[whoops, mis-pasted this] > so the question is what to do about it?
Identify actionable improvements. Maybe adding a branch somewhere and specializing on the most common powers is an acceptable solution. Best answer is fix LLVM bugs, some of this may go away when we switch to LLVM 3.5 which was just released, there's an open issue and several other problems so it won't be a simple version bump. > Everyone coming from Octave/Matlab will expect and should expect that X.^2 is usable and just as fast. Sure. Part of my point is because this is the only fast way to do things in Octave/Matlab, you're led to want to do things this way even in situations where it's not necessary or desirable. Do you actually need to make a new separate copy of the entire array? If not, you could square the elements in-place to save on allocation. Do you really need the squared value of every single element? If not, a reduction is a much faster way to get the information you're looking for. Expectations based on the way Octave/Matlab work are really not applicable to a system that works completely differently.
