2 % (c) The AQUA Project, Glasgow University, 1994-1996
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 @uniqueOf@ 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,
47 #include "HsVersions.h"
49 import {-# SOURCE #-} Name ( Name )
51 import Unique ( Uniquable(..), Unique, u2i, mkUniqueGrimily )
53 import GlaExts -- Lots of Int# operations
55 #if ! OMIT_NATIVE_CODEGEN
58 #define IF_NCG(a) {--}
62 %************************************************************************
64 \subsection{The @UniqFM@ type, and signatures for the functions}
66 %************************************************************************
68 We use @FiniteMaps@, with a (@uniqueOf@-able) @Unique@ as ``key''.
71 emptyUFM :: UniqFM elt
72 isNullUFM :: UniqFM elt -> Bool
73 unitUFM :: Uniquable key => key -> elt -> UniqFM elt
74 unitDirectlyUFM -- got the Unique already
75 :: Unique -> elt -> UniqFM elt
76 listToUFM :: Uniquable key => [(key,elt)] -> UniqFM elt
78 :: [(Unique, elt)] -> UniqFM elt
80 addToUFM :: Uniquable key => UniqFM elt -> key -> elt -> UniqFM elt
81 addListToUFM :: Uniquable key => UniqFM elt -> [(key,elt)] -> UniqFM elt
83 :: UniqFM elt -> Unique -> elt -> UniqFM elt
85 addToUFM_C :: Uniquable key => (elt -> elt -> elt)
86 -> UniqFM elt -> key -> elt -> UniqFM elt
87 addListToUFM_C :: Uniquable key => (elt -> elt -> elt)
88 -> UniqFM elt -> [(key,elt)]
91 delFromUFM :: Uniquable key => UniqFM elt -> key -> UniqFM elt
92 delListFromUFM :: Uniquable key => UniqFM elt -> [key] -> UniqFM elt
93 delFromUFM_Directly :: UniqFM elt -> Unique -> UniqFM elt
95 plusUFM :: UniqFM elt -> UniqFM elt -> UniqFM elt
97 plusUFM_C :: (elt -> elt -> elt)
98 -> UniqFM elt -> UniqFM elt -> UniqFM elt
100 minusUFM :: UniqFM elt -> UniqFM elt -> UniqFM elt
102 intersectUFM :: UniqFM elt -> UniqFM elt -> UniqFM elt
103 intersectUFM_C :: (elt -> elt -> elt)
104 -> UniqFM elt -> UniqFM elt -> UniqFM elt
105 foldUFM :: (elt -> a -> a) -> a -> UniqFM elt -> a
106 mapUFM :: (elt1 -> elt2) -> UniqFM elt1 -> UniqFM elt2
107 filterUFM :: (elt -> Bool) -> UniqFM elt -> UniqFM elt
109 sizeUFM :: UniqFM elt -> Int
110 elemUFM :: Uniquable key => key -> UniqFM elt -> Bool
112 lookupUFM :: Uniquable key => UniqFM elt -> key -> Maybe elt
113 lookupUFM_Directly -- when you've got the Unique already
114 :: UniqFM elt -> Unique -> Maybe elt
116 :: Uniquable key => UniqFM elt -> elt -> key -> elt
117 lookupWithDefaultUFM_Directly
118 :: UniqFM elt -> elt -> Unique -> elt
120 keysUFM :: UniqFM elt -> [Int] -- Get the keys
121 eltsUFM :: UniqFM elt -> [elt]
122 ufmToList :: UniqFM elt -> [(Unique, elt)]
125 %************************************************************************
127 \subsection{The @IdFinMap@ and @TyVarFinMap@ specialisations for Ids/TyVars}
129 %************************************************************************
132 -- Turn off for now, these need to be updated (SDM 4/98)
135 #ifdef __GLASGOW_HASKELL__
136 -- I don't think HBC was too happy about this (WDP 94/10)
139 addListToUFM :: UniqFM elt -> [(Name, elt)] -> UniqFM elt
142 addListToUFM_C :: (elt -> elt -> elt) -> UniqFM elt -> [(Name, elt)] -> UniqFM elt
