Zero out only that part of the hash table in use.
Obviously, the biggest gains are for the smallest inputs.
The WordsEncode1e1-8 benchmark should be unaffected (except possibly by
noise) because it never calls into encodeBlock.
name old speed new speed delta
WordsEncode1e1-8 678MB/s ± 0% 667MB/s ± 0% -1.74% (p=0.008 n=5+5)
WordsEncode1e2-8 90.1MB/s ± 0% 352.9MB/s ± 1% +291.85% (p=0.008 n=5+5)
WordsEncode1e3-8 295MB/s ± 0% 383MB/s ± 1% +29.67% (p=0.008 n=5+5)
WordsEncode1e4-8 276MB/s ± 0% 277MB/s ± 1% ~ (p=0.310 n=5+5)
WordsEncode1e5-8 248MB/s ± 0% 248MB/s ± 0% ~ (p=0.841 n=5+5)
WordsEncode1e6-8 295MB/s ± 0% 296MB/s ± 0% ~ (p=0.548 n=5+5)
RandomEncode-8 14.4GB/s ± 1% 14.4GB/s ± 2% ~ (p=1.000 n=5+5)
_ZFlat0-8 749MB/s ± 0% 748MB/s ± 0% ~ (p=0.151 n=5+5)
_ZFlat1-8 405MB/s ± 0% 406MB/s ± 0% +0.18% (p=0.032 n=4+5)
_ZFlat2-8 16.2GB/s ± 1% 16.1GB/s ± 1% ~ (p=0.421 n=5+5)
_ZFlat3-8 202MB/s ± 1% 604MB/s ± 0% +198.86% (p=0.008 n=5+5)
_ZFlat4-8 7.59GB/s ± 1% 7.62GB/s ± 1% ~ (p=0.548 n=5+5)
_ZFlat5-8 728MB/s ± 1% 729MB/s ± 0% ~ (p=0.548 n=5+5)
_ZFlat6-8 266MB/s ± 1% 267MB/s ± 0% ~ (p=0.548 n=5+5)
_ZFlat7-8 248MB/s ± 0% 248MB/s ± 0% ~ (p=0.881 n=5+5)
_ZFlat8-8 282MB/s ± 0% 282MB/s ± 0% ~ (p=0.556 n=4+5)
_ZFlat9-8 231MB/s ± 0% 231MB/s ± 0% +0.23% (p=0.032 n=5+5)
_ZFlat10-8 970MB/s ± 0% 972MB/s ± 0% ~ (p=0.421 n=5+5)
_ZFlat11-8 402MB/s ± 0% 401MB/s ± 0% ~ (p=0.730 n=5+5)
diff --git a/encode_amd64.s b/encode_amd64.s
index 56631f3..92f0a39 100644
--- a/encode_amd64.s
+++ b/encode_amd64.s
@@ -262,8 +262,11 @@
varTable:
// var table [maxTableSize]uint16
//
- // sizeof(table) is 32768 bytes, which is 2048 16-byte writes.
- MOVQ $2048, DX
+ // In the asm code, unlike the Go code, we can zero-initialize only the
+ // first tableSize elements. Each uint16 element is 2 bytes and each MOVOU
+ // writes 16 bytes, so we can do only tableSize/8 writes instead of the
+ // 2048 writes that would zero-initialize all of table's 32768 bytes.
+ SHRQ $3, DX
LEAQ table-32768(SP), BX
PXOR X0, X0
