mirrored from https://gitlab.haskell.org/ghc/ghc.git
/
SimplUtils.lhs
1420 lines (1175 loc) · 50.5 KB
/
SimplUtils.lhs
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
%
% (c) The AQUA Project, Glasgow University, 1993-1998
%
\section[SimplUtils]{The simplifier utilities}
\begin{code}
module SimplUtils (
-- Rebuilding
mkLam, mkCase, prepareAlts, bindCaseBndr,
-- Inlining,
preInlineUnconditionally, postInlineUnconditionally,
activeInline, activeRule, inlineMode,
-- The continuation type
SimplCont(..), DupFlag(..), LetRhsFlag(..),
contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
countValArgs, countArgs,
mkBoringStop, mkLazyArgStop, mkRhsStop, contIsRhsOrArg,
interestingCallContext, interestingArgContext,
interestingArg, mkArgInfo
) where
#include "HsVersions.h"
import SimplEnv
import DynFlags
import StaticFlags
import CoreSyn
import PprCore
import CoreFVs
import CoreUtils
import Literal
import CoreUnfold
import MkId
import Id
import NewDemand
import SimplMonad
import Type
import TyCon
import DataCon
import Unify ( dataConCannotMatch )
import VarSet
import BasicTypes
import Util
import Outputable
import List( nub )
\end{code}
%************************************************************************
%* *
The SimplCont type
%* *
%************************************************************************
A SimplCont allows the simplifier to traverse the expression in a
zipper-like fashion. The SimplCont represents the rest of the expression,
"above" the point of interest.
You can also think of a SimplCont as an "evaluation context", using
that term in the way it is used for operational semantics. This is the
way I usually think of it, For example you'll often see a syntax for
evaluation context looking like
C ::= [] | C e | case C of alts | C `cast` co
That's the kind of thing we are doing here, and I use that syntax in
the comments.
Key points:
* A SimplCont describes a *strict* context (just like
evaluation contexts do). E.g. Just [] is not a SimplCont
* A SimplCont describes a context that *does not* bind
any variables. E.g. \x. [] is not a SimplCont
\begin{code}
data SimplCont
= Stop -- An empty context, or hole, []
OutType -- Type of the result
LetRhsFlag
Bool -- True <=> There is something interesting about
-- the context, and hence the inliner
-- should be a bit keener (see interestingCallContext)
-- Two cases:
-- (a) This is the RHS of a thunk whose type suggests
-- that update-in-place would be possible
-- (b) This is an argument of a function that has RULES
-- Inlining the call might allow the rule to fire
| CoerceIt -- C `cast` co
OutCoercion -- The coercion simplified
SimplCont
| ApplyTo -- C arg
DupFlag
InExpr SimplEnv -- The argument and its static env
SimplCont
| Select -- case C of alts
DupFlag
InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
SimplCont
-- The two strict forms have no DupFlag, because we never duplicate them
| StrictBind -- (\x* \xs. e) C
InId [InBndr] -- let x* = [] in e
InExpr SimplEnv -- is a special case
SimplCont
| StrictArg -- e C
OutExpr OutType -- e and its type
(Bool,[Bool]) -- Whether the function at the head of e has rules,
SimplCont -- plus strictness flags for further args
data LetRhsFlag = AnArg -- It's just an argument not a let RHS
| AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
instance Outputable LetRhsFlag where
ppr AnArg = ptext SLIT("arg")
ppr AnRhs = ptext SLIT("rhs")
instance Outputable SimplCont where
ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty
ppr (ApplyTo dup arg se cont) = ((ptext SLIT("ApplyTo") <+> ppr dup <+> pprParendExpr arg)
{- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
ppr (StrictBind b _ _ _ cont) = (ptext SLIT("StrictBind") <+> ppr b) $$ ppr cont
ppr (StrictArg f _ _ cont) = (ptext SLIT("StrictArg") <+> ppr f) $$ ppr cont
ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
(nest 4 (ppr alts)) $$ ppr cont
ppr (CoerceIt co cont) = (ptext SLIT("CoerceIt") <+> ppr co) $$ ppr cont
data DupFlag = OkToDup | NoDup
instance Outputable DupFlag where
ppr OkToDup = ptext SLIT("ok")
ppr NoDup = ptext SLIT("nodup")
-------------------
mkBoringStop :: OutType -> SimplCont
mkBoringStop ty = Stop ty AnArg False
mkLazyArgStop :: OutType -> Bool -> SimplCont
mkLazyArgStop ty has_rules = Stop ty AnArg (canUpdateInPlace ty || has_rules)
mkRhsStop :: OutType -> SimplCont
mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
contIsRhsOrArg (Stop _ _ _) = True
contIsRhsOrArg (StrictBind {}) = True
contIsRhsOrArg (StrictArg {}) = True