145 addToUFM :: UniqFM elt -> Unique -> elt -> UniqFM elt
148 listToUFM :: [(Unique, elt)] -> UniqFM elt
151 lookupUFM :: UniqFM elt -> Name -> Maybe elt
152 , UniqFM elt -> Unique -> Maybe elt
155 #endif {- __GLASGOW_HASKELL__ -}
159 %************************************************************************
161 \subsection{Andy Gill's underlying @UniqFM@ machinery}
163 %************************************************************************
165 ``Uniq Finite maps'' are the heart and soul of the compiler's
166 lookup-tables/environments. Important stuff! It works well with
167 Dense and Sparse ranges.
168 Both @Uq@ Finite maps and @Hash@ Finite Maps
169 are built ontop of Int Finite Maps.
171 This code is explained in the paper:
173 A Gill, S Peyton Jones, B O'Sullivan, W Partain and Aqua Friends
174 "A Cheap balancing act that grows on a tree"
175 Glasgow FP Workshop, Sep 1994, pp??-??
178 %************************************************************************
180 \subsubsection{The @UniqFM@ type, and signatures for the functions}
182 %************************************************************************
184 @UniqFM a@ is a mapping from Unique to a.
186 First, the DataType itself; which is either a Node, a Leaf, or an Empty.
191 | LeafUFM FAST_INT ele
192 | NodeUFM FAST_INT -- the switching
193 FAST_INT -- the delta
197 -- for debugging only :-)
199 instance Text (UniqFM a) where
200 showsPrec _ (NodeUFM a b t1 t2) =
201 showString "NodeUFM " . shows (IBOX(a))
202 . showString " " . shows (IBOX(b))
203 . showString " (" . shows t1
204 . showString ") (" . shows t2
206 showsPrec _ (LeafUFM x a) = showString "LeafUFM " . shows (IBOX(x))
207 showsPrec _ (EmptyUFM) = id
211 %************************************************************************
213 \subsubsection{The @UniqFM@ functions}
215 %************************************************************************
217 First the ways of building a UniqFM.
221 unitUFM key elt = mkLeafUFM (u2i (uniqueOf key)) elt
222 unitDirectlyUFM key elt = mkLeafUFM (u2i key) elt
224 listToUFM key_elt_pairs
225 = addListToUFM_C use_snd EmptyUFM key_elt_pairs
227 listToUFM_Directly uniq_elt_pairs
228 = addListToUFM_directly_C use_snd EmptyUFM uniq_elt_pairs
231 Now ways of adding things to UniqFMs.
233 There is an alternative version of @addListToUFM_C@, that uses @plusUFM@,
234 but the semantics of this operation demands a linear insertion;
235 perhaps the version without the combinator function
236 could be optimised using it.
239 addToUFM fm key elt = addToUFM_C use_snd fm key elt
241 addToUFM_Directly fm u elt = insert_ele use_snd fm (u2i u) elt
243 addToUFM_C combiner fm key elt
244 = insert_ele combiner fm (u2i (uniqueOf key)) elt
246 addListToUFM fm key_elt_pairs = addListToUFM_C use_snd fm key_elt_pairs
247 addListToUFM_Directly fm uniq_elt_pairs = addListToUFM_directly_C use_snd fm uniq_elt_pairs
249 addListToUFM_C combiner fm key_elt_pairs
250 = foldl (\ fm (k, e) -> insert_ele combiner fm (u2i (uniqueOf k)) e)
253 addListToUFM_directly_C combiner fm uniq_elt_pairs
254 = foldl (\ fm (k, e) -> insert_ele combiner fm (u2i k) e)
258 Now ways of removing things from UniqFM.
261 delListFromUFM fm lst = foldl delFromUFM fm lst
263 delFromUFM fm key = delete fm (u2i (uniqueOf key))
264 delFromUFM_Directly fm u = delete fm (u2i u)
266 delete EmptyUFM _ = EmptyUFM
267 delete fm key = del_ele fm
269 del_ele :: UniqFM a -> UniqFM a
271 del_ele lf@(LeafUFM j _)
272 | j _EQ_ key = EmptyUFM
273 | otherwise = lf -- no delete!
