|Version 27 (modified by nomeata, 2 years ago) (diff)|
This is nomeata’s notepad about the nested CPR information:
- #1600 Main tickets where I mention progress.
Tickets with stuff that would make nested CPR better:
- #8598 CPR after IO (partly done)
- Everything in source:testsuite/tests/stranal/sigs/
- Paper-Writeup of CPR
- Shouldn’t nested CPR help a lot with Complex-heavy code? Is there something in nofib?
- Try passing CPR information from the scrunitee to the pattern variables. For that: Reverse flow of analysis for complex scrunitees (for simple, we want the demand coming from the body, for complex, this is not so important.)
- Use ticky-profiling to learn more about the effects of nested CPR.
- Look at DmdAnal-related [SLPJ-Tickets] and see which ones are affected by nested-cpr.
- Do not destroy join points or improve the code genrator (see below).
- Can we make sure more stuff gets the Converging flag, e.g. after a case of an unboxed value? Should case binders get the Converging flag? What about pattern match variables in strict data constructors? Unboxed values? See below.
Degradation exploration and explanation
At one point, I thought that a major contributor to increased allocations is nested-CPR’ing things returning String, causing them to return (# Char#, String #). But removing the CPR information from C# calls has zero effect on the allocations, both on master and on nested-cpr. It had very small (positive) effect on code size. Will have to look at Core...
Here are some case studies with extensive commenting of steps and results:
And here a summary of the problems identified, and solution attempts
- CPR kill join-points, because the wrapper does not completely cancel with anything else.
- Detecting join-points at the position of its binding is not enough.
- A recursive function can have a CPR-beneficial recursive call that makes CPR worthwhile, even if it does not help at the initial call. But it is also not unlikely that the recursive call is a tail-call, and CPR-ing has zero effect on that. Then it all depends on the external call.
Nested CPR is only sound if we know that the nested values are known to converge for sure. The semantics is clear: If f has CPR <...>m(tm(),), then in the body of case f x of (a,b), when entering a, we are guaranteed termination.
What is the semantics of an outer t? Given f with CPR <L>tm() and g with CPR <S>tm()? Does the latter even make sense? If so, should f undefined have CPR m() or tm()? Three possibilities:
- The convergence information a function is what holds if its strictness annotations are fulfilled: So if g x has tm() if x has t (possibly because it has previously been evaluated by the caller), otherwise m(). f x always has m () (presumably because x is _never_ entered when evaluating f.
- The convergence information a function is what holds always. This would in effect prevent <S>tm() from happening.
- The convergence information always holds, but special care is taken for unlifted types: I#, like any function expecting an unlifted parameter or free variable, would get <S>tm(). (For unlifted types, <L> and <S> are identical. One might turn that into a third way <#>, but unless there is more use to that than just clarification, we do not do that).
Clearly, 1. and 3. hold strictly more information than 2.: Under variant 2, <S>tm() would not occur, while the other variants allow that. Also, under 2, I# would not be known to terminate for sure, as it is strict. This would destroy any hope for nested CPR for things like (Int, Int).
I worked on 1, but it turned out to be too complicated. Notes at AdvancedConverges. So I’ll proceed with 3. now.
CPR can kill join points. Attempts to mitigate that:
Idea to fix this, and possibly more general benefits: http://www.haskell.org/pipermail/ghc-devs/2013-December/003481.html; prototype in branch wip/common-context.
- On its own, improvements are present but very small: http://www.haskell.org/pipermail/ghc-devs/2013-December/003500.html
- Enabling CPR for sum types in non-top-level-bindings (which is currently disabled due to worries abut lost join points) yields mixed results (min -3.8%, mean -0.0%, max 3.4%).
- Enabling sum types inside nested CPR: Also yields mixed, not very promising results (-6.9% / +0.0% / +11.3%).
Alternative: Detect join points during dmdAnal and make sure that their CPR info is not greater than that of the expression they are a join-point for. Would also fix #5075, see 5075#comment:19 for benchmark numbers.
- On its own, no changes.
- Enabling CPR for sumtypes: (min -3.8%, mean -0.0%, max 1.7%) (slightly better than with Common Context)
- Enabling sum types inside nested CPR: TBD
Unfortunately, naive approaches are not possible: We need to know if j is a joint point not only for let j = .. in ... j .., but also for expressions further out. Not nice.
Improvement to the code generator
It seems feasible to make the code generate generate better code for local functions that are not quite join-points any more, by jumping, passing both a continuation and a stack delta to the live variables. To be investigated.
Late lambda lifting
Might also help. Need to see if his branch can be merged onto master. (But I like the code generator idea better.)
- Should runSTRep be inlined (see ticket:1600#comment:34)?
- Can we use Terminates CPR information to eagerly evaluate thunks? Yes, and there is a small gain there: #8655
- But why no allocation change? Understand this better!
- Can we statically and/or dynamically count the number of thunks, and the number of CBV’ed thunks?
- Why is cacheprof not deterministic? (→ #8611)
- What became of Simon’s better-ho-cardinality branch? See better-ho-cardinality.
- Try vtunes to get better numbers.