Looking for ideas to lower binding overhead for lua-re2.

-- regexdna_re2.lua
re2 = require 'lua-re2'
function printelapse(start)
local elapse = os.clock() - start
print(elapse..'s')
end
-- omitted code
-- 'variants' is a table of regex dna patterns to match
-- 'seq' is a *really* long dna sequence string loaded from stdin
local start = os.clock()
for i = 1, 10 do
for _, p in ipairs(variants) do
io.write( string.format('%s %d\n', p, re2(p):countmatch(seq)) )
end
end
printelapse(start)


// lua-re2.cpp
// a temporary function in my binding code used in the benchmark above
static int re2_countmatch(lua_State *L)
{
lua_settop(L, 2);
RE2 *re2obj = luaRE2_checkobject(L, 1);
StringPiece subject( luaL_checkstring(L, 2) );

int c = 0;
while(RE2::FindAndConsume(&subject, *re2obj))
++c;

lua_pushinteger(L, c);
return 1;
}

# regexdna.py
from time import time
from sys import stdin, stdout
from re2 import sub, findall

def printelapse(start):
elapse = time() - start
print(str(elapse) + 's')

# omitted code
start = time()
for i in xrange(10):
for f in variants:
write(f.decode("utf-8") + ' ' + str(len(findall(f.decode("utf-8"), seq))) + '\n')
printelapse(start)

// regexdna.cpp
#include "re2/re2.h"
#include <iostream>
#include <memory>
#include <stdio.h>
#include <time.h>

int main(void)
{
// lots of omitted code
clock_t start = clock();
for(int it = 0; it < 10; ++it)
{
for (int i = 0; i < (int)(sizeof(pattern1) / sizeof(string)); ++i)
{
int count = 0;
RE2 pat(pattern1[i]);
StringPiece piece = str;

while (RE2::FindAndConsume(&piece, pat)) ++count;

printf("%s %d\n", pattern1[i].c_str(), count);
}
}
double elapse = (clock() - start)*1.0 / CLOCKS_PER_SEC;
cout << elapse << "s\n";
}



static int re2_countmatch(lua_State *L)
{
lua_settop(L, 2);
RE2 *re2obj = luaRE2_checkobject(L, 1);
static const string overhead_test = luaL_checkstring(L, 2);
StringPiece subject( overhead_test );

int c = 0;
while(RE2::FindAndConsume(&subject, *re2obj))
++c;

lua_pushinteger(L, c);
return 1;
}


• Mingw-gcc 4.6.3
• LuaJIT 2.02
• Python 2.7.3
• RE2 compiled from head using mingw above
• Windows 7 64-bit on Intel Q6600 quad-core.