Support `getproperty`/`setproperty!` in reductions

[SkyWorld117]

trafficstars

julia> using LoopVectorization

julia> mutable struct foo
           s
       end

julia> f = foo(0)
foo(0)

julia> m = rand(Float32, 4,4)
4×4 Matrix{Float32}:
 0.925989     0.8695    0.784866  0.0669142
 0.258075     0.165664  0.787071  0.697996
 0.219509     0.580573  0.298625  0.479805
 0.000333428  0.200458  0.858297  0.571481

julia> s = 0
0

julia> @avxt for i in eachindex(m)
           s += ifelse(m[i]>0.5, 1, 0)
       end

julia> s
8

julia> @avxt for i in eachindex(m)
           f.s += ifelse(m[i]>0.5, 1, 0)
       end
ERROR: LoadError: LHS not understood; only `:ref`s and `:tuple`s are currently supported.
f.s
Stacktrace:
 [1] add_assignment!(ls::LoopVectorization.LoopSet, LHS::Expr, RHS::Expr, elementbytes::Int64, position::Int64)
   @ LoopVectorization C:\Users\89639\.julia\packages\LoopVectorization\wL8Qh\src\modeling\graphs.jl:1148
 [2] push!(ls::LoopVectorization.LoopSet, ex::Expr, elementbytes::Int64, position::Int64)
   @ LoopVectorization C:\Users\89639\.julia\packages\LoopVectorization\wL8Qh\src\modeling\graphs.jl:1165
in expression starting at REPL[8]:1

[chriselrod]

The problem is the f.s +=, not the ifelse.

julia> mutable struct foo
           s
       end

julia> f = foo(0f0);

julia> m = rand(Float32, 4,4);

julia> s = 0f0;

julia> @avxt for i in eachindex(m)
           s += m[i]
       end

julia> s, sum(m)
(8.579453f0, 8.579453f0)

julia> @avxt for i in eachindex(m)
           f.s += m[i]
       end
ERROR: LoadError: LHS not understood; only `:ref`s and `:tuple`s are currently supported.
f.s
Stacktrace:
 [1] add_assignment!(ls::LoopVectorization.LoopSet, LHS::Expr, RHS::Expr, elementbytes::Int64, position::Int64)
   @ LoopVectorization ~/.julia/dev/LoopVectorization/src/modeling/graphs.jl:1148

I'd accept a PR adding support for this, but otherwise it will have to wait for the LoopVectorization rewrite.

Basically, to support this, either add the transform of

for i in eachindex(m)
    f.s += m[i]
end

into

s = 0.0
for i in eachindex(m)
    s += m[i]
end
f.s += s

before LoopVectorization does anything else with the expression, or add special handling in the parsing to interpret it and then again here to do the store when _turbo_! returns. The former approach is easier, so I'd suggest that one. Ideally, the PR would also handle something like f.s[1] += ....