contIsRhsOrArg other = False
-------------------
contIsDupable :: SimplCont -> Bool
contIsDupable (Stop _ _ _) = True
contIsDupable (ApplyTo OkToDup _ _ _) = True
contIsDupable (Select OkToDup _ _ _ _) = True
contIsDupable (CoerceIt _ cont) = contIsDupable cont
contIsDupable other = False
-------------------
contIsTrivial :: SimplCont -> Bool
contIsTrivial (Stop _ _ _) = True
contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
contIsTrivial other = False
-------------------
contResultType :: SimplCont -> OutType
contResultType (Stop to_ty _ _) = to_ty
contResultType (StrictArg _ _ _ cont) = contResultType cont
contResultType (StrictBind _ _ _ _ cont) = contResultType cont
contResultType (ApplyTo _ _ _ cont) = contResultType cont
contResultType (CoerceIt _ cont) = contResultType cont
contResultType (Select _ _ _ _ cont) = contResultType cont
-------------------
countValArgs :: SimplCont -> Int
countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
countValArgs other = 0
countArgs :: SimplCont -> Int
countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
countArgs other = 0
contArgs :: SimplCont -> ([OutExpr], SimplCont)
-- Uses substitution to turn each arg into an OutExpr
contArgs cont = go [] cont
where
go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
go args cont = (reverse args, cont)
dropArgs :: Int -> SimplCont -> SimplCont
dropArgs 0 cont = cont
dropArgs n (ApplyTo _ _ _ cont) = dropArgs (n-1) cont
dropArgs n other = pprPanic "dropArgs" (ppr n <+> ppr other)
\end{code}
\begin{code}
interestingArg :: OutExpr -> Bool
-- An argument is interesting if it has *some* structure
-- We are here trying to avoid unfolding a function that
-- is applied only to variables that have no unfolding
-- (i.e. they are probably lambda bound): f x y z
-- There is little point in inlining f here.
interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
-- Was: isValueUnfolding (idUnfolding v')
-- But that seems over-pessimistic
|| isDataConWorkId v
-- This accounts for an argument like
-- () or [], which is definitely interesting
interestingArg (Type _) = False
interestingArg (App fn (Type _)) = interestingArg fn
interestingArg (Note _ a) = interestingArg a
-- Idea (from Sam B); I'm not sure if it's a good idea, so commented out for now
-- interestingArg expr | isUnLiftedType (exprType expr)
-- -- Unlifted args are only ever interesting if we know what they are
-- = case expr of
-- Lit lit -> True
-- _ -> False
interestingArg other = True
-- Consider let x = 3 in f x
-- The substitution will contain (x -> ContEx 3), and we want to
-- to say that x is an interesting argument.
-- But consider also (\x. f x y) y
-- The substitution will contain (x -> ContEx y), and we want to say
-- that x is not interesting (assuming y has no unfolding)
\end{code}
Comment about interestingCallContext
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We want to avoid inlining an expression where there can't possibly be
any gain, such as in an argument position. Hence, if the continuation
is interesting (eg. a case scrutinee, application etc.) then we
inline, otherwise we don't.
Previously some_benefit used to return True only if the variable was
applied to some value arguments. This didn't work:
let x = _coerce_ (T Int) Int (I# 3) in
case _coerce_ Int (T Int) x of
I# y -> ....
we want to inline x, but can't see that it's a constructor in a case
scrutinee position, and some_benefit is False.
Another example:
dMonadST = _/\_ t -> :Monad (g1 _@_ t, g2 _@_ t, g3 _@_ t)
.... case dMonadST _@_ x0 of (a,b,c) -> ....
we'd really like to inline dMonadST here, but we *don't* want to
inline if the case expression is just
case x of y { DEFAULT -> ... }
since we can just eliminate this case instead (x is in WHNF). Similar
applies when x is bound to a lambda expression. Hence
contIsInteresting looks for case expressions with just a single
default case.
\begin{code}
interestingCallContext :: Bool -- False <=> no args at all
-> Bool -- False <=> no value args
-> SimplCont -> Bool
-- The "lone-variable" case is important. I spent ages
-- messing about with unsatisfactory varaints, but this is nice.
-- The idea is that if a variable appear all alone
-- as an arg of lazy fn, or rhs Stop
-- as scrutinee of a case Select
-- as arg of a strict fn ArgOf
-- then we should not inline it (unless there is some other reason,
-- e.g. is is the sole occurrence). We achieve this by making
-- interestingCallContext return False for a lone variable.
