2 % (c) The AQUA Project, Glasgow University, 1994-1998
4 \section[UniqFM]{Specialised finite maps, for things with @Uniques@}
6 Based on @FiniteMaps@ (as you would expect).
8 Basically, the things need to be in class @Uniquable@, and we use the
9 @getUnique@ method to grab their @Uniques@.
11 (A similar thing to @UniqSet@, as opposed to @Set@.)
15 UniqFM, -- abstract type
23 addListToUFM,addListToUFM_C,
25 addListToUFM_Directly,
41 lookupUFM, lookupUFM_Directly,
42 lookupWithDefaultUFM, lookupWithDefaultUFM_Directly,
47 #include "HsVersions.h"
49 import {-# SOURCE #-} Name ( Name )
51 import Unique ( Uniquable(..), Unique, u2i, mkUniqueGrimily )
53 import GlaExts -- Lots of Int# operations
56 #if ! OMIT_NATIVE_CODEGEN
59 #define IF_NCG(a) {--}
63 %************************************************************************
65 \subsection{The @UniqFM@ type, and signatures for the functions}
67 %************************************************************************
69 We use @FiniteMaps@, with a (@getUnique@-able) @Unique@ as ``key''.
72 emptyUFM :: UniqFM elt
73 isNullUFM :: UniqFM elt -> Bool
74 unitUFM :: Uniquable key => key -> elt -> UniqFM elt
75 unitDirectlyUFM -- got the Unique already
76 :: Unique -> elt -> UniqFM elt
77 listToUFM :: Uniquable key => [(key,elt)] -> UniqFM elt
79 :: [(Unique, elt)] -> UniqFM elt
81 addToUFM :: Uniquable key => UniqFM elt -> key -> elt -> UniqFM elt
82 addListToUFM :: Uniquable key => UniqFM elt -> [(key,elt)] -> UniqFM elt
84 :: UniqFM elt -> Unique -> elt -> UniqFM elt
86 addToUFM_C :: Uniquable key => (elt -> elt -> elt) -- old -> new -> result
89 -> UniqFM elt -- result
91 addListToUFM_C :: Uniquable key => (elt -> elt -> elt)
92 -> UniqFM elt -> [(key,elt)]
95 delFromUFM :: Uniquable key => UniqFM elt -> key -> UniqFM elt
96 delListFromUFM :: Uniquable key => UniqFM elt -> [key] -> UniqFM elt
97 delFromUFM_Directly :: UniqFM elt -> Unique -> UniqFM elt
99 plusUFM :: UniqFM elt -> UniqFM elt -> UniqFM elt
101 plusUFM_C :: (elt -> elt -> elt)
102 -> UniqFM elt -> UniqFM elt -> UniqFM elt
104 minusUFM :: UniqFM elt -> UniqFM elt -> UniqFM elt
106 intersectUFM :: UniqFM elt -> UniqFM elt -> UniqFM elt
107 intersectUFM_C :: (elt -> elt -> elt)
108 -> UniqFM elt -> UniqFM elt -> UniqFM elt
109 foldUFM :: (elt -> a -> a) -> a -> UniqFM elt -> a
110 mapUFM :: (elt1 -> elt2) -> UniqFM elt1 -> UniqFM elt2
111 filterUFM :: (elt -> Bool) -> UniqFM elt -> UniqFM elt
113 sizeUFM :: UniqFM elt -> Int
114 hashUFM :: UniqFM elt -> Int
115 elemUFM :: Uniquable key => key -> UniqFM elt -> Bool
117 lookupUFM :: Uniquable key => UniqFM elt -> key -> Maybe elt
118 lookupUFM_Directly -- when you've got the Unique already
119 :: UniqFM elt -> Unique -> Maybe elt
121 :: Uniquable key => UniqFM elt -> elt -> key -> elt
122 lookupWithDefaultUFM_Directly
123 :: UniqFM elt -> elt -> Unique -> elt
125 keysUFM :: UniqFM elt -> [Int] -- Get the keys
126 eltsUFM :: UniqFM elt -> [elt]
127 ufmToList :: UniqFM elt -> [(Unique, elt)]
130 %************************************************************************
132 \subsection{The @IdFinMap@ and @TyVarFinMap@ specialisations for Ids/TyVars}
134 %************************************************************************
137 -- Turn off for now, these need to be updated (SDM 4/98)
140 #ifdef __GLASGOW_HASKELL__
141 -- I don't think HBC was too happy about this (WDP 94/10)
