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,
40 lookupUFM, lookupUFM_Directly,
41 lookupWithDefaultUFM, lookupWithDefaultUFM_Directly,
46 #include "HsVersions.h"
48 import {-# SOURCE #-} Name ( Name )
50 import Unique ( Uniquable(..), Unique, u2i, mkUniqueGrimily )
52 import GlaExts -- Lots of Int# operations
54 #if ! OMIT_NATIVE_CODEGEN
57 #define IF_NCG(a) {--}
61 %************************************************************************
63 \subsection{The @UniqFM@ type, and signatures for the functions}
65 %************************************************************************
67 We use @FiniteMaps@, with a (@getUnique@-able) @Unique@ as ``key''.
70 emptyUFM :: UniqFM elt
71 isNullUFM :: UniqFM elt -> Bool
72 unitUFM :: Uniquable key => key -> elt -> UniqFM elt
73 unitDirectlyUFM -- got the Unique already
74 :: Unique -> elt -> UniqFM elt
75 listToUFM :: Uniquable key => [(key,elt)] -> UniqFM elt
77 :: [(Unique, elt)] -> UniqFM elt
79 addToUFM :: Uniquable key => UniqFM elt -> key -> elt -> UniqFM elt
80 addListToUFM :: Uniquable key => UniqFM elt -> [(key,elt)] -> UniqFM elt
82 :: UniqFM elt -> Unique -> elt -> UniqFM elt
84 addToUFM_C :: Uniquable key => (elt -> elt -> elt) -- old -> new -> result
87 -> UniqFM elt -- result
89 addListToUFM_C :: Uniquable key => (elt -> elt -> elt)
90 -> UniqFM elt -> [(key,elt)]
93 delFromUFM :: Uniquable key => UniqFM elt -> key -> UniqFM elt
94 delListFromUFM :: Uniquable key => UniqFM elt -> [key] -> UniqFM elt
95 delFromUFM_Directly :: UniqFM elt -> Unique -> UniqFM elt
97 plusUFM :: UniqFM elt -> UniqFM elt -> UniqFM elt
99 plusUFM_C :: (elt -> elt -> elt)
100 -> UniqFM elt -> UniqFM elt -> UniqFM elt
102 minusUFM :: UniqFM elt -> UniqFM elt -> UniqFM elt
104 intersectUFM :: UniqFM elt -> UniqFM elt -> UniqFM elt
105 intersectUFM_C :: (elt -> elt -> elt)
106 -> UniqFM elt -> UniqFM elt -> UniqFM elt
107 foldUFM :: (elt -> a -> a) -> a -> UniqFM elt -> a
108 mapUFM :: (elt1 -> elt2) -> UniqFM elt1 -> UniqFM elt2
109 filterUFM :: (elt -> Bool) -> UniqFM elt -> UniqFM elt
111 sizeUFM :: UniqFM elt -> Int
112 elemUFM :: Uniquable key => key -> UniqFM elt -> Bool
114 lookupUFM :: Uniquable key => UniqFM elt -> key -> Maybe elt
115 lookupUFM_Directly -- when you've got the Unique already
116 :: UniqFM elt -> Unique -> Maybe elt
118 :: Uniquable key => UniqFM elt -> elt -> key -> elt
119 lookupWithDefaultUFM_Directly
120 :: UniqFM elt -> elt -> Unique -> elt
122 keysUFM :: UniqFM elt -> [Int] -- Get the keys
123 eltsUFM :: UniqFM elt -> [elt]
124 ufmToList :: UniqFM elt -> [(Unique, elt)]
127 %************************************************************************
129 \subsection{The @IdFinMap@ and @TyVarFinMap@ specialisations for Ids/TyVars}
131 %************************************************************************
134 -- Turn off for now, these need to be updated (SDM 4/98)
137 #ifdef __GLASGOW_HASKELL__
138 -- I don't think HBC was too happy about this (WDP 94/10)
141 addListToUFM :: UniqFM elt -> [(Name, elt)] -> UniqFM elt
144 addListToUFM_C :: (elt -> elt -> elt) -> UniqFM elt -> [(Name, elt)] -> UniqFM elt
