Strange performance of a loop

[RoyiAvital]

@kristoffer.carlsson

The alignment is set here

https://github.com/JuliaLang/julia/blob/9ed628780776f59bd5e52fa3d7f8244d615a0301/src/julia_internal.h#L252-L253

and depends on the size of the array. Large arrays are already 64 bytes aligned:

julia> Int(pointer(rand(32))) % 64
48

julia> Int(pointer(rand(1024))) % 64
0

What you show is only the address of the 1st element.
What I suggest is making sure any Loop with vectorization won’t have “Anomaly” to take care of.

I meant something like Intel IPP.

If we define 1D array it will be padded to have size which is multiplication of 16 / 32 / 64 Bytes.
If you define 2D array it will be padded with rows which are also multiplication of 16 / 32 / 64 Bytes.

This way all loops will be able to be unrolled and vectorized with no issues about taking care of edge cases.