string_decoder: optimize write()

[mscdex]

By limiting property getting/setting to only where they are
absolutely necessary, we can achieve greater performance
especially with small utf8 inputs and any size base64 inputs.

ascii/binary are unaffected.

Example results for utf8 and base64 from the included benchmark:

misc/string-decode.js encoding=utf8 inlen=32 chunk=16 n=250000: ./iojs: 1946800 ./iojs-io.js-4655521: 1654400 ........... 17.67%
misc/string-decode.js encoding=utf8 inlen=32 chunk=64 n=250000: ./iojs: 2838500 ./iojs-io.js-4655521: 2545000 ........... 11.53%
misc/string-decode.js encoding=utf8 inlen=32 chunk=256 n=250000: ./iojs: 2822200 ./iojs-io.js-4655521: 2522400 .......... 11.88%
misc/string-decode.js encoding=utf8 inlen=32 chunk=1024 n=250000: ./iojs: 2829100 ./iojs-io.js-4655521: 2507800 ......... 12.81%
misc/string-decode.js encoding=utf8 inlen=128 chunk=16 n=250000: ./iojs: 508720 ./iojs-io.js-4655521: 499830 ............. 1.78%
misc/string-decode.js encoding=utf8 inlen=128 chunk=64 n=250000: ./iojs: 1016400 ./iojs-io.js-4655521: 931340 ............ 9.13%
misc/string-decode.js encoding=utf8 inlen=128 chunk=256 n=250000: ./iojs: 1250200 ./iojs-io.js-4655521: 1179200 .......... 6.02%
misc/string-decode.js encoding=utf8 inlen=128 chunk=1024 n=250000: ./iojs: 1258600 ./iojs-io.js-4655521: 1193400 ......... 5.46%
misc/string-decode.js encoding=utf8 inlen=1024 chunk=16 n=250000: ./iojs: 64944 ./iojs-io.js-4655521: 64852 .............. 0.14%
misc/string-decode.js encoding=utf8 inlen=1024 chunk=64 n=250000: ./iojs: 134380 ./iojs-io.js-4655521: 133250 ............ 0.85%
misc/string-decode.js encoding=utf8 inlen=1024 chunk=256 n=250000: ./iojs: 178810 ./iojs-io.js-4655521: 176190 ........... 1.49%
misc/string-decode.js encoding=utf8 inlen=1024 chunk=1024 n=250000: ./iojs: 215380 ./iojs-io.js-4655521: 212140 .......... 1.52%
misc/string-decode.js encoding=base64-utf8 inlen=32 chunk=16 n=250000: ./iojs: 407580 ./iojs-io.js-4655521: 307650 ...... 32.48%
misc/string-decode.js encoding=base64-utf8 inlen=32 chunk=64 n=250000: ./iojs: 1422000 ./iojs-io.js-4655521: 1120100 .... 26.96%
misc/string-decode.js encoding=base64-utf8 inlen=32 chunk=256 n=250000: ./iojs: 1416700 ./iojs-io.js-4655521: 1160400 ... 22.09%
misc/string-decode.js encoding=base64-utf8 inlen=32 chunk=1024 n=250000: ./iojs: 1339300 ./iojs-io.js-4655521: 1140600 .. 17.42%
misc/string-decode.js encoding=base64-utf8 inlen=128 chunk=16 n=250000: ./iojs: 93840 ./iojs-io.js-4655521: 84048 ....... 11.65%
misc/string-decode.js encoding=base64-utf8 inlen=128 chunk=64 n=250000: ./iojs: 383150 ./iojs-io.js-4655521: 308310 ..... 24.28%
misc/string-decode.js encoding=base64-utf8 inlen=128 chunk=256 n=250000: ./iojs: 1067600 ./iojs-io.js-4655521: 886690 ... 20.41%
misc/string-decode.js encoding=base64-utf8 inlen=128 chunk=1024 n=250000: ./iojs: 1011700 ./iojs-io.js-4655521: 926400 ... 9.21%
misc/string-decode.js encoding=base64-utf8 inlen=1024 chunk=16 n=250000: ./iojs: 13682 ./iojs-io.js-4655521: 10591 ...... 29.19%
misc/string-decode.js encoding=base64-utf8 inlen=1024 chunk=64 n=250000: ./iojs: 48976 ./iojs-io.js-4655521: 39895 ...... 22.76%
misc/string-decode.js encoding=base64-utf8 inlen=1024 chunk=256 n=250000: ./iojs: 141090 ./iojs-io.js-4655521: 120230 ... 17.35%
misc/string-decode.js encoding=base64-utf8 inlen=1024 chunk=1024 n=250000: ./iojs: 269880 ./iojs-io.js-4655521: 244070 .. 10.57%
misc/string-decode.js encoding=base64-ascii inlen=32 chunk=16 n=250000: ./iojs: 497550 ./iojs-io.js-4655521: 396600 ..... 25.45%
misc/string-decode.js encoding=base64-ascii inlen=32 chunk=64 n=250000: ./iojs: 1521700 ./iojs-io.js-4655521: 1174800 ... 29.52%
misc/string-decode.js encoding=base64-ascii inlen=32 chunk=256 n=250000: ./iojs: 1487500 ./iojs-io.js-4655521: 1174100 .. 26.69%
misc/string-decode.js encoding=base64-ascii inlen=32 chunk=1024 n=250000: ./iojs: 1401600 ./iojs-io.js-4655521: 1155800 . 21.27%
misc/string-decode.js encoding=base64-ascii inlen=128 chunk=16 n=250000: ./iojs: 142030 ./iojs-io.js-4655521: 114770 .... 23.75%
misc/string-decode.js encoding=base64-ascii inlen=128 chunk=64 n=250000: ./iojs: 499170 ./iojs-io.js-4655521: 396290 .... 25.96%
misc/string-decode.js encoding=base64-ascii inlen=128 chunk=256 n=250000: ./iojs: 1145200 ./iojs-io.js-4655521: 999490 .. 14.57%
misc/string-decode.js encoding=base64-ascii inlen=128 chunk=1024 n=250000: ./iojs: 1207200 ./iojs-io.js-4655521: 983240 . 22.77%
misc/string-decode.js encoding=base64-ascii inlen=1024 chunk=16 n=250000: ./iojs: 18163 ./iojs-io.js-4655521: 14729 ..... 23.31%
misc/string-decode.js encoding=base64-ascii inlen=1024 chunk=64 n=250000: ./iojs: 69888 ./iojs-io.js-4655521: 60726 ..... 15.09%
misc/string-decode.js encoding=base64-ascii inlen=1024 chunk=256 n=250000: ./iojs: 194780 ./iojs-io.js-4655521: 141810 .. 37.36%
misc/string-decode.js encoding=base64-ascii inlen=1024 chunk=1024 n=250000: ./iojs: 356430 ./iojs-io.js-4655521: 342270 .. 4.14%

