
name patterns as grammar rules or as Lua variables. When you
use a pattern inside another pattern, Lua expands its code, as if it
was a macro. So, if you have patterns that appear multiple times inside
others, and then those patterns are used inside other ones, etc., things
can get very large.

Quick comparison:
-----------------------------------------------------------
local P = lpeg.P"hello world"    -- pattern in a Lua variable
(P * P * P):ptree()
-- tree has 66 lines

local P = lpeg.V"P"         -- pattern in a grammar rule
lpeg.P{P * P * P, P = lpeg.P"hello world"}:ptree()
-- tree has 32 lines
-----------------------------------------------------------

-- Roberto