compress/lzw: don't use a closure in NewReader, which avoids having
to move some variables from the stack to the heap.
Sorted benchmark runs on my 2007-era Mac Mini (GOARCH=amd64, GOOS=linux):
Before:
lzw.BenchmarkDecoder 2000 878176 ns/op
lzw.BenchmarkDecoder 2000 878415 ns/op
lzw.BenchmarkDecoder 2000 880352 ns/op
lzw.BenchmarkDecoder 2000 898445 ns/op
lzw.BenchmarkDecoder 2000 901728 ns/op
After:
lzw.BenchmarkDecoder 2000 859065 ns/op
lzw.BenchmarkDecoder 2000 859402 ns/op
lzw.BenchmarkDecoder 2000 860035 ns/op
lzw.BenchmarkDecoder 2000 860555 ns/op
lzw.BenchmarkDecoder 2000 861109 ns/op
The ratio of before/after median times is 1.024.
The runtime.MemStats.Mallocs delta per loop drops from 109 to 104.
R=r, r2, dfc
CC=golang-dev
https://golang.org/cl/
4253043