perf: make quoted-printable decoding beat 0.9.0 on dense escapes too - #264

Comparing the 0.9.0 release against master turned up exactly one benchmark where the release still won: parse_qp_dense_escapes, by 58% on the x86 gate and 20% on an M4. This closes that gap and then some.

Why it was there

#229 replaced the quoted_printable crate with a run-copying decoder and measured −39%. That was true of the benchmarks that existed — but parse_qp_dense_escapes arrived later, with #223, so the one input the new decoder is worst at was never compared against the code it replaced.

Measured directly on =C3=A9 × 20000 (120 KB, an escape every three bytes):

The two changes

Both in vendor/mailparse/src/qp.rs.

1. Look under the cursor before reaching for memchr. After an escape the next byte is very often another = — every non-ASCII character encodes as two or three consecutive escapes — so memchr was called only to report a match at offset 0, paying its SIMD setup each time. On the dense fixture that was 40,000 calls. This is why the fix helps ordinary mail too, not just the pathological case.

2. Find the first dropped byte a chunk at a time. The rule-1 pre-scan used position, which cannot vectorise because it must stop at the first hit. It now reduces 32-byte chunks with |= and only falls back to a byte-wise scan inside the chunk that failed. is_kept is respelled as arithmetic (b - 0x20 < 0x5F, b - 9 < 2, b == 0x0D) so that reduction is branchless. On 120 KB with nothing to drop — nearly every real body — 63 µs → 5.5 µs.

Results

Local interleaved A/B, Apple M4, 3 rounds, controls within 3.1%:

Everything else is inside the noise floor. Against 0.9.0 the dense case is now 1.40× faster instead of 1.20× slower, and ordinary quoted-printable 2.6× faster.

Output is unchanged

is_kept was respelled, so the risk is that the set changed. is_kept_is_the_same_set checks all 256 bytes against the original matches! spelling. On top of that the existing crate-agreement tests still pass — the generated corpus at every length, all 65,536 =xy byte pairs, the per-rule cases and the real fixture — and qp_agreement fuzzed 7.3M executions against the crate with no disagreement.

upstream.patch is regenerated (check_vendored_mailparse.sh passes) and PATCH.md records both changes and why they are shaped that way, so the next mailparse hand-merge does not quietly drop them.