144 addListToUFM :: UniqFM elt -> [(Name, elt)] -> UniqFM elt
147 addListToUFM_C :: (elt -> elt -> elt) -> UniqFM elt -> [(Name, elt)] -> UniqFM elt
150 addToUFM :: UniqFM elt -> Unique -> elt -> UniqFM elt
153 listToUFM :: [(Unique, elt)] -> UniqFM elt
156 lookupUFM :: UniqFM elt -> Name -> Maybe elt
157 , UniqFM elt -> Unique -> Maybe elt
160 #endif {- __GLASGOW_HASKELL__ -}
164 %************************************************************************
166 \subsection{Andy Gill's underlying @UniqFM@ machinery}
168 %************************************************************************
170 ``Uniq Finite maps'' are the heart and soul of the compiler's
171 lookup-tables/environments. Important stuff! It works well with
172 Dense and Sparse ranges.
173 Both @Uq@ Finite maps and @Hash@ Finite Maps
174 are built ontop of Int Finite Maps.
176 This code is explained in the paper:
178 A Gill, S Peyton Jones, B O'Sullivan, W Partain and Aqua Friends
179 "A Cheap balancing act that grows on a tree"
180 Glasgow FP Workshop, Sep 1994, pp??-??
183 %************************************************************************
185 \subsubsection{The @UniqFM@ type, and signatures for the functions}
187 %************************************************************************
189 @UniqFM a@ is a mapping from Unique to a.
191 First, the DataType itself; which is either a Node, a Leaf, or an Empty.
196 | LeafUFM FAST_INT ele
197 | NodeUFM FAST_INT -- the switching
198 FAST_INT -- the delta
203 -- for debugging only :-)
204 instance Outputable (UniqFM a) where
205 ppr(NodeUFM a b t1 t2) =
206 sep [text "NodeUFM " <+> int IBOX(a) <+> int IBOX(b),
207 nest 1 (parens (ppr t1)),
208 nest 1 (parens (ppr t2))]
209 ppr (LeafUFM x a) = text "LeafUFM " <+> int IBOX(x)
210 ppr (EmptyUFM) = empty
212 -- and when not debugging the package itself...
213 instance Outputable a => Outputable (UniqFM a) where
214 ppr ufm = ppr (ufmToList ufm)
217 %************************************************************************
219 \subsubsection{The @UniqFM@ functions}
221 %************************************************************************
223 First the ways of building a UniqFM.
227 unitUFM key elt = mkLeafUFM (u2i (getUnique key)) elt
228 unitDirectlyUFM key elt = mkLeafUFM (u2i key) elt
230 listToUFM key_elt_pairs
231 = addListToUFM_C use_snd EmptyUFM key_elt_pairs
233 listToUFM_Directly uniq_elt_pairs
234 = addListToUFM_directly_C use_snd EmptyUFM uniq_elt_pairs
237 Now ways of adding things to UniqFMs.
239 There is an alternative version of @addListToUFM_C@, that uses @plusUFM@,
240 but the semantics of this operation demands a linear insertion;
241 perhaps the version without the combinator function
242 could be optimised using it.
245 addToUFM fm key elt = addToUFM_C use_snd fm key elt
247 addToUFM_Directly fm u elt = insert_ele use_snd fm (u2i u) elt
249 addToUFM_C combiner fm key elt
250 = insert_ele combiner fm (u2i (getUnique key)) elt
252 addListToUFM fm key_elt_pairs = addListToUFM_C use_snd fm key_elt_pairs
253 addListToUFM_Directly fm uniq_elt_pairs = addListToUFM_directly_C use_snd fm uniq_elt_pairs
255 addListToUFM_C combiner fm key_elt_pairs
256 = foldl (\ fm (k, e) -> insert_ele combiner fm (u2i (getUnique k)) e)
259 addListToUFM_directly_C combiner fm uniq_elt_pairs
260 = foldl (\ fm (k, e) -> insert_ele combiner fm (u2i k) e)