147 addToUFM :: UniqFM elt -> Unique -> elt -> UniqFM elt
150 listToUFM :: [(Unique, elt)] -> UniqFM elt
153 lookupUFM :: UniqFM elt -> Name -> Maybe elt
154 , UniqFM elt -> Unique -> Maybe elt
157 #endif {- __GLASGOW_HASKELL__ -}
161 %************************************************************************
163 \subsection{Andy Gill's underlying @UniqFM@ machinery}
165 %************************************************************************
167 ``Uniq Finite maps'' are the heart and soul of the compiler's
168 lookup-tables/environments. Important stuff! It works well with
169 Dense and Sparse ranges.
170 Both @Uq@ Finite maps and @Hash@ Finite Maps
171 are built ontop of Int Finite Maps.
173 This code is explained in the paper:
175 A Gill, S Peyton Jones, B O'Sullivan, W Partain and Aqua Friends
176 "A Cheap balancing act that grows on a tree"
177 Glasgow FP Workshop, Sep 1994, pp??-??
180 %************************************************************************
182 \subsubsection{The @UniqFM@ type, and signatures for the functions}
184 %************************************************************************
186 @UniqFM a@ is a mapping from Unique to a.
188 First, the DataType itself; which is either a Node, a Leaf, or an Empty.
193 | LeafUFM FAST_INT ele
194 | NodeUFM FAST_INT -- the switching
195 FAST_INT -- the delta
199 -- for debugging only :-)
201 instance Text (UniqFM a) where
202 showsPrec _ (NodeUFM a b t1 t2) =
203 showString "NodeUFM " . shows (IBOX(a))
204 . showString " " . shows (IBOX(b))
205 . showString " (" . shows t1
206 . showString ") (" . shows t2
208 showsPrec _ (LeafUFM x a) = showString "LeafUFM " . shows (IBOX(x))
209 showsPrec _ (EmptyUFM) = id
213 %************************************************************************
215 \subsubsection{The @UniqFM@ functions}
217 %************************************************************************
219 First the ways of building a UniqFM.
223 unitUFM key elt = mkLeafUFM (u2i (getUnique key)) elt
224 unitDirectlyUFM key elt = mkLeafUFM (u2i key) elt
226 listToUFM key_elt_pairs
227 = addListToUFM_C use_snd EmptyUFM key_elt_pairs
229 listToUFM_Directly uniq_elt_pairs
230 = addListToUFM_directly_C use_snd EmptyUFM uniq_elt_pairs
233 Now ways of adding things to UniqFMs.
235 There is an alternative version of @addListToUFM_C@, that uses @plusUFM@,
236 but the semantics of this operation demands a linear insertion;
237 perhaps the version without the combinator function
238 could be optimised using it.
241 addToUFM fm key elt = addToUFM_C use_snd fm key elt
243 addToUFM_Directly fm u elt = insert_ele use_snd fm (u2i u) elt
245 addToUFM_C combiner fm key elt
246 = insert_ele combiner fm (u2i (getUnique key)) elt
248 addListToUFM fm key_elt_pairs = addListToUFM_C use_snd fm key_elt_pairs
249 addListToUFM_Directly fm uniq_elt_pairs = addListToUFM_directly_C use_snd fm uniq_elt_pairs
251 addListToUFM_C combiner fm key_elt_pairs
252 = foldl (\ fm (k, e) -> insert_ele combiner fm (u2i (getUnique k)) e)
255 addListToUFM_directly_C combiner fm uniq_elt_pairs
256 = foldl (\ fm (k, e) -> insert_ele combiner fm (u2i k) e)
260 Now ways of removing things from UniqFM.
263 delListFromUFM fm lst = foldl delFromUFM fm lst
265 delFromUFM fm key = delete fm (u2i (getUnique key))
266 delFromUFM_Directly fm u = delete fm (u2i u)
268 delete EmptyUFM _ = EmptyUFM
269 delete fm key = del_ele fm
271 del_ele :: UniqFM a -> UniqFM a
273 del_ele lf@(LeafUFM j _)
274 | j _EQ_ key = EmptyUFM
275 | otherwise = lf -- no delete!