275 del_ele nd@(NodeUFM j p t1 t2)
277 = mkSLNodeUFM (NodeUFMData j p) (del_ele t1) t2
279 = mkLSNodeUFM (NodeUFMData j p) t1 (del_ele t2)
281 del_ele _ = panic "Found EmptyUFM FM when rec-deleting"
284 Now ways of adding two UniqFM's together.
287 plusUFM tr1 tr2 = plusUFM_C use_snd tr1 tr2
289 plusUFM_C f EmptyUFM tr = tr
290 plusUFM_C f tr EmptyUFM = tr
291 plusUFM_C f fm1 fm2 = mix_trees fm1 fm2
293 mix_trees (LeafUFM i a) t2 = insert_ele (flip f) t2 i a
294 mix_trees t1 (LeafUFM i a) = insert_ele f t1 i a
296 mix_trees left_t@(NodeUFM j p t1 t2) right_t@(NodeUFM j' p' t1' t2')
298 (ask_about_common_ancestor
302 -- Given a disjoint j,j' (p >^ p' && p' >^ p):
306 -- t1 t2 t1' t2' j j'
311 mix_branches (NewRoot nd False)
312 = mkLLNodeUFM nd left_t right_t
313 mix_branches (NewRoot nd True)
314 = mkLLNodeUFM nd right_t left_t
320 -- t1 t2 t1' t2' t1 + t1' t2 + t2'
322 mix_branches (SameRoot)
323 = mkSSNodeUFM (NodeUFMData j p)
326 -- Now the 4 different other ways; all like this:
328 -- Given j >^ j' (and, say, j > j')
332 -- t1 t2 t1' t2' t1 t2 + j'
335 mix_branches (LeftRoot Leftt) -- | trace "LL" True
338 (mix_trees t1 right_t)
341 mix_branches (LeftRoot Rightt) -- | trace "LR" True
345 (mix_trees t2 right_t)
347 mix_branches (RightRoot Leftt) -- | trace "RL" True
350 (mix_trees left_t t1')
353 mix_branches (RightRoot Rightt) -- | trace "RR" True
357 (mix_trees left_t t2')
359 mix_trees _ _ = panic "EmptyUFM found when inserting into plusInt"
362 And ways of subtracting them. First the base cases,
363 then the full D&C approach.
366 minusUFM EmptyUFM _ = EmptyUFM
367 minusUFM t1 EmptyUFM = t1
368 minusUFM fm1 fm2 = minus_trees fm1 fm2
371 -- Notice the asymetry of subtraction
373 minus_trees lf@(LeafUFM i a) t2 =
378 minus_trees t1 (LeafUFM i _) = delete t1 i
380 minus_trees left_t@(NodeUFM j p t1 t2) right_t@(NodeUFM j' p' t1' t2')
382 (ask_about_common_ancestor
386 -- Given a disjoint j,j' (p >^ p' && p' >^ p):
390 -- t1 t2 t1' t2' t1 t2
395 minus_branches (NewRoot nd _) = left_t
401 -- t1 t2 t1' t2' t1 + t1' t2 + t2'
403 minus_branches (SameRoot)
404 = mkSSNodeUFM (NodeUFMData j p)
407 -- Now the 4 different other ways; all like this:
408 -- again, with asymatry
411 -- The left is above the right
413 minus_branches (LeftRoot Leftt)
416 (minus_trees t1 right_t)
418 minus_branches (LeftRoot Rightt)
422 (minus_trees t2 right_t)
425 -- The right is above the left
427 minus_branches (RightRoot Leftt)
428 = minus_trees left_t t1'
429 minus_branches (RightRoot Rightt)
430 = minus_trees left_t t2'
432 minus_trees _ _ = panic "EmptyUFM found when insering into plusInt"