--
-- Why? At least in the case-scrutinee situation, turning
-- let x = (a,b) in case x of y -> ...
-- into
-- let x = (a,b) in case (a,b) of y -> ...
-- and thence to
-- let x = (a,b) in let y = (a,b) in ...
-- is bad if the binding for x will remain.
--
-- Another example: I discovered that strings
-- were getting inlined straight back into applications of 'error'
-- because the latter is strict.
-- s = "foo"
-- f = \x -> ...(error s)...
-- Fundamentally such contexts should not ecourage inlining because
-- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
-- so there's no gain.
--
-- However, even a type application or coercion isn't a lone variable.
-- Consider
-- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
-- We had better inline that sucker! The case won't see through it.
--
-- For now, I'm treating treating a variable applied to types
-- in a *lazy* context "lone". The motivating example was
-- f = /\a. \x. BIG
-- g = /\a. \y. h (f a)
-- There's no advantage in inlining f here, and perhaps
-- a significant disadvantage. Hence some_val_args in the Stop case
interestingCallContext some_args some_val_args cont
= interesting cont
where
interesting (Select {}) = some_args
interesting (ApplyTo {}) = True -- Can happen if we have (coerce t (f x)) y
-- Perhaps True is a bit over-keen, but I've
-- seen (coerce f) x, where f has an INLINE prag,
-- So we have to give some motivaiton for inlining it
interesting (StrictArg {}) = some_val_args
interesting (StrictBind {}) = some_val_args -- ??
interesting (Stop ty _ interesting) = some_val_args && interesting
interesting (CoerceIt _ cont) = interesting cont
-- If this call is the arg of a strict function, the context
-- is a bit interesting. If we inline here, we may get useful
-- evaluation information to avoid repeated evals: e.g.
-- x + (y * z)
-- Here the contIsInteresting makes the '*' keener to inline,
-- which in turn exposes a constructor which makes the '+' inline.
-- Assuming that +,* aren't small enough to inline regardless.
--
-- It's also very important to inline in a strict context for things
-- like
-- foldr k z (f x)
-- Here, the context of (f x) is strict, and if f's unfolding is
-- a build it's *great* to inline it here. So we must ensure that
-- the context for (f x) is not totally uninteresting.
-------------------
mkArgInfo :: Id
-> Int -- Number of value args
-> SimplCont -- Context of the cal
-> (Bool, [Bool]) -- Arg info
-- The arg info consists of
-- * A Bool indicating if the function has rules (recursively)
-- * A [Bool] indicating strictness for each arg
-- The [Bool] is usually infinite, but if it is finite it
-- guarantees that the function diverges after being given
-- that number of args
mkArgInfo fun n_val_args call_cont
= (interestingArgContext fun call_cont, fun_stricts)
where
vanilla_stricts, fun_stricts :: [Bool]
vanilla_stricts = repeat False
fun_stricts
= case splitStrictSig (idNewStrictness fun) of
(demands, result_info)
| not (demands `lengthExceeds` n_val_args)
-> -- Enough args, use the strictness given.
-- For bottoming functions we used to pretend that the arg
-- is lazy, so that we don't treat the arg as an
-- interesting context. This avoids substituting
-- top-level bindings for (say) strings into
-- calls to error. But now we are more careful about
-- inlining lone variables, so its ok (see SimplUtils.analyseCont)
if isBotRes result_info then
map isStrictDmd demands -- Finite => result is bottom
else
map isStrictDmd demands ++ vanilla_stricts
other -> vanilla_stricts -- Not enough args, or no strictness
interestingArgContext :: Id -> SimplCont -> Bool
-- If the argument has form (f x y), where x,y are boring,
-- and f is marked INLINE, then we don't want to inline f.
-- But if the context of the argument is
-- g (f x y)
-- where g has rules, then we *do* want to inline f, in case it
-- exposes a rule that might fire. Similarly, if the context is
-- h (g (f x x))
-- where h has rules, then we do want to inline f.
-- The interesting_arg_ctxt flag makes this happen; if it's
-- set, the inliner gets just enough keener to inline f
-- regardless of how boring f's arguments are, if it's marked INLINE
--
-- The alternative would be to *always* inline an INLINE function,
-- regardless of how boring its context is; but that seems overkill
-- For example, it'd mean that wrapper functions were always inlined
interestingArgContext fn cont
= idHasRules fn || go cont
where
go (Select {}) = False
go (ApplyTo {}) = False
go (StrictArg {}) = True
go (StrictBind {}) = False -- ??
go (CoerceIt _ c) = go c
go (Stop _ _ interesting) = interesting
-------------------
canUpdateInPlace :: Type -> Bool
-- Consider let x = <wurble> in ...