264 Now ways of removing things from UniqFM.
267 delListFromUFM fm lst = foldl delFromUFM fm lst
269 delFromUFM fm key = delete fm (u2i (getUnique key))
270 delFromUFM_Directly fm u = delete fm (u2i u)
272 delete EmptyUFM _ = EmptyUFM
273 delete fm key = del_ele fm
275 del_ele :: UniqFM a -> UniqFM a
277 del_ele lf@(LeafUFM j _)
278 | j _EQ_ key = EmptyUFM
279 | otherwise = lf -- no delete!
281 del_ele nd@(NodeUFM j p t1 t2)
283 = mkSLNodeUFM (NodeUFMData j p) (del_ele t1) t2
285 = mkLSNodeUFM (NodeUFMData j p) t1 (del_ele t2)
287 del_ele _ = panic "Found EmptyUFM FM when rec-deleting"
290 Now ways of adding two UniqFM's together.
293 plusUFM tr1 tr2 = plusUFM_C use_snd tr1 tr2
295 plusUFM_C f EmptyUFM tr = tr
296 plusUFM_C f tr EmptyUFM = tr
297 plusUFM_C f fm1 fm2 = mix_trees fm1 fm2
299 mix_trees (LeafUFM i a) t2 = insert_ele (flip f) t2 i a
300 mix_trees t1 (LeafUFM i a) = insert_ele f t1 i a
302 mix_trees left_t@(NodeUFM j p t1 t2) right_t@(NodeUFM j' p' t1' t2')
304 (ask_about_common_ancestor
308 -- Given a disjoint j,j' (p >^ p' && p' >^ p):
312 -- t1 t2 t1' t2' j j'
317 mix_branches (NewRoot nd False)
318 = mkLLNodeUFM nd left_t right_t
319 mix_branches (NewRoot nd True)
320 = mkLLNodeUFM nd right_t left_t
326 -- t1 t2 t1' t2' t1 + t1' t2 + t2'
328 mix_branches (SameRoot)
329 = mkSSNodeUFM (NodeUFMData j p)
332 -- Now the 4 different other ways; all like this:
334 -- Given j >^ j' (and, say, j > j')
338 -- t1 t2 t1' t2' t1 t2 + j'
341 mix_branches (LeftRoot Leftt) -- | trace "LL" True
344 (mix_trees t1 right_t)
347 mix_branches (LeftRoot Rightt) -- | trace "LR" True
351 (mix_trees t2 right_t)
353 mix_branches (RightRoot Leftt) -- | trace "RL" True
356 (mix_trees left_t t1')
359 mix_branches (RightRoot Rightt) -- | trace "RR" True
363 (mix_trees left_t t2')
365 mix_trees _ _ = panic "EmptyUFM found when inserting into plusInt"
368 And ways of subtracting them. First the base cases,
369 then the full D&C approach.
372 minusUFM EmptyUFM _ = EmptyUFM
373 minusUFM t1 EmptyUFM = t1
374 minusUFM fm1 fm2 = minus_trees fm1 fm2
377 -- Notice the asymetry of subtraction
379 minus_trees lf@(LeafUFM i a) t2 =
384 minus_trees t1 (LeafUFM i _) = delete t1 i
386 minus_trees left_t@(NodeUFM j p t1 t2) right_t@(NodeUFM j' p' t1' t2')
388 (ask_about_common_ancestor
392 -- Given a disjoint j,j' (p >^ p' && p' >^ p):
396 -- t1 t2 t1' t2' t1 t2
401 minus_branches (NewRoot nd _) = left_t
407 -- t1 t2 t1' t2' t1 + t1' t2 + t2'
409 minus_branches (SameRoot)
410 = mkSSNodeUFM (NodeUFMData j p)
413 -- Now the 4 different other ways; all like this:
414 -- again, with asymatry
417 -- The left is above the right
419 minus_branches (LeftRoot Leftt)
422 (minus_trees t1 right_t)
424 minus_branches (LeftRoot Rightt)
428 (minus_trees t2 right_t)
431 -- The right is above the left
433 minus_branches (RightRoot Leftt)
434 = minus_trees left_t t1'
435 minus_branches (RightRoot Rightt)
436 = minus_trees left_t t2'
438 minus_trees _ _ = panic "EmptyUFM found when insering into plusInt"
