Force avx to "avx = false" when use LHS scattering, unless the index is checked to be unique.

[N5N3]

using Tullio
using LoopVectorization
J = [1,3,1,1]
A = [i^2 for i in 1:10]

Example 1(wrong):

@tullio B[J[i],k] := A[k]
@tullio B[J[i],k] += A[k]
# 3×10 Array{Int64,2}:
#  1  4  9  16  25  36  49  64  81  100
#  0  0  0   0   0   0   0   0   0    0
#  1  4  9  16  25  36  49  64  81  100

Example 2(right):

@tullio B[J[i],k] := A[k]
@tullio avx=false B[J[i],k] += A[k]
# 3×10 Array{Int64,2}:
#  4  16  36  64  100  144  196  256  324  400
#  0   0   0   0    0    0    0    0    0    0
#  2   8  18  32   50   72   98  128  162  200

As @avx can't tell the user all the misusing, maybe the default set should be avx = false.

[chriselrod]

LoopVectorization issue:
JuliaSIMD/LoopVectorization.jl#145

The problem there is that if you try to write to the same memory location multiple times simultaneously, the answer is going to be undefined (but probably equal to one of those simultaneous write attempts).
This could also be considered a cumsum-like example, in that correct value of memory-reads are also changing between loop iterations.

I'd say this is a problem in documentation. @N5N3 should've been able to figure out that this is expected behavior.

[mcabbott]

Thanks for finding this. There is some logic to prevent multi-threading from messing this up, with += it sets unsafeind = [:i] which hides i from threader. It's possible this ought to also disable @avx entirely, I just need to remember how things work. The goal is certainly to make operations safe by default.

[N5N3]

So the SubArray ver is correct only because the size is small and the Multi-threads is not open. If true, I think this might be warned somewhere in the document. As it seem's impossible to check the indices of the SubArray.