Reduction vector optimization could be expensive for simple expressions because it uses several additional instructions per vector. For example, simple test shows that only int[] multiply has benefit from this optimization (see below). We need to restrict reduction optimization only to cases when it is beneficial.


On SB (AVX1):

java -XX:-SuperWordReductions -XX:-TieredCompilation -Xbatch -XX:CompileCommand=exclude,Reduction::main Reduction
Warmup ...
 Warmup is done in 1160 msec
sum int: 121
sum long: 120
sum float: 362
sum double: 353
mul int: 356
mul long: 357
mul float: 593
mul double: 588

java -XX:+SuperWordReductions -XX:-TieredCompilation -Xbatch -XX:CompileCommand=exclude,Reduction::main Reduction
Warmup ...
 Warmup is done in 1216 msec
sum int: 123
sum long: 119
sum float: 436
sum double: 407
mul int: 204
mul long: 349
mul float: 669
mul double: 639


On NHM (AVX2):

java -XX:-SuperWordReductions -XX:-TieredCompilation -Xbatch -XX:CompileCommand=exclude,Reduction::main Reduction
Warmup ...
 Warmup is done in 1408 msec
sum int: 139
sum long: 139
sum float: 412
sum double: 412
mul int: 412
mul long: 412
mul float: 686
mul double: 686

java -XX:+SuperWordReductions -XX:-TieredCompilation -Xbatch -XX:CompileCommand=exclude,Reduction::main Reduction
Warmup ...
 Warmup is done in 1340 msec
sum int: 123
sum long: 139
sum float: 412
sum double: 412
mul int: 209
mul long: 412
mul float: 686
mul double: 687