277 del_ele nd@(NodeUFM j p t1 t2)
279 = mkSLNodeUFM (NodeUFMData j p) (del_ele t1) t2
281 = mkLSNodeUFM (NodeUFMData j p) t1 (del_ele t2)
283 del_ele _ = panic "Found EmptyUFM FM when rec-deleting"
286 Now ways of adding two UniqFM's together.
289 plusUFM tr1 tr2 = plusUFM_C use_snd tr1 tr2
291 plusUFM_C f EmptyUFM tr = tr
292 plusUFM_C f tr EmptyUFM = tr
293 plusUFM_C f fm1 fm2 = mix_trees fm1 fm2
295 mix_trees (LeafUFM i a) t2 = insert_ele (flip f) t2 i a
296 mix_trees t1 (LeafUFM i a) = insert_ele f t1 i a
298 mix_trees left_t@(NodeUFM j p t1 t2) right_t@(NodeUFM j' p' t1' t2')
300 (ask_about_common_ancestor
304 -- Given a disjoint j,j' (p >^ p' && p' >^ p):
308 -- t1 t2 t1' t2' j j'
313 mix_branches (NewRoot nd False)
314 = mkLLNodeUFM nd left_t right_t
315 mix_branches (NewRoot nd True)
316 = mkLLNodeUFM nd right_t left_t
322 -- t1 t2 t1' t2' t1 + t1' t2 + t2'
324 mix_branches (SameRoot)
325 = mkSSNodeUFM (NodeUFMData j p)
328 -- Now the 4 different other ways; all like this:
330 -- Given j >^ j' (and, say, j > j')
334 -- t1 t2 t1' t2' t1 t2 + j'
337 mix_branches (LeftRoot Leftt) -- | trace "LL" True
340 (mix_trees t1 right_t)
343 mix_branches (LeftRoot Rightt) -- | trace "LR" True
347 (mix_trees t2 right_t)
349 mix_branches (RightRoot Leftt) -- | trace "RL" True
352 (mix_trees left_t t1')
355 mix_branches (RightRoot Rightt) -- | trace "RR" True
359 (mix_trees left_t t2')
361 mix_trees _ _ = panic "EmptyUFM found when inserting into plusInt"
364 And ways of subtracting them. First the base cases,
365 then the full D&C approach.
368 minusUFM EmptyUFM _ = EmptyUFM
369 minusUFM t1 EmptyUFM = t1
370 minusUFM fm1 fm2 = minus_trees fm1 fm2
373 -- Notice the asymetry of subtraction
375 minus_trees lf@(LeafUFM i a) t2 =
380 minus_trees t1 (LeafUFM i _) = delete t1 i
382 minus_trees left_t@(NodeUFM j p t1 t2) right_t@(NodeUFM j' p' t1' t2')
384 (ask_about_common_ancestor
388 -- Given a disjoint j,j' (p >^ p' && p' >^ p):
392 -- t1 t2 t1' t2' t1 t2
397 minus_branches (NewRoot nd _) = left_t
403 -- t1 t2 t1' t2' t1 + t1' t2 + t2'
405 minus_branches (SameRoot)
406 = mkSSNodeUFM (NodeUFMData j p)
409 -- Now the 4 different other ways; all like this:
410 -- again, with asymatry
413 -- The left is above the right
415 minus_branches (LeftRoot Leftt)
418 (minus_trees t1 right_t)
420 minus_branches (LeftRoot Rightt)
424 (minus_trees t2 right_t)
427 -- The right is above the left
429 minus_branches (RightRoot Leftt)
430 = minus_trees left_t t1'
431 minus_branches (RightRoot Rightt)
432 = minus_trees left_t t2'
434 minus_trees _ _ = panic "EmptyUFM found when insering into plusInt"
