Make SubArrays immutable

[simonster]

Since #8867 this actually makes a difference in the optimizations LLVM is able to perform. On the benchmark from #8809, the effect is quite dramatic. The code is:

function f1(x,v)
    for j=1:length(x)
        @inbounds x[j] += v
    end
end
let
    n = 600
    tn = 10^7
    x = randn(n,3)
    v = 1.00001
    xslice = sub(x,:, 2)
    @time for i=1:tn
        f1(xslice,v)
    end
end

Before, this took ~9 seconds to execute. With this change, it takes ~2.5. However, LLVM 3.3 still won't vectorize it, even though it vectorizes with Arrays and ArrayViews. It looks like there's an extra mul in the inner loop.

[andreasnoack]

👍

[timholy]

I remember thinking about this and deciding not to do it, but for the life of me I can't remember why. It might have been some intermediate testing thing, or something. Let's do it.

[timholy]

And it would be really, really great to figure out the vectorization thing. Do we have any diagnostic tools that will tell us why it didn't vectorize?

[simonster]

I think that it's not vectorizing (at least on my system; from #9318 it sounds like other systems may vectorize non-unit strides) because of the multiplication by V.stride1 here. Otherwise the inner loop IR is identical to the IR for the non-vectorized branch for Arrays and ArrayViews. Perhaps we can avoid emitting that when we can determine that V.stride1 will be 1 based on the types?