These benchmarks demonstrate that
the nilcheckelim pass is roughly O(n^2):

BenchmarkNilCheckDeep1    	 2000000	       741 ns/op	   1.35 MB/s
BenchmarkNilCheckDeep10   	 1000000	      2237 ns/op	   4.47 MB/s
BenchmarkNilCheckDeep100  	   20000	     60713 ns/op	   1.65 MB/s
BenchmarkNilCheckDeep1000 	     200	   7925198 ns/op	   0.13 MB/s
BenchmarkNilCheckDeep10000	       1	1220104252 ns/op	   0.01 MB/s

Profiling suggests that building the
dominator tree is also O(n^2),
and before size factors take over,
considerably more expensive than nilcheckelim.

Change-Id: If966b38ec52243a25f355dab871300d29db02e16
Reviewed-on: https://go-review.googlesource.com/11520Reviewed-by: Keith Randall <khr@golang.org>