437 And taking the intersection of two UniqFM's.
440 intersectUFM t1 t2 = intersectUFM_C use_snd t1 t2
442 intersectUFM_C f EmptyUFM _ = EmptyUFM
443 intersectUFM_C f _ EmptyUFM = EmptyUFM
444 intersectUFM_C f fm1 fm2 = intersect_trees fm1 fm2
446 intersect_trees (LeafUFM i a) t2 =
449 Just b -> mkLeafUFM i (f a b)
451 intersect_trees t1 (LeafUFM i a) =
454 Just b -> mkLeafUFM i (f b a)
456 intersect_trees left_t@(NodeUFM j p t1 t2) right_t@(NodeUFM j' p' t1' t2')
458 (ask_about_common_ancestor
462 -- Given a disjoint j,j' (p >^ p' && p' >^ p):
465 -- / \ + / \ ==> EmptyUFM
470 intersect_branches (NewRoot nd _) = EmptyUFM
476 -- t1 t2 t1' t2' t1 x t1' t2 x t2'
478 intersect_branches (SameRoot)
479 = mkSSNodeUFM (NodeUFMData j p)
480 (intersect_trees t1 t1')
481 (intersect_trees t2 t2')
482 -- Now the 4 different other ways; all like this:
484 -- Given j >^ j' (and, say, j > j')
488 -- t1 t2 t1' t2' t1' t2'
490 -- This does cut down the search space quite a bit.
492 intersect_branches (LeftRoot Leftt)
493 = intersect_trees t1 right_t
494 intersect_branches (LeftRoot Rightt)
495 = intersect_trees t2 right_t
496 intersect_branches (RightRoot Leftt)
497 = intersect_trees left_t t1'
498 intersect_branches (RightRoot Rightt)
499 = intersect_trees left_t t2'
501 intersect_trees x y = panic ("EmptyUFM found when intersecting trees")
504 Now the usual set of `collection' operators, like map, fold, etc.
507 foldUFM f a (NodeUFM _ _ t1 t2) = foldUFM f (foldUFM f a t2) t1
508 foldUFM f a (LeafUFM _ obj) = f obj a
509 foldUFM f a EmptyUFM = a
513 mapUFM fn EmptyUFM = EmptyUFM
514 mapUFM fn fm = map_tree fn fm
516 filterUFM fn EmptyUFM = EmptyUFM
517 filterUFM fn fm = filter_tree fn fm
520 Note, this takes a long time, O(n), but
521 because we dont want to do this very often, we put up with this.
522 O'rable, but how often do we look at the size of
527 sizeUFM (NodeUFM _ _ t1 t2) = sizeUFM t1 + sizeUFM t2
528 sizeUFM (LeafUFM _ _) = 1
530 isNullUFM EmptyUFM = True
534 looking up in a hurry is the {\em whole point} of this binary tree lark.
535 Lookup up a binary tree is easy (and fast).
538 elemUFM key fm = case lookUp fm (u2i (getUnique key)) of
542 lookupUFM fm key = lookUp fm (u2i (getUnique key))
543 lookupUFM_Directly fm key = lookUp fm (u2i key)
545 lookupWithDefaultUFM fm deflt key
546 = case lookUp fm (u2i (getUnique key)) of
550 lookupWithDefaultUFM_Directly fm deflt key
551 = case lookUp fm (u2i key) of
555 lookUp EmptyUFM _ = Nothing
556 lookUp fm i = lookup_tree fm
558 lookup_tree :: UniqFM a -> Maybe a
560 lookup_tree (LeafUFM j b)
562 | otherwise = Nothing
563 lookup_tree (NodeUFM j p t1 t2)
564 | j _GT_ i = lookup_tree t1
565 | otherwise = lookup_tree t2
567 lookup_tree EmptyUFM = panic "lookup Failed"
570 folds are *wonderful* things.
