Foreign memory access hot loop benchmark

[Maurizio Cimadamore]

On 19/11/2020 10:24, Antoine Chambille wrote:
> But for the specific benchmark we're looking at, the one with manual 
> unrolling (AddBenchmark.unrolledMHI_v2), I don't feel like the low 
> performance is due to the absence of auto-vectorization. As Vlad 
> recently mentioned, auto-vectorization is never enabled when Unsafe of 
> VarHandle is used. Also the 20x speed drop is very large, more than 
> the typical boost of auto-vectorization. Doesn't it look like 
> something more basic like the absence of inlining, or a Java method 
> not being replaced with its intrinsic ?

Sorry, I overlooked AddBenchmark.unrolledMHI_v2 - it is indeed a lot 
slower than the base case, which doesn't make a lot of sense. My guess 
is that the recently added optimization for double/float fail in this 
specific case, so we end up with same performances as before.

Can you try with longs (instead of doubles) and see what happens?

Thanks
Maurizio