Before:
BenchmarkHeaderWriteSubset  500000  2354 ns/op  197 B/op  2 allocs/op
After:
BenchmarkHeaderWriteSubset 1000000  2085 ns/op    0 B/op  0 allocs/op

Fixes #3761

R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/7508043