On 06/06/16 21:44, Jeff Hain wrote:
> With multiply it's faster when not inlined, but slower when inlined.
> For some reason the score error is smaller with multiply.
The other thing to bear in mind is that performance depends on how the
method is used. For example, if a benchmark tests uses positive
numbers the negative case will be moved out of line and all we're left
with is a single test and branch instruction:
tbnz x10, #63, deoptimize_label
This instruction will usually be correctly predicted so costs very
little, maybe nothing. So you will get different results depending on
the banchmark, and the benchmark thus needs to be representative of
real-world use. Whatever that is! This optimization game is hard.
:-)
Andrew.
