perf(lexer): a word position asks about zsh groups without measuring the file

[LESdylan]

What & why

Issue #135: sourcing a 3.6k-line rc cost 181 ms, against 52 ms in bash, and all of it was CPU time.

Under callgrind, 54% of every instruction retired while sourcing the hellishrc_plugins framework (3.5k lines) was strlen, and all of it came from one line:

int	xg_alt_group(const char *at)
{
	return (xg_alt_group_n(at, at + ft_strlen(at), false));
}

At every byte of every word, the lexer asks whether a zsh alternation group (a|b) starts there. at points into the whole input, so each question first measured the rest of the file. That made lexing quadratic in the file's size, in either dialect.

Change (src/execution/case_match_ext3.c): xg_alt_group_n answers 0 unless *at is ( and the zsh dialect is on. Those two tests now come first, and the end of the input is measured only when both pass. The answers are the same by construction.

Instructions retired to source that framework: 457 M → 208 M. Wall time, OPT=1 build, minimum of 7 runs:

	bash 5.3.9	develop	this PR
-n, 6000 small functions	32 ms	631 ms	28 ms
source the framework	24 ms	100 ms	74 ms
source 6000 trivial lines	16 ms	54 ms	24 ms

Refs #135. It does not close it. What remains is linear, and elsewhere:

• a chunk is re-lexed each time an open construct grows past a heredoc line;
• the heredoc pre-scan lexes line by line;
• the lexer does lookaheads on every byte.

How I verified it

• A scaling test that fails on develop. tests/parse_scaling_test.py gains a word-heavy case: 2000 lines against 16000, through both -n and source, with no fast pass.
  • On develop's ASan build the ratios are 39.6 and 22.0, which is quadratic.
  • On this branch they are 7.0 and 8.0: linear, and under the new 20.0 bound.
  • The existing cases pass anything under 1.5 s without comparing, and this quadratic stayed under that at their sizes, which is why they never caught it.
• Local gates on this commit (ASan debug build):
  • golden tests/tester: 5363/5363;
  • tests/run_scripts.sh against bash --posix: 125/125;
  • verify_alloc.sh: identical output on both heaps;
  • alloc_stress.sh: all clean;
  • tests/pty_suite.sh: 111 ok, 6 skipped, 4 failed. None of the four failures is this change:
    • prompt_compat_matrix, prompt_drift_matrix and prompt_jobs_badge expect the non-root %/$ prompt and get #, because the container runs as root. They fail the same way on develop, and CI runs them as a normal user.
    • hxp_framework_test hit the 420 s per-file limit on this 4-core container. develop's binary takes 440 s for it on the same machine, and CI's runners finish it inside the limit.
• norminette is OK on the touched file.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RAmeHfJNm7XjYbMNrQqvkG

---

Generated by Claude Code