zstd: Add amd64 match length assembly (#824)

* zstd: Add amd64 match length assembly

Copied from the S2 implementation. 5-10% faster.

zstd/blockdec.go

@@ -592,7 +592,7 @@ func (b *blockDec) prepareSequences(in []byte, hist *history) (err error) {
 				}
 				seq.fse.setRLE(symb)
 				if debugDecoder {
-					printf("RLE set to %+v, code: %v", symb, v)
+					printf("RLE set to 0x%x, code: %v", symb, v)
 				}
 			case compModeFSE:
 				println("Reading table for", tableIndex(i))

zstd/matchlen_amd64.go

@@ -0,0 +1,16 @@
+//go:build amd64 && !appengine && !noasm && gc
+// +build amd64,!appengine,!noasm,gc
+
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+
+package zstd
+
+// matchLen returns how many bytes match in a and b
+//
+// It assumes that:
+//
+//	len(a) <= len(b) and len(a) > 0
+//
+//go:noescape
+func matchLen(a []byte, b []byte) int

zstd/matchlen_amd64.s

@@ -0,0 +1,68 @@
+// Copied from S2 implementation.
+
+//go:build !appengine && !noasm && gc && !noasm
+
+#include "textflag.h"
+
+// func matchLen(a []byte, b []byte) int
+// Requires: BMI
+TEXT ·matchLen(SB), NOSPLIT, $0-56
+	MOVQ a_base+0(FP), AX
+	MOVQ b_base+24(FP), CX
+	MOVQ a_len+8(FP), DX
+
+	// matchLen
+	XORL SI, SI
+	CMPL DX, $0x08
+	JB   matchlen_match4_standalone
+
+matchlen_loopback_standalone:
+	MOVQ  (AX)(SI*1), BX
+	XORQ  (CX)(SI*1), BX
+	TESTQ BX, BX
+	JZ    matchlen_loop_standalone
+
+#ifdef GOAMD64_v3
+	TZCNTQ BX, BX
+#else
+	BSFQ BX, BX
+#endif
+	SARQ $0x03, BX
+	LEAL (SI)(BX*1), SI
+	JMP  gen_match_len_end
+
+matchlen_loop_standalone:
+	LEAL -8(DX), DX
+	LEAL 8(SI), SI
+	CMPL DX, $0x08
+	JAE  matchlen_loopback_standalone
+
+matchlen_match4_standalone:
+	CMPL DX, $0x04
+	JB   matchlen_match2_standalone
+	MOVL (AX)(SI*1), BX
+	CMPL (CX)(SI*1), BX
+	JNE  matchlen_match2_standalone
+	LEAL -4(DX), DX
+	LEAL 4(SI), SI
+
+matchlen_match2_standalone:
+	CMPL DX, $0x02
+	JB   matchlen_match1_standalone
+	MOVW (AX)(SI*1), BX
+	CMPW (CX)(SI*1), BX
+	JNE  matchlen_match1_standalone
+	LEAL -2(DX), DX
+	LEAL 2(SI), SI
+
+matchlen_match1_standalone:
+	CMPL DX, $0x01
+	JB   gen_match_len_end
+	MOVB (AX)(SI*1), BL
+	CMPB (CX)(SI*1), BL
+	JNE  gen_match_len_end
+	INCL SI
+
+gen_match_len_end:
+	MOVQ SI, ret+48(FP)
+	RET

zstd/matchlen_generic.go

@@ -0,0 +1,33 @@
+//go:build !amd64 || appengine || !gc || noasm
+// +build !amd64 appengine !gc noasm
+
+// Copyright 2019+ Klaus Post. All rights reserved.
+// License information can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"encoding/binary"
+	"math/bits"
+)
+
+// matchLen returns the maximum common prefix length of a and b.
+// a must be the shortest of the two.
+func matchLen(a, b []byte) (n int) {
+	for ; len(a) >= 8 && len(b) >= 8; a, b = a[8:], b[8:] {
+		diff := binary.LittleEndian.Uint64(a) ^ binary.LittleEndian.Uint64(b)
+		if diff != 0 {
+			return n + bits.TrailingZeros64(diff)>>3
+		}
+		n += 8
+	}
+
+	for i := range a {
+		if a[i] != b[i] {
+			break
+		}
+		n++
+	}
+	return n
+
+}

zstd/zstd.go

@@ -9,7 +9,6 @@ import (
 	"errors"
 	"log"
 	"math"
-	"math/bits"
 )
 
 // enable debug printing
@@ -106,27 +105,6 @@ func printf(format string, a ...interface{}) {
 	}
 }
 
-// matchLen returns the maximum common prefix length of a and b.
-// a must be the shortest of the two.
-func matchLen(a, b []byte) (n int) {
-	for ; len(a) >= 8 && len(b) >= 8; a, b = a[8:], b[8:] {
-		diff := binary.LittleEndian.Uint64(a) ^ binary.LittleEndian.Uint64(b)
-		if diff != 0 {
-			return n + bits.TrailingZeros64(diff)>>3
-		}
-		n += 8
-	}
-
-	for i := range a {
-		if a[i] != b[i] {
-			break
-		}
-		n++
-	}
-	return n
-
-}
-
 func load3232(b []byte, i int32) uint32 {
 	return binary.LittleEndian.Uint32(b[:len(b):len(b)][i:])
 }