These tricks based on tests of parser generated by pointlander/peg and pigeon with same grammer on my computer (12600kf, win11). baseline is pointlander/peg(~0.3s). First version of pigeon with same grammar took 4.5s.
Never use state, extermely slow. (-2.2s for not use state, -0.9s for disable state by option).
Remove vals := make([]any, 0, len(seq.exprs)) from parseSeqExpr or just collect !nil values (-0.55s)
Don't enable -optimize-basic-latin, it seems not improved. But basicLatinChars [128]bool is always a part of charClassMatcher, remove it. (-0.06s)
Map access of ruleRefExpr takes ~0.2s, replace it by index take ~0.05s. (-0.15s)
Reduce the use of label syntax, especially in situations where your label matched but then the sentence match fails, leading to another path to match(backtracking). A few label without backtracking is fast.
Remove pushV and popV and parser.vstack, use p.vstack = make(map[string]any) to instead. (-0.06s, it seems took long time(~50%) in cpu profiler, but i did't got so much performance improved by remove it.)
parser.rstack takes ~0.02s. (i didn't remove it)
Finally, we can got a parser cost 2x times by a parser generated by pointlander/peg. My current version take 0.5s to pass test, it's a huge boost compare to original version(4.5s), but still much slower than pointlander/peg.
It's a ok trade off for me, because code generated by pigeon is more simple and clear. And easy to port builder to other languages.
If anyone know how to make it faster, please tell me.
On linux(github action, i did't test on a real linux env), pointlander/peg is much faster. Even the optimized verison of pigeon, cost 4x when cold start, 10x when warn start.
These tricks based on tests of parser generated by pointlander/peg and pigeon with same grammer on my computer (12600kf, win11). baseline is pointlander/peg(~0.3s). First version of pigeon with same grammar took 4.5s.
vals := make([]any, 0, len(seq.exprs))from parseSeqExpr or just collect !nil values (-0.55s)-optimize-basic-latin, it seems not improved. ButbasicLatinChars [128]boolis always a part of charClassMatcher, remove it. (-0.06s)ruleRefExprtakes ~0.2s, replace it by index take ~0.05s. (-0.15s)pushVandpopVand parser.vstack, usep.vstack = make(map[string]any)to instead. (-0.06s, it seems took long time(~50%) in cpu profiler, but i did't got so much performance improved by remove it.)parser.rstacktakes ~0.02s. (i didn't remove it)Finally, we can got a parser cost 2x times by a parser generated by pointlander/peg. My current version take 0.5s to pass test, it's a huge boost compare to original version(4.5s), but still much slower than pointlander/peg. It's a ok trade off for me, because code generated by pigeon is more simple and clear. And easy to port builder to other languages.
If anyone know how to make it faster, please tell me.