435 And taking the intersection of two UniqFM's.
438 intersectUFM t1 t2 = intersectUFM_C use_snd t1 t2
440 intersectUFM_C f EmptyUFM _ = EmptyUFM
441 intersectUFM_C f _ EmptyUFM = EmptyUFM
442 intersectUFM_C f fm1 fm2 = intersect_trees fm1 fm2
444 intersect_trees (LeafUFM i a) t2 =
447 Just b -> mkLeafUFM i (f a b)
449 intersect_trees t1 (LeafUFM i a) =
452 Just b -> mkLeafUFM i (f b a)
454 intersect_trees left_t@(NodeUFM j p t1 t2) right_t@(NodeUFM j' p' t1' t2')
456 (ask_about_common_ancestor
460 -- Given a disjoint j,j' (p >^ p' && p' >^ p):
463 -- / \ + / \ ==> EmptyUFM
468 intersect_branches (NewRoot nd _) = EmptyUFM
474 -- t1 t2 t1' t2' t1 x t1' t2 x t2'
476 intersect_branches (SameRoot)
477 = mkSSNodeUFM (NodeUFMData j p)
478 (intersect_trees t1 t1')
479 (intersect_trees t2 t2')
480 -- Now the 4 different other ways; all like this:
482 -- Given j >^ j' (and, say, j > j')
486 -- t1 t2 t1' t2' t1' t2'
488 -- This does cut down the search space quite a bit.
490 intersect_branches (LeftRoot Leftt)
491 = intersect_trees t1 right_t
492 intersect_branches (LeftRoot Rightt)
493 = intersect_trees t2 right_t
494 intersect_branches (RightRoot Leftt)
495 = intersect_trees left_t t1'
496 intersect_branches (RightRoot Rightt)
497 = intersect_trees left_t t2'
499 intersect_trees x y = panic ("EmptyUFM found when intersecting trees")
502 Now the usual set of `collection' operators, like map, fold, etc.
505 foldUFM f a (NodeUFM _ _ t1 t2) = foldUFM f (foldUFM f a t2) t1
506 foldUFM f a (LeafUFM _ obj) = f obj a
507 foldUFM f a EmptyUFM = a
511 mapUFM fn EmptyUFM = EmptyUFM
512 mapUFM fn fm = map_tree fn fm
514 filterUFM fn EmptyUFM = EmptyUFM
515 filterUFM fn fm = filter_tree fn fm
518 Note, this takes a long time, O(n), but
519 because we dont want to do this very often, we put up with this.
520 O'rable, but how often do we look at the size of
525 sizeUFM (NodeUFM _ _ t1 t2) = sizeUFM t1 + sizeUFM t2
526 sizeUFM (LeafUFM _ _) = 1
528 isNullUFM EmptyUFM = True
532 looking up in a hurry is the {\em whole point} of this binary tree lark.
533 Lookup up a binary tree is easy (and fast).
536 elemUFM key fm = case lookUp fm (u2i (uniqueOf key)) of
540 lookupUFM fm key = lookUp fm (u2i (uniqueOf key))
541 lookupUFM_Directly fm key = lookUp fm (u2i key)
543 lookupWithDefaultUFM fm deflt key
544 = case lookUp fm (u2i (uniqueOf key)) of
548 lookupWithDefaultUFM_Directly fm deflt key
549 = case lookUp fm (u2i key) of
553 lookUp EmptyUFM _ = Nothing
554 lookUp fm i = lookup_tree fm
556 lookup_tree :: UniqFM a -> Maybe a
558 lookup_tree (LeafUFM j b)
560 | otherwise = Nothing
561 lookup_tree (NodeUFM j p t1 t2)
562 | j _GT_ i = lookup_tree t1
563 | otherwise = lookup_tree t2
565 lookup_tree EmptyUFM = panic "lookup Failed"
568 folds are *wonderful* things.
