-- | Graph Coloring. -- This is a generic graph coloring library, abstracted over the type of -- the node keys, nodes and colors. -- module GraphColor ( module GraphBase, module GraphOps, module GraphPpr, colorGraph ) where import GraphBase import GraphOps import GraphPpr import Unique import UniqFM import UniqSet import Outputable import Data.Maybe import Data.List -- | Try to color a graph with this set of colors. -- Uses Chaitin's algorithm to color the graph. -- The graph is scanned for nodes which are deamed 'trivially colorable'. These nodes -- are pushed onto a stack and removed from the graph. -- Once this process is complete the graph can be colored by removing nodes from -- the stack (ie in reverse order) and assigning them colors different to their neighbors. -- colorGraph :: ( Uniquable k, Uniquable cls, Uniquable color , Eq cls, Ord k , Outputable k, Outputable cls, Outputable color) => Bool -- ^ whether to do iterative coalescing -> Int -- ^ how many times we've tried to color this graph so far. -> UniqFM (UniqSet color) -- ^ map of (node class -> set of colors available for this class). -> Triv k cls color -- ^ fn to decide whether a node is trivially colorable. -> (Graph k cls color -> k) -- ^ fn to choose a node to potentially leave uncolored if nothing is trivially colorable. -> Graph k cls color -- ^ the graph to color. -> ( Graph k cls color -- the colored graph. , UniqSet k -- the set of nodes that we couldn't find a color for. , UniqFM k ) -- map of regs (r1 -> r2) that were coaleced -- r1 should be replaced by r2 in the source colorGraph iterative spinCount colors triv spill graph0 = let -- If we're not doing iterative coalescing then do an aggressive coalescing first time -- around and then conservative coalescing for subsequent passes. -- -- Aggressive coalescing is a quick way to get rid of many reg-reg moves. However, if -- there is a lot of register pressure and we do it on every round then it can make the -- graph less colorable and prevent the algorithm from converging in a sensible number -- of cycles. -- (graph_coalesced, kksCoalesce1) = if iterative then (graph0, []) else if spinCount == 0 then coalesceGraph True triv graph0 else coalesceGraph False triv graph0 -- run the scanner to slurp out all the trivially colorable nodes -- (and do coalescing if iterative coalescing is enabled) (ksTriv, ksProblems, kksCoalesce2) = colorScan iterative triv spill graph_coalesced -- If iterative coalescing is enabled, the scanner will coalesce the graph as does its business. -- We need to apply all the coalescences found by the scanner to the original -- graph before doing assignColors. -- -- Because we've got the whole, non-pruned graph here we turn on aggressive coalecing -- to force all the (conservative) coalescences found during scanning. -- (graph_scan_coalesced, _) = mapAccumL (coalesceNodes True triv) graph_coalesced kksCoalesce2 -- color the trivially colorable nodes -- during scanning, keys of triv nodes were added to the front of the list as they were found -- this colors them in the reverse order, as required by the algorithm. (graph_triv, ksNoTriv) = assignColors colors graph_scan_coalesced ksTriv -- try and color the problem nodes -- problem nodes are the ones that were left uncolored because they weren't triv. -- theres a change we can color them here anyway. (graph_prob, ksNoColor) = assignColors colors graph_triv ksProblems -- if the trivially colorable nodes didn't color then something is probably wrong -- with the provided triv function. -- in if not $ null ksNoTriv then pprPanic "colorGraph: trivially colorable nodes didn't color!" -- empty ( empty $$ text "ksTriv = " <> ppr ksTriv $$ text "ksNoTriv = " <> ppr ksNoTriv $$ text "colors = " <> ppr colors $$ empty $$ dotGraph (\_ -> text "white") triv graph_triv) else ( graph_prob , mkUniqSet ksNoColor -- the nodes that didn't color (spills) , if iterative then (listToUFM kksCoalesce2) else (listToUFM kksCoalesce1)) -- | Scan through the conflict graph separating out trivially colorable and -- potentially uncolorable (problem) nodes. -- -- Checking whether a node is trivially colorable or not is a resonably expensive operation, -- so after a triv node is found and removed from the graph it's no good to return to the 'start' -- of the graph and recheck a bunch of nodes that will probably still be non-trivially colorable. -- -- To ward against this, during each pass through the graph we collect up a list of triv nodes -- that were found, and only remove them once we've finished the pass. The more nodes we can delete -- at once the more likely it is that nodes we've already checked will become trivially colorable -- for the next pass. -- -- TODO: add work lists to finding triv nodes is easier. -- If we've just scanned the graph, and removed triv nodes, then the only -- nodes that we need to rescan are the ones we've removed edges from. colorScan :: ( Uniquable k, Uniquable cls, Uniquable color , Ord k, Eq cls , Outputable k, Outputable cls) => Bool -- ^ whether to do iterative coalescing -> Triv k cls color -- ^ fn to decide whether a node is trivially colorable -> (Graph k cls color -> k) -- ^ fn to choose a node to potentially leave uncolored if nothing is trivially colorable. -> Graph k cls color -- ^ the graph to scan -> ([k], [k], [(k, k)]) -- triv colorable nodes, problem nodes, pairs of nodes to coalesce colorScan iterative triv spill graph = colorScan_spin iterative triv spill graph [] [] [] colorScan_spin :: ( Uniquable k, Uniquable cls, Uniquable color , Ord k, Eq cls , Outputable k, Outputable cls) => Bool -> Triv k cls color -> (Graph k cls color -> k) -> Graph k cls color -> [k] -> [k] -> [(k, k)] -> ([k], [k], [(k, k)]) colorScan_spin iterative triv spill graph ksTriv ksSpill kksCoalesce -- if the graph is empty then we're done | isNullUFM $ graphMap graph = (ksTriv, ksSpill, reverse kksCoalesce) -- Simplify: -- Look for trivially colorable nodes. -- If we can find some then remove them from the graph and go back for more. -- | nsTrivFound@(_:_) <- scanGraph (\node -> triv (nodeClass node) (nodeConflicts node) (nodeExclusions node) -- for iterative coalescing we only want non-move related -- nodes here && (not iterative || isEmptyUniqSet (nodeCoalesce node))) $ graph , ksTrivFound <- map nodeId nsTrivFound , graph2 <- foldr (\k g -> let Just g' = delNode k g in g') graph ksTrivFound = colorScan_spin iterative triv spill graph2 (ksTrivFound ++ ksTriv) ksSpill kksCoalesce -- Coalesce: -- If we're doing iterative coalescing and no triv nodes are available -- then it's time for a coalescing pass. | iterative = case coalesceGraph False triv graph of -- we were able to coalesce something -- go back to Simplify and see if this frees up more nodes to be trivially colorable. (graph2, kksCoalesceFound @(_:_)) -> colorScan_spin iterative triv spill graph2 ksTriv ksSpill (reverse kksCoalesceFound ++ kksCoalesce) -- Freeze: -- nothing could be coalesced (or was triv), -- time to choose a node to freeze and give up on ever coalescing it. (graph2, []) -> case freezeOneInGraph graph2 of -- we were able to freeze something -- hopefully this will free up something for Simplify (graph3, True) -> colorScan_spin iterative triv spill graph3 ksTriv ksSpill kksCoalesce -- we couldn't find something to freeze either -- time for a spill (graph3, False) -> colorScan_spill iterative triv spill graph3 ksTriv ksSpill kksCoalesce -- spill time | otherwise = colorScan_spill iterative triv spill graph ksTriv ksSpill kksCoalesce -- Select: -- we couldn't find any triv nodes or things to freeze