-
Simon Peyton Jones authored
This patch is the result of Ilya Sergey's internship at MSR. It constitutes a thorough overhaul and simplification of the demand analyser. It makes a solid foundation on which we can now build. Main changes are * Instead of having one combined type for Demand, a Demand is now a pair (JointDmd) of - a StrDmd and - an AbsDmd. This allows strictness and absence to be though about quite orthogonally, and greatly reduces brain melt-down. * Similarly in the DmdResult type, it's a pair of - a PureResult (indicating only divergence/non-divergence) - a CPRResult (which deals only with the CPR property * In IdInfo, the strictnessInfo field contains a StrictSig, not a Maybe StrictSig demandInfo field contains a Demand, not a Maybe Demand We don't need Nothing (to indicate no strictness/demand info) any more; topSig/topDmd will do. * Remove "boxity" analysis entirely. This was an attempt to avoid "reboxing", but it added complexity, is extremely ad-hoc, and makes very little difference in practice. * Remove the "unboxing strategy" computation. This was an an attempt to ensure that a worker didn't get zillions of arguments by unboxing big tuples. But in fact removing it DRAMATICALLY reduces allocation in an inner loop of the I/O library (where the threshold argument-count had been set just too low). It's exceptional to have a zillion arguments and I don't think it's worth the complexity, especially since it turned out to have a serious performance hit. * Remove quite a bit of ad-hoc cruft * Move worthSplittingFun, worthSplittingThunk from WorkWrap to Demand. This allows JointDmd to be fully abstract, examined only inside Demand. Everything else really follows from these changes. All of this is really just refactoring, so we don't expect big performance changes, but acutally the numbers look quite good. Here is a full nofib run with some highlights identified: Program Size Allocs Runtime Elapsed TotalMem -------------------------------------------------------------------------------- expert -2.6% -15.5% 0.00 0.00 +0.0% fluid -2.4% -7.1% 0.01 0.01 +0.0% gg -2.5% -28.9% 0.02 0.02 -33.3% integrate -2.6% +3.2% +2.6% +2.6% +0.0% mandel2 -2.6% +4.2% 0.01 0.01 +0.0% nucleic2 -2.0% -16.3% 0.11 0.11 +0.0% para -2.6% -20.0% -11.8% -11.7% +0.0% parser -2.5% -17.9% 0.05 0.05 +0.0% prolog -2.6% -13.0% 0.00 0.00 +0.0% puzzle -2.6% +2.2% +0.8% +0.8% +0.0% sorting -2.6% -35.9% 0.00 0.00 +0.0% treejoin -2.6% -52.2% -9.8% -9.9% +0.0% -------------------------------------------------------------------------------- Min -2.7% -52.2% -11.8% -11.7% -33.3% Max -1.8% +4.2% +10.5% +10.5% +7.7% Geometric Mean -2.5% -2.8% -0.4% -0.5% -0.4% Things to note * Binary sizes are smaller. I don't know why, but it's good. * Allocation is sometiemes a *lot* smaller. I believe that all the big numbers (I checked treejoin, gg, sorting) arise from one place, namely a function GHC.IO.Encoding.UTF8.utf8_decode, which is strict in two Buffers both of which have several arugments. Not w/w'ing both arguments (which is what we did before) has a big effect. So the big win in actually somewhat accidental, gained by removing the "unboxing strategy" code. * A couple of benchmarks allocate slightly more. This turns out to be due to reboxing (integrate). But the biggest increase is mandel2, and *that* turned out also to be a somewhat accidental loss of CSE, and pointed the way to doing better CSE: see Trac #7596. * Runtimes are never very reliable, but seem to improve very slightly. All in all, a good piece of work. Thank you Ilya!
0831a12e