571 eltsUFM fm = foldUFM (:) [] fm
573 ufmToList fm = fold_tree (\ iu elt rest -> (mkUniqueGrimily iu, elt) : rest) [] fm
575 keysUFM fm = fold_tree (\ iu elt rest -> IBOX(iu) : rest) [] fm
577 fold_tree f a (NodeUFM _ _ t1 t2) = fold_tree f (fold_tree f a t2) t1
578 fold_tree f a (LeafUFM iu obj) = f iu obj a
579 fold_tree f a EmptyUFM = a
582 %************************************************************************
584 \subsubsection{The @UniqFM@ type, and its functions}
586 %************************************************************************
588 You should always use these to build the tree.
589 There are 4 versions of mkNodeUFM, depending on
590 the strictness of the two sub-tree arguments.
591 The strictness is used *both* to prune out
592 empty trees, *and* to improve performance,
593 stoping needless thunks lying around.
594 The rule of thumb (from experence with these trees)
595 is make thunks strict, but data structures lazy.
596 If in doubt, use mkSSNodeUFM, which has the `strongest'
597 functionality, but may do a few needless evaluations.
600 mkLeafUFM :: FAST_INT -> a -> UniqFM a
601 mkLeafUFM i a = LeafUFM i a
603 -- The *ONLY* ways of building a NodeUFM.
605 mkSSNodeUFM (NodeUFMData j p) EmptyUFM t2 = t2
606 mkSSNodeUFM (NodeUFMData j p) t1 EmptyUFM = t1
607 mkSSNodeUFM (NodeUFMData j p) t1 t2
608 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
611 mkSLNodeUFM (NodeUFMData j p) EmptyUFM t2 = t2
612 mkSLNodeUFM (NodeUFMData j p) t1 t2
613 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
616 mkLSNodeUFM (NodeUFMData j p) t1 EmptyUFM = t1
617 mkLSNodeUFM (NodeUFMData j p) t1 t2
618 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
621 mkLLNodeUFM (NodeUFMData j p) t1 t2
622 = ASSERT(correctNodeUFM (IBOX(j)) (IBOX(p)) t1 t2)
632 correctNodeUFM j p t1 t2
633 = correct (j-p) (j-1) p t1 && correct j ((j-1)+p) p t2
635 correct low high _ (LeafUFM i _)
636 = low <= IBOX(i) && IBOX(i) <= high
637 correct low high above_p (NodeUFM j p _ _)
638 = low <= IBOX(j) && IBOX(j) <= high && above_p > IBOX(p)
639 correct _ _ _ EmptyUFM = panic "EmptyUFM stored inside a tree"
642 Note: doing SAT on this by hand seems to make it worse. Todo: Investigate,
643 and if necessary do $\lambda$ lifting on our functions that are bound.
653 insert_ele f EmptyUFM i new = mkLeafUFM i new
655 insert_ele f (LeafUFM j old) i new
657 mkLLNodeUFM (getCommonNodeUFMData
662 | j _EQ_ i = mkLeafUFM j (f old new)
664 mkLLNodeUFM (getCommonNodeUFMData
670 insert_ele f n@(NodeUFM j p t1 t2) i a
672 = if (i _GE_ (j _SUB_ p))
673 then mkSLNodeUFM (NodeUFMData j p) (insert_ele f t1 i a) t2
674 else mkLLNodeUFM (getCommonNodeUFMData
680 = if (i _LE_ ((j _SUB_ ILIT(1)) _ADD_ p))
681 then mkLSNodeUFM (NodeUFMData j p) t1 (insert_ele f t2 i a)
682 else mkLLNodeUFM (getCommonNodeUFMData
692 map_tree f (NodeUFM j p t1 t2)
693 = mkSSNodeUFM (NodeUFMData j p) (map_tree f t1) (map_tree f t2)
694 map_tree f (LeafUFM i obj)
695 = mkLeafUFM i (f obj)
697 map_tree f _ = panic "map_tree failed"
701 filter_tree f nd@(NodeUFM j p t1 t2)
702 = mkSSNodeUFM (NodeUFMData j p) (filter_tree f t1) (filter_tree f t2)
704 filter_tree f lf@(LeafUFM i obj)
706 | otherwise = EmptyUFM
707 filter_tree f _ = panic "filter_tree failed"