or coalesce, -- and the graph isn't empty yet.. We'll have to choose a spill -- candidate and leave it uncolored. -- colorScan_spill :: ( Uniquable k, Uniquable cls, Uniquable color , Ord k, Eq cls , Outputable k, Outputable cls) => Bool -> Triv k cls color -> (Graph k cls color -> k) -> Graph k cls color -> [k] -> [k] -> [(k, k)] -> ([k], [k], [(k, k)]) colorScan_spill iterative triv spill graph ksTriv ksSpill kksCoalesce = let kSpill = spill graph Just graph' = delNode kSpill graph in colorScan_spin iterative triv spill graph' ksTriv (kSpill : ksSpill) kksCoalesce -- | Try to assign a color to all these nodes. assignColors :: ( Uniquable k, Uniquable cls, Uniquable color , Outputable cls) => UniqFM (UniqSet color) -- ^ map of (node class -> set of colors available for this class). -> Graph k cls color -- ^ the graph -> [k] -- ^ nodes to assign a color to. -> ( Graph k cls color -- the colored graph , [k]) -- the nodes that didn't color. assignColors colors graph ks = assignColors' colors graph [] ks where assignColors' _ graph prob [] = (graph, prob) assignColors' colors graph prob (k:ks) = case assignColor colors k graph of -- couldn't color this node Nothing -> assignColors' colors graph (k : prob) ks -- this node colored ok, so do the rest Just graph' -> assignColors' colors graph' prob ks assignColor colors u graph | Just c <- selectColor colors graph u = Just (setColor u c graph) | otherwise = Nothing -- | Select a color for a certain node -- taking into account preferences, neighbors and exclusions. -- returns Nothing if no color can be assigned to this node. -- selectColor :: ( Uniquable k, Uniquable cls, Uniquable color , Outputable cls) => UniqFM (UniqSet color) -- ^ map of (node class -> set of colors available for this class). -> Graph k cls color -- ^ the graph -> k -- ^ key of the node to select a color for. -> Maybe color selectColor colors graph u = let -- lookup the node Just node = lookupNode graph u -- lookup the available colors for the class of this node. colors_avail = case lookupUFM colors (nodeClass node) of Nothing -> pprPanic "selectColor: no colors available for class " (ppr (nodeClass node)) Just cs -> cs -- find colors we can't use because they're already being used -- by a node that conflicts with this one. Just nsConflicts = sequence $ map (lookupNode graph) $ nonDetEltsUniqSet $ nodeConflicts node -- See Note [Unique Determinism and code generation] colors_conflict = mkUniqSet $ catMaybes $ map nodeColor nsConflicts -- the prefs of our neighbors colors_neighbor_prefs = mkUniqSet $ concat $ map nodePreference nsConflicts -- colors that are still valid for us colors_ok_ex = minusUniqSet colors_avail (nodeExclusions node) colors_ok = minusUniqSet colors_ok_ex colors_conflict -- the colors that we prefer, and are still ok colors_ok_pref = intersectUniqSets (mkUniqSet $ nodePreference node) colors_ok -- the colors that we could choose while being nice to our neighbors colors_ok_nice = minusUniqSet colors_ok colors_neighbor_prefs -- the best of all possible worlds.. colors_ok_pref_nice = intersectUniqSets colors_ok_nice colors_ok_pref -- make the decision chooseColor -- everyone is happy, yay! | not $ isEmptyUniqSet colors_ok_pref_nice , c : _ <- filter (\x -> elementOfUniqSet x colors_ok_pref_nice) (nodePreference node) = Just c -- we've got one of our preferences | not $ isEmptyUniqSet colors_ok_pref , c : _ <- filter (\x -> elementOfUniqSet x colors_ok_pref) (nodePreference node) = Just c -- it wasn't a preference, but it was still ok | not $ isEmptyUniqSet colors_ok , c : _ <- nonDetEltsUniqSet colors_ok -- See Note [Unique Determinism and code generation] = Just c -- no colors were available for us this time. -- looks like we're going around the loop again.. | otherwise = Nothing in chooseColor