Determine how to maximally elide intermediate allocations in the spectral broadcasting code

[jakebolewski]

@simonbyrne mentioned that under some circumstances with StaticArrays:

function apply_slab()
    out = MArray{}...
    ....
    return SMatrix(out)
end

will not heap allocate the temporary intermediate array if the compiler can infer the lifetime of the mutable array pointer. It doesn't look like that optimization is currently always triggering and needs to be further investigated. In general we probably need to re-think slab allocation optimization under the new operator structure.

[simonbyrne]

@vchuravy any idea on what we should look at here?

[jakebolewski]

It looks like Trixi.jl pre-allocates the temporaries in thread local array data structures

[jakebolewski]

Removing the above form (just returning the MAray) for more complex examples is a ~10-15% slowdown so it is doing something but not the extent of what we expected.

[vchuravy]

cc: @aviatesk (This is a prime example why the IPO based escape analysis is important).

Right currently the following invariants must hold for allocation elision to be triggered.

1. The object shall not escape the current function
2. If the object is passed to a function as an argument -- said function must be inlined.
3. No use of the object in an error branch (looks like an escape in LLVM)

For the GPU code we used to guarantee 2. through KernelAbstractions and 3. through the use of @inbounds, to avoid escape though bounds errors.