710 %************************************************************************
712 \subsubsection{The @UniqFM@ type, and signatures for the functions}
714 %************************************************************************
718 This is the information that is held inside a NodeUFM, packaged up for
723 = NodeUFMData FAST_INT
727 This is the information used when computing new NodeUFMs.
730 data Side = Leftt | Rightt -- NB: avoid 1.3 names "Left" and "Right"
732 = LeftRoot Side -- which side is the right down ?
733 | RightRoot Side -- which side is the left down ?
734 | SameRoot -- they are the same !
735 | NewRoot NodeUFMData -- here's the new, common, root
736 Bool -- do you need to swap left and right ?
739 This specifies the relationship between NodeUFMData and CalcNodeUFMData.
742 indexToRoot :: FAST_INT -> NodeUFMData
746 l = (ILIT(1) :: FAST_INT)
748 NodeUFMData (((i `shiftR_` l) `shiftL_` l) _ADD_ ILIT(1)) l
750 getCommonNodeUFMData :: NodeUFMData -> NodeUFMData -> NodeUFMData
752 getCommonNodeUFMData (NodeUFMData i p) (NodeUFMData i2 p2)
753 | p _EQ_ p2 = getCommonNodeUFMData_ p j j2
754 | p _LT_ p2 = getCommonNodeUFMData_ p2 (j _QUOT_ (p2 _QUOT_ p)) j2
755 | otherwise = getCommonNodeUFMData_ p j (j2 _QUOT_ (p _QUOT_ p2))
757 l = (ILIT(1) :: FAST_INT)
758 j = i _QUOT_ (p `shiftL_` l)
759 j2 = i2 _QUOT_ (p2 `shiftL_` l)
761 getCommonNodeUFMData_ :: FAST_INT -> FAST_INT -> FAST_INT -> NodeUFMData
763 getCommonNodeUFMData_ p j j_
765 = NodeUFMData (((j `shiftL_` l) _ADD_ l) _MUL_ p) p
767 = getCommonNodeUFMData_ (p `shiftL_` l) (j `shiftR_` l) (j_ `shiftR_` l)
769 ask_about_common_ancestor :: NodeUFMData -> NodeUFMData -> CommonRoot
771 ask_about_common_ancestor x@(NodeUFMData j p) y@(NodeUFMData j2 p2)
772 | j _EQ_ j2 = SameRoot
774 = case getCommonNodeUFMData x y of
775 nd@(NodeUFMData j3 p3)
776 | j3 _EQ_ j -> LeftRoot (decideSide (j _GT_ j2))
777 | j3 _EQ_ j2 -> RightRoot (decideSide (j _LT_ j2))
778 | otherwise -> NewRoot nd (j _GT_ j2)
780 decideSide :: Bool -> Side
781 decideSide True = Leftt
782 decideSide False = Rightt
785 This might be better in Util.lhs ?
788 Now the bit twiddling functions.
790 shiftL_ :: FAST_INT -> FAST_INT -> FAST_INT
791 shiftR_ :: FAST_INT -> FAST_INT -> FAST_INT
793 #if __GLASGOW_HASKELL__
794 {-# INLINE shiftL_ #-}
795 {-# INLINE shiftR_ #-}
796 shiftL_ n p = word2Int#((int2Word# n) `shiftL#` p)
797 shiftR_ n p = word2Int#((int2Word# n) `shiftr` p)
799 shiftr x y = shiftRA# x y
802 shiftL_ n p = n * (2 ^ p)
803 shiftR_ n p = n `quot` (2 ^ p)
809 use_snd :: a -> b -> b