Feat/fasterer

[ribasushi]

The core change is this

go-base36/base36.go

Lines 50 to 59 in 986b8a8

	// It is crucial to make this as short as possible, especially for
	// the usual case of CIDs.
	bufsz = zcnt +
	// This is an integer simplification of
	// ceil(log(256)/log(36))
	(bufsz-zcnt)*277/179 + 1
	// Note: pools *DO NOT* help, the overhead of zeroing
	// kills any performance gain to be had
	out := make([]byte, bufsz)

We now allocate just as much as we need, and iterate over the buffer in place: once to breakdown into values, and a second time to recode into an alphabet.

• before

go-multibase$ go test ./... -count=1 -bench .
goos: darwin
goarch: amd64
pkg: github.com/multiformats/go-multibase
BenchmarkRoundTrip/base36-8         	  326816	      3599 ns/op
BenchmarkRoundTrip/base36upper-8    	  332366	      3571 ns/op
BenchmarkRoundTrip/base58btc-8      	  481983	      2551 ns/op
BenchmarkRoundTrip/base58flickr-8   	  470001	      2556 ns/op
BenchmarkEncode/base36-8            	  435662	      2788 ns/op
BenchmarkEncode/base36upper-8       	  440614	      2788 ns/op
BenchmarkEncode/base58btc-8         	  770272	      1568 ns/op
BenchmarkEncode/base58flickr-8      	  700389	      1562 ns/op
BenchmarkDecode/base36-8            	 1606472	       745 ns/op
BenchmarkDecode/base36upper-8       	 1614858	       744 ns/op
BenchmarkDecode/base58btc-8         	 1245900	       952 ns/op
BenchmarkDecode/base58flickr-8      	 1259892	       956 ns/op
PASS

• after

go-multibase$ go test ./... -count=1 -bench .
goos: darwin
goarch: amd64
pkg: github.com/multiformats/go-multibase
BenchmarkRoundTrip/base36-8         	  432175	      2750 ns/op
BenchmarkRoundTrip/base36upper-8    	  436525	      2723 ns/op
BenchmarkRoundTrip/base58btc-8      	  466893	      2553 ns/op
BenchmarkRoundTrip/base58flickr-8   	  475454	      2547 ns/op
BenchmarkEncode/base36-8            	  601210	      1933 ns/op
BenchmarkEncode/base36upper-8       	  629655	      1938 ns/op
BenchmarkEncode/base58btc-8         	  791280	      1572 ns/op
BenchmarkEncode/base58flickr-8      	  795505	      1555 ns/op
BenchmarkDecode/base36-8            	 1558412	       766 ns/op
BenchmarkDecode/base36upper-8       	 1561417	       762 ns/op
BenchmarkDecode/base58btc-8         	 1250644	       954 ns/op
BenchmarkDecode/base58flickr-8      	 1263541	       954 ns/op
PASS

P.S. If you look closely you see that the decoding became marginally slower. To the best of my ability to reason it has to do with the total compiled object size. That is - if you revert 986b8a84d74, then the decoding number goes back to where it was, even though the reverted commit does not touch the decode method 🤯

[ribasushi]

👍
I'll tag this as part of multiformats/go-multibase#37 soon, want to let b58 bake a bit more...