Hi Csaba,
Thanks for your presentation, that's a nice high-level overview of what you're up to.
A few thoughts:
- Whole-program optimization sounds great, but also very ambitious, given the amount of code GHC generates today. I'd be amazed to see advances in that area, though, and your >100-module CFA performance incites hope!
- I wonder if going through GRIN results in a more efficient mapping to hardware. I recently found that the code GHC generates is dominated by administrative traffic from and to the heap [1]. I suspect that you can have big wins here if you manage to convey better call stack, heap and alias information to LLVM.
- The Control Analysis+specialisation approach sounds pretty similar to doing Constructor Specialisation [2] for Lambdas (cf. 6.2) if you also inline the function for which you specialise afterwards. I sunk many hours into making that work reliably, fast and without code bloat in the past, to no avail. Frankly, if you can do it in GRIN, I don't see why we couldn't do it in Core. But maybe we can learn from the GRIN implementation afterwards and maybe rethink SpecConstr. Maybe the key is not to inline the function for which we specialise? But then you don't gain that much...
- I follow the Counting Immutable Beans [3] stuff quite closely (Sebastian is a colleague of mine) and hope that it is applicable to Haskell some day. But I think using Perceus, like any purely RC-based memory management scheme, means that you can't have cycles in your heap, so no loopy thunks (such as constant-space `ones = 1:ones`) and mutability. I think that makes a pretty huge difference for many use cases. Sebastian also told me that they have to adapt their solutions to the cycle restriction from time to time, so far always successfully. But it comes at a cost: You have to adapt the code you want to write into a form that works.
I only read the slides, apologies if some of my points were invalidated by something you said.
Keep up the good work!
Cheers,
Sebastian