Operations with loop variables

[kose-y]

I have been playing around with this package, and it looks like there are many unimplemented corner cases involving operations with loop variables.

With version 0.6.27:

using LoopVectorization
function rshift_i!(out)
    n = length(out)
    @avx for i in 1:n
        out[i] = out[i] << i
    end
end

n = 512
out = Array{Int}(undef, n)
rshift_i!(out)

MethodError: no method matching <<(::VectorizationBase._MM{4,Int64}, ::VectorizationBase.SVec{4,Int64})
Closest candidates are:
  <<(!Matched::VectorizationBase.Static{N}, ::Any) where N at /home/kose/.julia/packages/VectorizationBase/OOwYB/src/static.jl:89
  <<(!Matched::VectorizationBase.Mask{W,U} where U<:Unsigned, ::Any) where W at /home/kose/.julia/packages/VectorizationBase/OOwYB/src/masks.jl:34
  <<(!Matched::VectorizationBase.SVec{W,T}, ::VectorizationBase.SVec{W,T}) where {W, T<:Union{Int128, Int16, Int32, Int64, Int8, UInt128, UInt16, UInt32, UInt64, UInt8, Ptr}} at /home/kose/.julia/packages/SIMDPirates/thWlM/src/SIMDPirates.jl:80
  ...

Stacktrace:
 [1] macro expansion at ./gcutils.jl:91 [inlined]
 [2] rshift_i!(::Array{Int64,1}) at ./In[78]:3
 [3] top-level scope at In[80]:3

function one_plus_i!(out)
    n = length(out)
    @avx for i in 1:n
        out[i] = 1 + i
    end
end

n = 512
out = Array{Int}(undef, n)
one_plus_i!(out)

MethodError: no method matching vstore!(::Ptr{Int64}, ::VectorizationBase._MM{4,Int64}, ::VectorizationBase._MM{4,Int64})
Closest candidates are:
  vstore!(::Ptr{T}, !Matched::Tuple{Vararg{VecElement{T},W}}, ::I, !Matched::Val{Aligned}, !Matched::Val{Nontemporal}) where {W, T, I, Aligned, Nontemporal} at /home/kose/.julia/packages/SIMDPirates/thWlM/src/memory.jl:329
  vstore!(::Ptr{T}, !Matched::Union{Tuple{Vararg{VecElement{T},W}}, VectorizationBase.AbstractStructVec{W,T}}, ::VectorizationBase._MM{W,I} where I<:Number) where {W, T} at /home/kose/.julia/packages/SIMDPirates/thWlM/src/memory.jl:474
  vstore!(::Ptr{T}, !Matched::Union{Tuple{Vararg{VecElement{T},W}}, VectorizationBase.AbstractStructVec{W,T}}, ::VectorizationBase._MM{W,I} where I<:Number, !Matched::Unsigned) where {W, T} at /home/kose/.julia/packages/SIMDPirates/thWlM/src/memory.jl:474
  ...

Stacktrace:
 [1] vstore! at /home/kose/.julia/packages/VectorizationBase/OOwYB/src/vectorizable.jl:197 [inlined]
 [2] macro expansion at ./gcutils.jl:91 [inlined]
 [3] one_plus_i!(::Array{Int64,1}) at ./In[72]:3
 [4] top-level scope at In[74]:3

[chriselrod]

I added these examples to the test suite.

Please file new issues as they come up!
PRs are welcome too.

For reference, the actual fixes were in SIMDPirates.jl.
Eventually, it'll need a refactoring with a focus on better type promotion behavior.

[chriselrod]

I cannot reproduce the test failures on Macs on travis.
I have neither a Mac, nor a CPU without AVX2. The latter is probably the problem. AVX2 added the SIMD shift operations. Travis's Macs do not have them.

The problem is probably that shifts are sometimes undefined when they exceed the number of bits of the element,