mirror of
https://github.com/gohugoio/hugo.git
synced 2026-08-24 15:28:54 +00:00
Bjørn Erik Pedersen
65c82178b7
The old implementation copied the incoming bytes into a fixed-size buffer
one byte at a time, and for every byte ran bytes.Contains for every
pattern over the whole buffer. That's O(n·patterns) scans of the buffer,
which showed up badly on large output (a full rendered page).
Scan each not-yet-matched pattern once per Write with bytes.Contains over
the chunk itself, in place, rather than per byte. The only carried state
is a small boundary window (the last maxPatternLen-1 bytes) joined with
the head of the next chunk, so a pattern straddling a Write boundary is
still detected. The chunk is never copied, so the extra allocation is
bounded by the longest pattern and independent of the output size. Once
all patterns have matched we mark done and drop the buffer.
patternLen summed the pattern lengths (to size the old buffer); the
boundary window only needs the longest pattern, so it's renamed
maxPatternLen and returns the max.
```bash
│ benchcmp.bench │ fix-hasbytewriter.bench │
│ sec/op │ sec/op vs base │
HasBytesWriter-10 2473.095µ ± ∞ ¹ 6.114µ ± ∞ ¹ -99.75% (p=0.029 n=4)
¹ need >= 6 samples for confidence interval at level 0.95
│ benchcmp.bench │ fix-hasbytewriter.bench │
│ B/op │ B/op vs base │
HasBytesWriter-10 48.00 ± ∞ ¹ 128.00 ± ∞ ¹ +166.67% (p=0.029 n=4)
¹ need >= 6 samples for confidence interval at level 0.95
│ benchcmp.bench │ fix-hasbytewriter.bench │
│ allocs/op │ allocs/op vs base │
```