573 eltsUFM fm = foldUFM (:) [] fm
575 ufmToList fm = fold_tree (\ iu elt rest -> (mkUniqueGrimily iu, elt) : rest) [] fm
577 keysUFM fm = fold_tree (\ iu elt rest -> IBOX(iu) : rest) [] fm
579 fold_tree f a (NodeUFM _ _ t1 t2) = fold_tree f (fold_tree f a t2) t1
580 fold_tree f a (LeafUFM iu obj) = f iu obj a
581 fold_tree f a EmptyUFM = a
584 %************************************************************************
586 \subsubsection{The @UniqFM@ type, and its functions}
588 %************************************************************************
590 You should always use these to build the tree.
591 There are 4 versions of mkNodeUFM, depending on
592 the strictness of the two sub-tree arguments.
593 The strictness is used *both* to prune out
594 empty trees, *and* to improve performance,
595 stoping needless thunks lying around.
596 The rule of thumb (from experence with these trees)
597 is make thunks strict, but data structures lazy.
598 If in doubt, use mkSSNodeUFM, which has the `strongest'
599 functionality, but may do a few needless evaluations.
602 mkLeafUFM :: FAST_INT -> a -> UniqFM a
603 mkLeafUFM i a = LeafUFM i a
605 -- The *ONLY* ways of building a NodeUFM.
607 mkSSNodeUFM (NodeUFMData j p) EmptyUFM t2 = t2
608 mkSSNodeUFM (NodeUFMData j p) t1 EmptyUFM = t1
609 mkSSNodeUFM (NodeUFMData j p) t1 t2
610 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
613 mkSLNodeUFM (NodeUFMData j p) EmptyUFM t2 = t2
614 mkSLNodeUFM (NodeUFMData j p) t1 t2
615 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
618 mkLSNodeUFM (NodeUFMData j p) t1 EmptyUFM = t1
619 mkLSNodeUFM (NodeUFMData j p) t1 t2
620 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
623 mkLLNodeUFM (NodeUFMData j p) t1 t2
624 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
634 correctNodeUFM j p t1 t2
635 = correct (j-p) (j-1) p t1 && correct j ((j-1)+p) p t2
637 correct low high _ (LeafUFM i _)
638 = low <= IBOX(i) && IBOX(i) <= high
639 correct low high above_p (NodeUFM j p _ _)
640 = low <= IBOX(j) && IBOX(j) <= high && above_p > IBOX(p)
641 correct _ _ _ EmptyUFM = panic "EmptyUFM stored inside a tree"
644 Note: doing SAT on this by hand seems to make it worse. Todo: Investigate,
645 and if necessary do $\lambda$ lifting on our functions that are bound.
655 insert_ele f EmptyUFM i new = mkLeafUFM i new
657 insert_ele f (LeafUFM j old) i new
659 mkLLNodeUFM (getCommonNodeUFMData
664 | j _EQ_ i = mkLeafUFM j (f old new)
666 mkLLNodeUFM (getCommonNodeUFMData
672 insert_ele f n@(NodeUFM j p t1 t2) i a
674 = if (i _GE_ (j _SUB_ p))
675 then mkSLNodeUFM (NodeUFMData j p) (insert_ele f t1 i a) t2
676 else mkLLNodeUFM (getCommonNodeUFMData
682 = if (i _LE_ ((j _SUB_ ILIT(1)) _ADD_ p))
683 then mkLSNodeUFM (NodeUFMData j p) t1 (insert_ele f t2 i a)
684 else mkLLNodeUFM (getCommonNodeUFMData
694 map_tree f (NodeUFM j p t1 t2)
695 = mkSSNodeUFM (NodeUFMData j p) (map_tree f t1) (map_tree f t2)
696 map_tree f (LeafUFM i obj)
697 = mkLeafUFM i (f obj)
699 map_tree f _ = panic "map_tree failed"
703 filter_tree f nd@(NodeUFM j p t1 t2)
704 = mkSSNodeUFM (NodeUFMData j p) (filter_tree f t1) (filter_tree f t2)
706 filter_tree f lf@(LeafUFM i obj)
708 | otherwise = EmptyUFM
709 filter_tree f _ = panic "filter_tree failed"