[rvagg]

nice, I like those utf8 numbers, any idea how much property getting is impacting this?

LGTM

[mscdex]

I made these changes awhile ago so I can't recall for sure, but this patch is almost entirely made up of changes to avoid property lookups, so I would say that it is impacting things the most (especially in the loop in write()) versus some of the other minor non-property-related changes.

[mscdex]

/cc @iojs/collaborators

[micnic]

LGTM

[petkaantonov]

This should be implemented in C, we will need utf8 decoder in the future for buffer.toString() which is currently delegating to V8's utf8 decoder.

[mscdex]

@petkaantonov Why? Is V8 dropping their UTF-8 decoder soon?

[petkaantonov]

@mscdex No but instead of using NewFromUtf8 (which v8 internally decodes and copies to jsheap), we would decode it ourselves into an external buffer which doesn't take space in jsheap and also allows string sizes bigger than 290mb or so. See #649

[chrisdickinson]

I wonder if V8 will unroll the subsequent loop (or if there's value in doing so ourselves.)

[mscdex]

You mean for the i = 3 case? I didn't try that since I figured people would frown about code duplication between the i = 3 and i = buflen cases. I also did not see what kind performance difference it would make though.

[chrisdickinson]

This LGTM.

This should be implemented in C, we will need utf8 decoder in the future for buffer.toString() which is currently delegating to V8's utf8 decoder.

We'll still have to support this module if we move to a C decoder, so making it faster is probably worthwhile.

[mscdex]

Landed in 8a94581.