On 6g/linux:
benchmark                     old ns/op    new ns/op    delta
BenchmarkFDCT                      4606         4241   -7.92%
BenchmarkIDCT                      4187         3923   -6.31%
BenchmarkDecodeBaseline         3154864      3170224   +0.49%
BenchmarkDecodeProgressive      4072812      4017132   -1.37%
BenchmarkEncode                39406920     34596760  -12.21%

Stack requirements before (from 'go tool 6g -S'):
(scan.go:37) TEXT    (*decoder).processSOS+0(SB),$1352-32
(writer.go:448) TEXT    (*encoder).writeSOS+0(SB),$5344-24

after:
(scan.go:37) TEXT    (*decoder).processSOS+0(SB),$1064-32
(writer.go:448) TEXT    (*encoder).writeSOS+0(SB),$2520-24

Also, in encoder.writeSOS, re-use the yBlock scratch buffer for Cb and
Cr. This reduces the stack requirement slightly, but also avoids an
unlucky coincidence where a BenchmarkEncode stack split lands between
encoder.writeByte and bufio.Writer.WriteByte, which occurs very often
during Huffman encoding and is otherwise disasterous for the final
benchmark number. FWIW, the yBlock re-use *without* the s/int/int32/
change does not have a noticable effect on the benchmarks.

R=r
CC=golang-dev, rsc
https://golang.org/cl/6823043