441 And taking the intersection of two UniqFM's.
444 intersectUFM t1 t2 = intersectUFM_C use_snd t1 t2
446 intersectUFM_C f EmptyUFM _ = EmptyUFM
447 intersectUFM_C f _ EmptyUFM = EmptyUFM
448 intersectUFM_C f fm1 fm2 = intersect_trees fm1 fm2
450 intersect_trees (LeafUFM i a) t2 =
453 Just b -> mkLeafUFM i (f a b)
455 intersect_trees t1 (LeafUFM i a) =
458 Just b -> mkLeafUFM i (f b a)
460 intersect_trees left_t@(NodeUFM j p t1 t2) right_t@(NodeUFM j' p' t1' t2')
462 (ask_about_common_ancestor
466 -- Given a disjoint j,j' (p >^ p' && p' >^ p):
469 -- / \ + / \ ==> EmptyUFM
474 intersect_branches (NewRoot nd _) = EmptyUFM
480 -- t1 t2 t1' t2' t1 x t1' t2 x t2'
482 intersect_branches (SameRoot)
483 = mkSSNodeUFM (NodeUFMData j p)
484 (intersect_trees t1 t1')
485 (intersect_trees t2 t2')
486 -- Now the 4 different other ways; all like this:
488 -- Given j >^ j' (and, say, j > j')
492 -- t1 t2 t1' t2' t1' t2'
494 -- This does cut down the search space quite a bit.
496 intersect_branches (LeftRoot Leftt)
497 = intersect_trees t1 right_t
498 intersect_branches (LeftRoot Rightt)
499 = intersect_trees t2 right_t
500 intersect_branches (RightRoot Leftt)
501 = intersect_trees left_t t1'
502 intersect_branches (RightRoot Rightt)
503 = intersect_trees left_t t2'
505 intersect_trees x y = panic ("EmptyUFM found when intersecting trees")
508 Now the usual set of `collection' operators, like map, fold, etc.
511 foldUFM f a (NodeUFM _ _ t1 t2) = foldUFM f (foldUFM f a t2) t1
512 foldUFM f a (LeafUFM _ obj) = f obj a
513 foldUFM f a EmptyUFM = a
517 mapUFM fn EmptyUFM = EmptyUFM
518 mapUFM fn fm = map_tree fn fm
520 filterUFM fn EmptyUFM = EmptyUFM
521 filterUFM fn fm = filter_tree fn fm
524 Note, this takes a long time, O(n), but
525 because we dont want to do this very often, we put up with this.
526 O'rable, but how often do we look at the size of
531 sizeUFM (NodeUFM _ _ t1 t2) = sizeUFM t1 + sizeUFM t2
532 sizeUFM (LeafUFM _ _) = 1
534 isNullUFM EmptyUFM = True
537 -- hashing is used in VarSet.uniqAway, and should be fast
538 -- We use a cheap and cheerful method for now
540 hashUFM (NodeUFM n _ _ _) = IBOX(n)
541 hashUFM (LeafUFM n _) = IBOX(n)
544 looking up in a hurry is the {\em whole point} of this binary tree lark.
545 Lookup up a binary tree is easy (and fast).
548 elemUFM key fm = case lookUp fm (u2i (getUnique key)) of
552 lookupUFM fm key = lookUp fm (u2i (getUnique key))
553 lookupUFM_Directly fm key = lookUp fm (u2i key)
555 lookupWithDefaultUFM fm deflt key
556 = case lookUp fm (u2i (getUnique key)) of
560 lookupWithDefaultUFM_Directly fm deflt key
561 = case lookUp fm (u2i key) of
565 lookUp EmptyUFM _ = Nothing
566 lookUp fm i = lookup_tree fm
568 lookup_tree :: UniqFM a -> Maybe a
570 lookup_tree (LeafUFM j b)
572 | otherwise = Nothing
573 lookup_tree (NodeUFM j p t1 t2)
574 | j _GT_ i = lookup_tree t1
575 | otherwise = lookup_tree t2
577 lookup_tree EmptyUFM = panic "lookup Failed"
580 folds are *wonderful* things.
