Profiling revealed two main causes of this slowdown. One is the garbage collector invocations themselves; Epoch is designed to have a "greedy" collector which trips frequently and collects aggressively. Unfortunately, the cost of walking the stack and heap for live objects is nontrivial, so tripping the GC all the time leads to lower performance throughput.
It's worth noting that in standard apps like Era, the difference isn't even visible; I've yet to see Era hitch noticeably with the fixed garbage collector. But for something that wants to peg a CPU core and go flat-out as fast as possible, triggering collection cycles all the time is a bad idea.
The other major cause of performance woes is register spilling. Without going into painful amounts of detail, one of the requirements of LLVM's garbage collection support is that you can no longer store pointers to GC-able objects in CPU registers. Instead, all objects that might be collected must be stored in memory on the machine stack. Moreover, any time you want to update those values, it has to be done on the stack, so there is overhead in moving from register values back onto the stack every time you touch a variable that might be garbage collected.
This may not seem like much, but it turns out to be pretty painful in a benchmark that relies heavily on computations done on objects. Since every object can potentially be collected by the GC, every object's pointer must live on the stack at all times. This winds up being a substantial perf hit.
On the upside, both problems can be fixed with a single tactical strike. The key observation is that many Epoch functions never do anything that generates garbage. A typical Epoch program is comprised of many small, nearly-trivial functions; and if those functions can't possibly create any garbage, there's no purpose checking the GC when they finish.
As a bonus, if a function never needs to call the GC, and never calls any other functions which need the GC, we can turn off register spilling for that function!
The Shootout: Epoch vs. C++
Tonight I implemented a quick little function tag called [nogc]. This tag tells the JITter two things: first, the function cannot generate garbage, and therefore should not invoke the garbage collector; and secondly, because the GC is never invoked from a [nogc] function, register spilling can be disabled in the body of that function.
The results are duly impressive: the raytracer is back to running at ~26.5FPS on my dev machine - a full 4x speedup over the worst-case performance it was exhibiting earlier today.
But it gets even cooler.
Out of sheer curiosity, I wrote a port of the Epoch raytracer in C++. It's not optimal C++, to be fair, and cuts a few corners to make it function more similarly to the Epoch program. But, line for line, it does basically exactly what the Epoch version does.
The C++ port of the raytracer runs at only 23FPS.