code review 6533048: compress/flate: move the history buffer out of the deco...

compress/flate: move the history buffer out of the decompressor struct.

I'm not exactly sure why there's a performance gain, but it seems like
an easy win. Maybe it's a cache line thing. Maybe it's that
unsafe.Sizeof(decompressor{}) drops to below unmappedzero, so that
checkref/checkoffset don't need to insert TESTB instructions. Maybe
it's less noise for the conservative garbage collector. Maybe it's
something else.

compress/flate benchmarks:
BenchmarkDecodeDigitsSpeed1e4          378628       349906   -7.59%
BenchmarkDecodeDigitsSpeed1e5         3481976      3204898   -7.96%
BenchmarkDecodeDigitsSpeed1e6        34419500     31750660   -7.75%
BenchmarkDecodeDigitsDefault1e4        362317       335562   -7.38%
BenchmarkDecodeDigitsDefault1e5       3290032      3107624   -5.54%
BenchmarkDecodeDigitsDefault1e6      30542540     28937480   -5.26%
BenchmarkDecodeDigitsCompress1e4       362803       335158   -7.62%
BenchmarkDecodeDigitsCompress1e5      3294512      3114526   -5.46%
BenchmarkDecodeDigitsCompress1e6     30514940     28927090   -5.20%
BenchmarkDecodeTwainSpeed1e4           412818       389521   -5.64%
BenchmarkDecodeTwainSpeed1e5          3475780      3288908   -5.38%
BenchmarkDecodeTwainSpeed1e6         33629640     31931420   -5.05%
BenchmarkDecodeTwainDefault1e4         369736       348850   -5.65%
BenchmarkDecodeTwainDefault1e5        2861050      2721383   -4.88%
BenchmarkDecodeTwainDefault1e6       27120120     25862050   -4.64%
BenchmarkDecodeTwainCompress1e4        372057       350822   -5.71%
BenchmarkDecodeTwainCompress1e5       2855109      2718664   -4.78%
BenchmarkDecodeTwainCompress1e6      26987010     26336030   -2.41%

image/png benchmarks:
BenchmarkDecodeGray               1841839      1802251   -2.15%
BenchmarkDecodeNRGBAGradient      7115318      6933280   -2.56%
BenchmarkDecodeNRGBAOpaque        6135892      6013284   -2.00%
BenchmarkDecodePaletted           1153313      1114302   -3.38%
BenchmarkDecodeRGB                5619404      5511190   -1.93%

[nigeltao, 2012-09-19 02:56:05]

Hello rsc@golang.org (cc: golang-dev@googlegroups.com),

I'd like you to review this change to
https://code.google.com/p/go

[r, 2012-09-19 03:00:18]

i'd be curious how it compares if the buffer was the last field of the struct

[r, 2012-09-19 03:00:38]

(as an array again)

[nigeltao, 2012-09-19 03:09:25]

On 19 September 2012 13:00,  <r@golang.org> wrote:
> i'd be curious how it compares if the buffer was the last field of the
> struct

Yeah, I tried that: the speed-up was about half as good (i.e. 2-3%
instead of 5-7%) as the indirect through a pointer.

[nigeltao, 2012-09-24 07:58:18]

*** Submitted as http://code.google.com/p/go/source/detail?r=3d7777fe2e8a ***

compress/flate: move the history buffer out of the decompressor struct.

I'm not exactly sure why there's a performance gain, but it seems like
an easy win. Maybe it's a cache line thing. Maybe it's that
unsafe.Sizeof(decompressor{}) drops to below unmappedzero, so that
checkref/checkoffset don't need to insert TESTB instructions. Maybe
it's less noise for the conservative garbage collector. Maybe it's
something else.

compress/flate benchmarks:
BenchmarkDecodeDigitsSpeed1e4          378628       349906   -7.59%
BenchmarkDecodeDigitsSpeed1e5         3481976      3204898   -7.96%
BenchmarkDecodeDigitsSpeed1e6        34419500     31750660   -7.75%
BenchmarkDecodeDigitsDefault1e4        362317       335562   -7.38%
BenchmarkDecodeDigitsDefault1e5       3290032      3107624   -5.54%
BenchmarkDecodeDigitsDefault1e6      30542540     28937480   -5.26%
BenchmarkDecodeDigitsCompress1e4       362803       335158   -7.62%
BenchmarkDecodeDigitsCompress1e5      3294512      3114526   -5.46%
BenchmarkDecodeDigitsCompress1e6     30514940     28927090   -5.20%
BenchmarkDecodeTwainSpeed1e4           412818       389521   -5.64%
BenchmarkDecodeTwainSpeed1e5          3475780      3288908   -5.38%
BenchmarkDecodeTwainSpeed1e6         33629640     31931420   -5.05%
BenchmarkDecodeTwainDefault1e4         369736       348850   -5.65%
BenchmarkDecodeTwainDefault1e5        2861050      2721383   -4.88%
BenchmarkDecodeTwainDefault1e6       27120120     25862050   -4.64%
BenchmarkDecodeTwainCompress1e4        372057       350822   -5.71%
BenchmarkDecodeTwainCompress1e5       2855109      2718664   -4.78%
BenchmarkDecodeTwainCompress1e6      26987010     26336030   -2.41%

image/png benchmarks:
BenchmarkDecodeGray               1841839      1802251   -2.15%
BenchmarkDecodeNRGBAGradient      7115318      6933280   -2.56%
BenchmarkDecodeNRGBAOpaque        6135892      6013284   -2.00%
BenchmarkDecodePaletted           1153313      1114302   -3.38%
BenchmarkDecodeRGB                5619404      5511190   -1.93%

R=rsc, r
CC=golang-dev
http://codereview.appspot.com/6533048