583 eltsUFM fm = foldUFM (:) [] fm
585 ufmToList fm = fold_tree (\ iu elt rest -> (mkUniqueGrimily iu, elt) : rest) [] fm
587 keysUFM fm = fold_tree (\ iu elt rest -> IBOX(iu) : rest) [] fm
589 fold_tree f a (NodeUFM _ _ t1 t2) = fold_tree f (fold_tree f a t2) t1
590 fold_tree f a (LeafUFM iu obj) = f iu obj a
591 fold_tree f a EmptyUFM = a
594 %************************************************************************
596 \subsubsection{The @UniqFM@ type, and its functions}
598 %************************************************************************
600 You should always use these to build the tree.
601 There are 4 versions of mkNodeUFM, depending on
602 the strictness of the two sub-tree arguments.
603 The strictness is used *both* to prune out
604 empty trees, *and* to improve performance,
605 stoping needless thunks lying around.
606 The rule of thumb (from experence with these trees)
607 is make thunks strict, but data structures lazy.
608 If in doubt, use mkSSNodeUFM, which has the `strongest'
609 functionality, but may do a few needless evaluations.
612 mkLeafUFM :: FAST_INT -> a -> UniqFM a
613 mkLeafUFM i a = LeafUFM i a
615 -- The *ONLY* ways of building a NodeUFM.
617 mkSSNodeUFM (NodeUFMData j p) EmptyUFM t2 = t2
618 mkSSNodeUFM (NodeUFMData j p) t1 EmptyUFM = t1
619 mkSSNodeUFM (NodeUFMData j p) t1 t2
620 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
623 mkSLNodeUFM (NodeUFMData j p) EmptyUFM t2 = t2
624 mkSLNodeUFM (NodeUFMData j p) t1 t2
625 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
628 mkLSNodeUFM (NodeUFMData j p) t1 EmptyUFM = t1
629 mkLSNodeUFM (NodeUFMData j p) t1 t2
630 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
633 mkLLNodeUFM (NodeUFMData j p) t1 t2
634 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
644 correctNodeUFM j p t1 t2
645 = correct (j-p) (j-1) p t1 && correct j ((j-1)+p) p t2
647 correct low high _ (LeafUFM i _)
648 = low <= IBOX(i) && IBOX(i) <= high
649 correct low high above_p (NodeUFM j p _ _)
650 = low <= IBOX(j) && IBOX(j) <= high && above_p > IBOX(p)
651 correct _ _ _ EmptyUFM = panic "EmptyUFM stored inside a tree"
654 Note: doing SAT on this by hand seems to make it worse. Todo: Investigate,
655 and if necessary do $\lambda$ lifting on our functions that are bound.
665 insert_ele f EmptyUFM i new = mkLeafUFM i new
667 insert_ele f (LeafUFM j old) i new
669 mkLLNodeUFM (getCommonNodeUFMData
674 | j _EQ_ i = mkLeafUFM j (f old new)
676 mkLLNodeUFM (getCommonNodeUFMData
682 insert_ele f n@(NodeUFM j p t1 t2) i a
684 = if (i _GE_ (j _SUB_ p))
685 then mkSLNodeUFM (NodeUFMData j p) (insert_ele f t1 i a) t2
686 else mkLLNodeUFM (getCommonNodeUFMData
692 = if (i _LE_ ((j _SUB_ ILIT(1)) _ADD_ p))
693 then mkLSNodeUFM (NodeUFMData j p) t1 (insert_ele f t2 i a)
694 else mkLLNodeUFM (getCommonNodeUFMData
704 map_tree f (NodeUFM j p t1 t2)
705 = mkSSNodeUFM (NodeUFMData j p) (map_tree f t1) (map_tree f t2)
706 map_tree f (LeafUFM i obj)
707 = mkLeafUFM i (f obj)
709 map_tree f _ = panic "map_tree failed"
713 filter_tree f nd@(NodeUFM j p t1 t2)
714 = mkSSNodeUFM (NodeUFMData j p) (filter_tree f t1) (filter_tree f t2)
716 filter_tree f lf@(LeafUFM i obj)
718 | otherwise = EmptyUFM
719 filter_tree f _ = panic "filter_tree failed"
