perf: compiled-runtime output flushing and match-table transcoding (1.5x)

[ddehilster]

What

Two changes to the compiled-analyzer runtime, both found by profiling rather than by reading the code.

1. Arun::out flushed the stream on every <<. NLP-ENGINE-499/505 added that so print-debugging survives a crash, with a comment saying the cost was negligible "for typical diagnostic volumes". parse-en-us writes ~6 MB across 19 files per run (a 2.35 MB out.xml), one token at a time, so it became a WriteFile plus a file-position query per write.

Durability now comes from a new Parse::flushostrs() called from Parse::finPass, so every completed pass's output is on disk if the process dies — O(passes) instead of O(writes). NLP_FLUSH_EVERY_WRITE=1 restores the old behaviour for bisecting a crash within one pass.

2. find_str_nocase rebuilt an icu::UnicodeString for every element of a match list on every call, even though every caller passes one of the static const tables emitted into the generated analyzer code. Now cached in a direct-mapped table keyed on the array address.

Profile (before)

Compiled -COMPILED run of parse-en-us, x86 RelWithDebInfo, 37.5 KB input, ~7k stack samples:

symbol	self
ntdll!ZwWriteFile	25.5%
ntdll!ZwQueryInformationFile	13.2%
find_str_nocase (+ICU under it)	~18% inclusive
Var::find + Ipair::getKey + Delt<Ipair>::getData	~1%

Arun::out was 34.7% inclusive, ostream::flush 32.1%, fflush 32.0%.

That last row is worth noting: the string-keyed variable lookup (Var::find's linear list scan with _tcscmp) looks like the obvious villain when reading the generated code, but it is about 1% of runtime.

Results

Interleaved A/B, 8 runs per arm, parse-en-us. The machine had ~35% background load, so min and p25 are the meaningful columns:

	min	p25	median
flush every write (old)	17.19s	17.84s	19.59s
flush per pass (new)	11.69s	11.72s	13.03s

1.47x – 1.52x.

After the fix, find_str_nocase self time went 13.4% → 4.5%. (A first attempt using std::map for that cache was a complete wash — the tree lookup cost as much as the transcoding it saved. The direct-mapped version is what paid off.)

Correctness

All 18 analyzer output files are byte-identical before and after. Only dbg.log differs, because it contains the per-pass timings.

Not in this PR

run!_chkstk is ~6% of self time. The generator emits one giant matchRule<N> per pass containing a switch over every rule, so the frame is sized for the union of all rules and every rule execution probes it. Confirmed the frames exceed a page: building with /Gs1048576 crashes with an access violation. The fix is emitting one function per rule, which is a change in the analyzer code generator — separate PR.

🤖 Generated with Claude Code