crypto/cipher: optimize safeXORBytes

[josharian]

The discussion at #31586 includes several techniques for optimizing xor of byte slices, including using encoding/binary (with native endianness) and loop unrolling. It'd be nice to apply them to safeXORBytes, which is used on many non-amd64 platforms. An optimized safeXORBytes might even be fast enough that we could delete fastXORBytes.

Somewhat related: #30553

cc @nhooyr

[josharian]

(Aside: I've wanted encoding/binary.NativeEndian to exist on multiple occasions. Here's another one.)

[nhooyr]

Will give this a shot on the weekend.

[nhooyr]

(Aside: I've wanted encoding/binary.NativeEndian to exist on multiple occasions. Here's another one.)

I agree this would be very useful. Some discussion in the past: https://groups.google.com/forum/#!topic/golang-nuts/3GEzwKfRRQw

[josharian]

Will give this a shot on the weekend.

Awesome. No rush, in any case—the freeze just started, so nothing will go in for three months.

Thanks for the pointer to previous discussion of native endianness. I’ll file a proposal to discuss.

[bcmills]

If you implement a generic, fast xor function, it could probably stand to be used in x/crypto/sha3 — the current implementation there is somewhat dubious due to alignment assumptions.

(See CL 203837.)

CC @FiloSottile

[josharian]

@bcmills good point, thanks. If we put it in a place that both x/crypto and crypto/cipher can reach it, that'd presumably also resolve #30553.

[josharian]

Proposal at #35398.

[nhooyr]

Added in nhooyr@24ce00e

Results against xorBytesSSE2 on amd64:

$ /Users/nhooyr/src/golang/go/bin/go test -run=_ -bench=XORBytes
goos: darwin
goarch: amd64
pkg: crypto/cipher
BenchmarkXORBytes/8Bytes-8    	176461338	         6.70 ns/op	1194.11 MB/s
BenchmarkXORBytes/128Bytes-8  	100000000	        11.6 ns/op	11062.06 MB/s
BenchmarkXORBytes/2048Bytes-8 	12895827	        95.2 ns/op	21517.42 MB/s
BenchmarkXORBytes/32768Bytes-8         	  627291	      1596 ns/op	20527.27 MB/s
BenchmarkFastSafeXORBytes/8Bytes-8     	147817434	         7.92 ns/op	1009.53 MB/s
BenchmarkFastSafeXORBytes/128Bytes-8   	66896612	        17.2 ns/op	7438.88 MB/s
BenchmarkFastSafeXORBytes/2048Bytes-8  	 6217075	       185 ns/op	11069.10 MB/s
BenchmarkFastSafeXORBytes/32768Bytes-8 	  397446	      3007 ns/op	10897.15 MB/s
PASS
ok  	crypto/cipher	12.014s

So safe fast is still about 2x as slow. Still 10x faster in my testing than basic simple xor though.