722 %************************************************************************
724 \subsubsection{The @UniqFM@ type, and signatures for the functions}
726 %************************************************************************
730 This is the information that is held inside a NodeUFM, packaged up for
735 = NodeUFMData FAST_INT
739 This is the information used when computing new NodeUFMs.
742 data Side = Leftt | Rightt -- NB: avoid 1.3 names "Left" and "Right"
744 = LeftRoot Side -- which side is the right down ?
745 | RightRoot Side -- which side is the left down ?
746 | SameRoot -- they are the same !
747 | NewRoot NodeUFMData -- here's the new, common, root
748 Bool -- do you need to swap left and right ?
751 This specifies the relationship between NodeUFMData and CalcNodeUFMData.
754 indexToRoot :: FAST_INT -> NodeUFMData
758 l = (ILIT(1) :: FAST_INT)
760 NodeUFMData (((i `shiftR_` l) `shiftL_` l) _ADD_ ILIT(1)) l
762 getCommonNodeUFMData :: NodeUFMData -> NodeUFMData -> NodeUFMData
764 getCommonNodeUFMData (NodeUFMData i p) (NodeUFMData i2 p2)
765 | p _EQ_ p2 = getCommonNodeUFMData_ p j j2
766 | p _LT_ p2 = getCommonNodeUFMData_ p2 (j _QUOT_ (p2 _QUOT_ p)) j2
767 | otherwise = getCommonNodeUFMData_ p j (j2 _QUOT_ (p _QUOT_ p2))
769 l = (ILIT(1) :: FAST_INT)
770 j = i _QUOT_ (p `shiftL_` l)
771 j2 = i2 _QUOT_ (p2 `shiftL_` l)
773 getCommonNodeUFMData_ :: FAST_INT -> FAST_INT -> FAST_INT -> NodeUFMData
775 getCommonNodeUFMData_ p j j_
777 = NodeUFMData (((j `shiftL_` l) _ADD_ l) _MUL_ p) p
779 = getCommonNodeUFMData_ (p `shiftL_` l) (j `shiftR_` l) (j_ `shiftR_` l)
781 ask_about_common_ancestor :: NodeUFMData -> NodeUFMData -> CommonRoot
783 ask_about_common_ancestor x@(NodeUFMData j p) y@(NodeUFMData j2 p2)
784 | j _EQ_ j2 = SameRoot
786 = case getCommonNodeUFMData x y of
787 nd@(NodeUFMData j3 p3)
788 | j3 _EQ_ j -> LeftRoot (decideSide (j _GT_ j2))
789 | j3 _EQ_ j2 -> RightRoot (decideSide (j _LT_ j2))
790 | otherwise -> NewRoot nd (j _GT_ j2)
792 decideSide :: Bool -> Side
793 decideSide True = Leftt
794 decideSide False = Rightt
797 This might be better in Util.lhs ?
800 Now the bit twiddling functions.
802 shiftL_ :: FAST_INT -> FAST_INT -> FAST_INT
803 shiftR_ :: FAST_INT -> FAST_INT -> FAST_INT
805 #if __GLASGOW_HASKELL__
806 {-# INLINE shiftL_ #-}
807 {-# INLINE shiftR_ #-}
808 shiftL_ n p = word2Int#((int2Word# n) `shiftL#` p)
809 shiftR_ n p = word2Int#((int2Word# n) `shiftr` p)
811 shiftr x y = shiftRL# x y
814 shiftL_ n p = n * (2 ^ p)
815 shiftR_ n p = n `quot` (2 ^ p)
821 use_snd :: a -> b -> b