base64: avoid compliler boundary checks in the hot loop

Unfortunately compiler is not capable to realise that boundary checks
are not necessary on each src[si+N] access.
Selection of sub-slice helps.

benchmark                        old MB/s     new MB/s     speedup
BenchmarkEncodeToString-8        678.24       675.38       1.00x
BenchmarkDecodeString/2-8        111.61       109.20       0.98x
BenchmarkDecodeString/4-8        183.22       195.49       1.07x
BenchmarkDecodeString/8-8        249.11       249.52       1.00x
BenchmarkDecodeString/64-8       653.02       697.99       1.07x
BenchmarkDecodeString/8192-8     950.17       1028.51      1.08x

Author: kirillx

src/encoding/base64/base64.go

@@ -488,15 +488,16 @@ func (enc *Encoding) Decode(dst, src []byte) (n int, err error) {
 
 	si := 0
 	for strconv.IntSize >= 64 && len(src)-si >= 8 && len(dst)-n >= 8 {
+		src2 := src[si : si+8]
 		if dn, ok := assemble64(
-			enc.decodeMap[src[si+0]],
-			enc.decodeMap[src[si+1]],
-			enc.decodeMap[src[si+2]],
-			enc.decodeMap[src[si+3]],
-			enc.decodeMap[src[si+4]],
-			enc.decodeMap[src[si+5]],
-			enc.decodeMap[src[si+6]],
-			enc.decodeMap[src[si+7]],
+			enc.decodeMap[src2[0]],
+			enc.decodeMap[src2[1]],
+			enc.decodeMap[src2[2]],
+			enc.decodeMap[src2[3]],
+			enc.decodeMap[src2[4]],
+			enc.decodeMap[src2[5]],
+			enc.decodeMap[src2[6]],
+			enc.decodeMap[src2[7]],
 		); ok {
 			binary.BigEndian.PutUint64(dst[n:], dn)
 			n += 6
@@ -512,11 +513,12 @@ func (enc *Encoding) Decode(dst, src []byte) (n int, err error) {
 	}
 
 	for len(src)-si >= 4 && len(dst)-n >= 4 {
+		src2 := src[si : si+4]
 		if dn, ok := assemble32(
-			enc.decodeMap[src[si+0]],
-			enc.decodeMap[src[si+1]],
-			enc.decodeMap[src[si+2]],
-			enc.decodeMap[src[si+3]],
+			enc.decodeMap[src2[0]],
+			enc.decodeMap[src2[1]],
+			enc.decodeMap[src2[2]],
+			enc.decodeMap[src2[3]],
 		); ok {
 			binary.BigEndian.PutUint32(dst[n:], dn)
 			n += 3