712 %************************************************************************
714 \subsubsection{The @UniqFM@ type, and signatures for the functions}
716 %************************************************************************
720 This is the information that is held inside a NodeUFM, packaged up for
725 = NodeUFMData FAST_INT
729 This is the information used when computing new NodeUFMs.
732 data Side = Leftt | Rightt -- NB: avoid 1.3 names "Left" and "Right"
734 = LeftRoot Side -- which side is the right down ?
735 | RightRoot Side -- which side is the left down ?
736 | SameRoot -- they are the same !
737 | NewRoot NodeUFMData -- here's the new, common, root
738 Bool -- do you need to swap left and right ?
741 This specifies the relationship between NodeUFMData and CalcNodeUFMData.
744 indexToRoot :: FAST_INT -> NodeUFMData
748 l = (ILIT(1) :: FAST_INT)
750 NodeUFMData (((i `shiftR_` l) `shiftL_` l) _ADD_ ILIT(1)) l
752 getCommonNodeUFMData :: NodeUFMData -> NodeUFMData -> NodeUFMData
754 getCommonNodeUFMData (NodeUFMData i p) (NodeUFMData i2 p2)
755 | p _EQ_ p2 = getCommonNodeUFMData_ p j j2
756 | p _LT_ p2 = getCommonNodeUFMData_ p2 (j _QUOT_ (p2 _QUOT_ p)) j2
757 | otherwise = getCommonNodeUFMData_ p j (j2 _QUOT_ (p _QUOT_ p2))
759 l = (ILIT(1) :: FAST_INT)
760 j = i _QUOT_ (p `shiftL_` l)
761 j2 = i2 _QUOT_ (p2 `shiftL_` l)
763 getCommonNodeUFMData_ :: FAST_INT -> FAST_INT -> FAST_INT -> NodeUFMData
765 getCommonNodeUFMData_ p j j_
767 = NodeUFMData (((j `shiftL_` l) _ADD_ l) _MUL_ p) p
769 = getCommonNodeUFMData_ (p `shiftL_` l) (j `shiftR_` l) (j_ `shiftR_` l)
771 ask_about_common_ancestor :: NodeUFMData -> NodeUFMData -> CommonRoot
773 ask_about_common_ancestor x@(NodeUFMData j p) y@(NodeUFMData j2 p2)
774 | j _EQ_ j2 = SameRoot
776 = case getCommonNodeUFMData x y of
777 nd@(NodeUFMData j3 p3)
778 | j3 _EQ_ j -> LeftRoot (decideSide (j _GT_ j2))
779 | j3 _EQ_ j2 -> RightRoot (decideSide (j _LT_ j2))
780 | otherwise -> NewRoot nd (j _GT_ j2)
782 decideSide :: Bool -> Side
783 decideSide True = Leftt
784 decideSide False = Rightt
787 This might be better in Util.lhs ?
790 Now the bit twiddling functions.
792 shiftL_ :: FAST_INT -> FAST_INT -> FAST_INT
793 shiftR_ :: FAST_INT -> FAST_INT -> FAST_INT
795 #if __GLASGOW_HASKELL__
796 {-# INLINE shiftL_ #-}
797 {-# INLINE shiftR_ #-}
798 shiftL_ n p = word2Int#((int2Word# n) `shiftL#` p)
799 shiftR_ n p = word2Int#((int2Word# n) `shiftr` p)
801 shiftr x y = shiftRL# x y
804 shiftL_ n p = n * (2 ^ p)
805 shiftR_ n p = n `quot` (2 ^ p)
811 use_snd :: a -> b -> b