I spent the past few days rewriting some Lua code [1] into C. While I find LPEG (Lua Parsing Expression Grammar) [2] to be convenient, it is not necessarily fast [3]. Normally this isn't an issue but in this case, I was calling LPEG (Lua Parsing Expression Grammar) for each character in a blog post.
Fortunately, it was fairly straight forward porting the code to C. The code goes through text a [DELETED-character-DELETED] codepoint at a time. If it's a whitespace character or a hyphen, I mark the current position as a possible breakpoint for the text; otherwise I ignore combining characters (they don't count towards the line length). Then, when I reach past the number of characters I want for a line, I copy out the string from the beginning of the “line” to the marked breakpoint (and if there isn't a breakpoint, there is no good place to break the line so I will break the line at the line length—not much else to do), then mark the beginning of the next line and continue until the end of the text.
The hardest part was figuring out how to classify each character I needed. In the end, I pull out each Unicode codepoint from UTF-8 (Unicode Transformation Format—8-bit) [4] and look through an array to classify the codepoint as whitespace, a hyphen or a combining character; if they aren't in the table, it just a normal character.
As a sanity check, I reran the original profiling test [5]:
Table: Lines of Lua code executed to serve a request gopher (original) 457035 gopher (new) 18246 gemini (just because) 22661
Much better. And most of the 457,035 lines of code being executed are now hidden behind C. Now to make sure the code is actually faster, I profiled the new wrapt() function:
local wraptx = wrapt
local function wrapt(...)
local start = rdtsc() [6]
local res = wraptx(...)
local stop = rdtsc()
syslog('notice',"wrapt()=%d",stop-start)
return res
end
with the decently sized request [7] I used before (each line is a call to wrapt()):
Table: Runtime (lower is better) #Lua code C code 43330 11810 43440 12000 45300 12220 48100 12020 48680 13690 49260 12650 54140 12270 54650 12460 58530 12130 59760 14180 61100 15480 65440 14970 67920 15810 68750 15310 69920 17170 69960 17780 70740 16510 75640 16750 78870 19170 83200 18190 87090 17290 89070 23360 91440 19560 101800 21520 102460 21060 103790 22180 106000 22400 106010 21870 112960 21160 115300 21870 115980 23130 118690 24980 122550 23960 122710 24550 127610 23830 129580 24670 130120 24930 140580 26570 141930 25210 157640 27050 168000 32250
Excellent! The new code is three to five times faster. Now to just sit back and see how the new code fares over the next few days.
[2] https://www.inf.puc-rio.br/~roberto/lpeg/