Minor performance improvements

[giltho]

I got Claude to write performance improvements for Obol. It tried like 10 different things over 1h30 and only 2 of them make a vague improvement (5% total). It says the low-hanging fruits are really within Charon (and were already implemented in AeneasVerif/charon#1173.

It's not much, and I'm not sure we can do much better because, apparently, 55% of runtime is spent in const-eval.

Here's the summary written by Claude, I went through this and it doesn't sound banana and 5% is always good.

Summary

Two small caches in translate_raw_span that together cut ~4.6% off translation time on a real-world workload.

Spans are translated once per MIR statement, terminator, local, and item — so this is one of the hottest paths in the driver. Each call was
re-running path normalization, FileName hashing, and SourceMap::lookup_char_pos for spans we'd already seen.

• ddc49d4 — cache FileId by SourceFile stable id. translate_filename walks path components and strips sysroot/cargo prefixes;
  register_file then hashes a FileName. Caching on rustc's StableSourceFileId (already 1:1 with the source file) skips all of that after
  the first hit per file. Falls through to register_file on miss, so dedup semantics are unchanged.
• a9bbacf — memoize translate_raw_span by ty::Span. Many statements share the same expansion span (inlining, macros). Caching the
  full SpanData skips both the file lookup and the two lookup_char_pos calls.

Measurements

signalapp/SparsePostQuantumRatchet, hyperfine with --prepare 'touch src/lib.rs', 20 runs side-by-side:

Version	Mean	vs baseline
main	6.232 s ± 0.068	1.00×
after ddc49d4	~5.94 s	1.05×
after a9bbacf (this branch)	5.944 s ± 0.056	1.046× faster (~3σ)

Profile confirms the mechanism: translate_span_from_smir inclusive drops from 4.92% → 0.43% across the two commits.

[N1ark]

nice