Changes between Version 2 and Version 3 of Inlining


Ignore:
Timestamp:
Mar 24, 2009 12:18:04 AM (6 years ago)
Author:
claus
Comment:

add examples/open issues

Legend:

Unmodified
Added
Removed
Modified
  • Inlining

    v2 v3  
    106106       `PEEL/UNROLL` have their own limits, applicable to the whole recursion group, creating intrinsic loop breakers when the counters run out. Every `PEEL` or `UNROLL` action creates calls with smaller counters in the inlined copies, if the calls go into the same recursion.
    107107
     108== Examples ==
     109
     110Examples of hand-unrolled/-peeled loops abound in optimized code. `unroll2.hs` demonstrates the usual worker/wrapper split (+static argument transform), which is part of the standard recommendations for writing code with good performance, and is used all through the standard libraries (here, `PEEL` is more important than `UNROLL`. `unroll0.hs` shows an extreme example with a near trivial loop body, so that the loop administration overhead is relatively expensive (`PEEL` makes the loop body available to the loop combinator worker, `UNROLL` reduces loop administration overhead, and reassociation rules enact constant folding). `unroll1.hs` is an example of loop unrolling activating existing optimization `RULES` (array fusion in this case; note that a [http://www.haskell.org/pipermail/haskell-cafe/2009-March/057295.html fixpoint fusion] could be even more beneficial than the finite unrolling fusion here).
     111
     112== Open Issues ==
     113
     114 - how to handle programs using implicit recursions (via combinators)? One could `PEEL/UNROLL` the definitions of said combinators, but that would give no control at the call sites (like a single setting for all loops). One possibility is to allow `INLINE PEEL/UNROLL` at the call sites, interpreted as making and unfolding __copies__ of the combinator definitions?
     115
     116 - while, eg, `Hugs` just ignores `INLINE` pragmas, recent `GHC` versions have taken to reporting parse errors instead of warnings when encountering unknown forms of `INLINE`
     117
     118 - how does Core-level recursion unfolding interact with backend-level loop transformations (once those come into existence)?
     119
    108120== Further References ==
    109121