184.108.40.206 Mar 2014
- Remove UTF8 character from hoopl.cabal to workaround issue in GHC 7.8.1’s build system
220.127.116.11 Mar 2014
Bundled with GHC 7.8.1
Update to Cabal format 1.10
Lots of API changes; mainly a new API for working with Blocks
Summary of refactorings:
Blockdatatype and all the operations on Blocks. It seemed like a good idea to collect all this stuff together in one place.
Compiler.Hoopl.Graphnow has the operations on Graphs.
Compiler.Hoopl.GraphUtilare no more; their contents have been moved to other homes. (and a bit of duplicated functionality has been removed).
I removed the newtypes around
Label, these are now just type synonyms. The newtype wrappers were costing some performance in GHC, because in cases like
mapToListthe newtype isn’t optimised away.
This change might be controversial. Feel free to complain.
- Optimisations to the Dataflow algorithms. I’m not actually using this implementation of Dataflow in GHC any more, instead I have a local copy specialised to our monad, for speed. Nevertheless I’ve put some of the optimisations I’m using in the GHC version into the generic library version too.
Summary of API changes:
mapGraphBlocks(was previously called
graphMapBlocks, and not exported)
mapBlock'(strict version of
New API for working with Blocks:
-- ** Predicates on Blocks , isEmptyBlock -- ** Constructing blocks , emptyBlock, blockCons, blockSnoc , blockJoinHead, blockJoinTail, blockJoin, blockJoinAny , blockAppend -- ** Deconstructing blocks , firstNode, lastNode, endNodes , blockSplitHead, blockSplitTail, blockSplit, blockSplitAny -- ** Modifying blocks , replaceFirstNode, replaceLastNode -- ** Converting to and from lists , blockToList, blockFromList -- ** Maps and folds , mapBlock, mapBlock', mapBlock3' , foldBlockNodesF, foldBlockNodesF3 , foldBlockNodesB, foldBlockNodesB3
mapMaybeC(no need: we have
Block constructors are no longer exported (use the block API instead)
blockToNodeList'''(use the block API instead)
lookupBlock, (I don’t know what any of these are for, if they’re still important we could reinstate)
Compiler.Hoopl.GHC is now Compiler.Hoopl.Internals and exports some more stuff.
Label is not a newtype; type Label = Unique
Unique is not a newtype: type Unique = Int (these newtypes were adding overhead)
blockMapNodes3 is now mapBlock3’
Lots of internal refactoring and tidying up
- Re-export runWithFuel as per Justin Bailey
- Uploaded to Hackage by Ian Lynagh; appears to contain updates that use Safe Haskell if GHC >= 7.2 (thanks David Terei)
- Version changed with no record of update; never uploaded to Hackage
- Eliminate warning about nonexhaustive pattern match (thanks Edward Yang)
- Works with GHC 7 (thanks Edward Yang)
cabal sdistnow sort of works (and is added to validate)
- Matches the camera-ready Haskell’10 paper
- Major reorganization per simonpj visit to Tufts 20 April 2010 Collections
- Added function to fold over nodes in graph.
- Pointed type replaces WithTop and WithBot, which are now synonyms.
The type of AGraph is now abstract. It is now recommended to create AGraphs with just three functions:
<*> concatenation |*><*| splicing at a closed point addBlocks add C/C blocks out of line
There are new utility functions in modules Util and XUtil, all exported by Compiler.Hoopl. Here’s a selection:
-- | A utility function so that a transfer function for a first -- node can be given just a fact; we handle the lookup. This -- function is planned to be made obsolete by changes in the dataflow -- interface. firstXfer :: Edges n => (n C O -> f -> f) -> (n C O -> FactBase f -> f) firstXfer xfer n fb = xfer n $ fromJust $ lookupFact fb $ entryLabel n -- | This utility function handles a common case in which a transfer function -- produces a single fact out of a last node, which is then distributed -- over the outgoing edges. distributeXfer :: Edges n => (n O C -> f -> f) -> (n O C -> f -> FactBase f) distributeXfer xfer n f = mkFactBase [ (l, xfer n f) | l <- successors n ] -- | This utility function handles a common case in which a transfer function -- for a last node takes the incoming fact unchanged and simply distributes -- that fact over the outgoing edges. distributeFact :: Edges n => n O C -> f -> FactBase f -- | Function 'graphMapBlocks' enables a change of representation of blocks, -- nodes, or both. It lifts a polymorphic block transform into a polymorphic -- graph transform. When the block representation stabilizes, a similar -- function should be provided for blocks. graphMapBlocks :: forall block n block' n' e x . (forall e x . block n e x -> block' n' e x) -> (Graph' block n e x -> Graph' block' n' e x) postorder_dfs :: Edges (block n) => Graph' block n O x -> [block n C C]
There are quite a few other variations related to depth-first traversal.
There is a new module Compiler.Hoopl.Pointed, which uses GADTs to
enable you to add a Top or Bot element to a lattice, or both, all
using the same type. (Types
WithTop in XUtil,
exported by Compiler.Hoopl, do similar jobs, but I think they are
inferior. Your opinion is solicited.)
I added a function showGraph to print Graphs. Right now it requires a polymorphic node-showing function as argument. When we change the Block representation we may get enough static type information that we can simply have an instance declaration for
instance (Show n C O, Show n O O, Show n O C) => Show (Graph n e x)
At present, I don’t see how to achieve such a declaration.
John added new functions
debugBwdJoins to extend
passes with debugging information.
We added primed versions of the analyzeAndRewrite functions, which operate on Graph, not Body. Recommended, at least for first-timers.
Not all maps keyed by Label are FactBases, so there is now a new set of names of functions to manipulate LabelMaps that do not contain dataflow facts.
Split pass and rewrite-function combinators into Compiler.Hoopl.Combinators.
Changed order of blocks for forward and backward analysis. These changes have not been tested, because we don’t have a true regression suite yet.
Graph and Body types now have more polymorphic variants Graph’ and Body’.
Lots of experiments with zippers.
ForwardTransfer will become an abstract type, and clients will have two ways to create ForwardTransfers: as now, with a single, polymorphic transfer function; or with a triple of monomorphic transfer functions. The implementation will use monomorphic functions, which will enable more useful combinators on passes, including adding more debugging information and enabling us to combine passes.
Perhaps we should provide splicing and
addBlocks on Graph?
Change of block representation to have three monomorphic unit constructors and one polymorphic concatenation constructor.
Graph body to be represented by a finite map; add functions to check for duplicate labels.
- Added a bunch of zipper experiments.
- Existing clients should not be affected.
- More expressive debugging support
- retract arfGraph and normalization; export analyzeAndRewriterFwd’
- Expose arfGraph and normalization functions
- Debugging support
- Rationalized AGraph splicing functions.
- Restrict clients so they see much less, including hiding the value constructors for Body.