[LLVMdev] [RFC] Heuristic for complete loop unrolling

[Owen Anderson]

> On Jan 24, 2015, at 5:02 AM, Sean Silva <chisophugis at gmail.com> wrote:
> 
> I actually think the general problem that this thread has been talking about (how to determine if a particular optimization is going to enable further optimizations) is very interesting in general, but I would take this particular example with a grain of salt because is a very specific stylized computation.

I see variations of this optimization come up in many contexts in graphics programming, not just convolutions.  It's very common in code that operates on 4x4 matrices, which comes up frequently in vertex shaders.

Convolutions and stencils are particularly interesting in that there are *even more* optimizations you can do for them, typically with the goal of improving cache locality.  But what Michael is proposing is useful for other classes of programs as well.

-Owen