-- If <wurble> returns an explicit constructor, we might be able
-- to do update in place. So we treat even a thunk RHS context
-- as interesting if update in place is possible. We approximate
-- this by seeing if the type has a single constructor with a
-- small arity. But arity zero isn't good -- we share the single copy
-- for that case, so no point in sharing.
canUpdateInPlace ty
| not opt_UF_UpdateInPlace = False
| otherwise
= case splitTyConApp_maybe ty of
Nothing -> False
Just (tycon, _) -> case tyConDataCons_maybe tycon of
Just [dc] -> arity == 1 || arity == 2
where
arity = dataConRepArity dc
other -> False
\end{code}
%************************************************************************
%* *
\subsection{Decisions about inlining}
%* *
%************************************************************************
Inlining is controlled partly by the SimplifierMode switch. This has two
settings:
SimplGently (a) Simplifying before specialiser/full laziness
(b) Simplifiying inside INLINE pragma
(c) Simplifying the LHS of a rule
(d) Simplifying a GHCi expression or Template
Haskell splice
SimplPhase n Used at all other times
The key thing about SimplGently is that it does no call-site inlining.
Before full laziness we must be careful not to inline wrappers,
because doing so inhibits floating
e.g. ...(case f x of ...)...
==> ...(case (case x of I# x# -> fw x#) of ...)...
==> ...(case x of I# x# -> case fw x# of ...)...
and now the redex (f x) isn't floatable any more.
The no-inlining thing is also important for Template Haskell. You might be
compiling in one-shot mode with -O2; but when TH compiles a splice before
running it, we don't want to use -O2. Indeed, we don't want to inline
anything, because the byte-code interpreter might get confused about
unboxed tuples and suchlike.
INLINE pragmas
~~~~~~~~~~~~~~
SimplGently is also used as the mode to simplify inside an InlineMe note.
\begin{code}
inlineMode :: SimplifierMode
inlineMode = SimplGently
\end{code}
It really is important to switch off inlinings inside such
expressions. Consider the following example
let f = \pq -> BIG
in
let g = \y -> f y y
{-# INLINE g #-}
in ...g...g...g...g...g...
Now, if that's the ONLY occurrence of f, it will be inlined inside g,
and thence copied multiple times when g is inlined.
This function may be inlinined in other modules, so we
don't want to remove (by inlining) calls to functions that have
specialisations, or that may have transformation rules in an importing
scope.
E.g. {-# INLINE f #-}
f x = ...g...
and suppose that g is strict *and* has specialisations. If we inline
g's wrapper, we deny f the chance of getting the specialised version
of g when f is inlined at some call site (perhaps in some other
module).
It's also important not to inline a worker back into a wrapper.
A wrapper looks like
wraper = inline_me (\x -> ...worker... )
Normally, the inline_me prevents the worker getting inlined into
the wrapper (initially, the worker's only call site!). But,
if the wrapper is sure to be called, the strictness analyser will
mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
continuation. That's why the keep_inline predicate returns True for
ArgOf continuations. It shouldn't do any harm not to dissolve the
inline-me note under these circumstances.
Note that the result is that we do very little simplification
inside an InlineMe.
all xs = foldr (&&) True xs
any p = all . map p {-# INLINE any #-}
Problem: any won't get deforested, and so if it's exported and the
importer doesn't use the inlining, (eg passes it as an arg) then we
won't get deforestation at all. We havn't solved this problem yet!
preInlineUnconditionally
~~~~~~~~~~~~~~~~~~~~~~~~
@preInlineUnconditionally@ examines a bndr to see if it is used just
once in a completely safe way, so that it is safe to discard the
binding inline its RHS at the (unique) usage site, REGARDLESS of how
big the RHS might be. If this is the case we don't simplify the RHS
first, but just inline it un-simplified.
This is much better than first simplifying a perhaps-huge RHS and then
inlining and re-simplifying it. Indeed, it can be at least quadratically
better. Consider
x1 = e1
x2 = e2[x1]
x3 = e3[x2]
...etc...
xN = eN[xN-1]
We may end up simplifying e1 N times, e2 N-1 times, e3 N-3 times etc.
This can happen with cascades of functions too:
f1 = \x1.e1
f2 = \xs.e2[f1]
f3 = \xs.e3[f3]
...etc...
THE MAIN INVARIANT is this:
---- preInlineUnconditionally invariant -----
IF preInlineUnconditionally chooses to inline x = <rhs>
THEN doing the inlining should not change the occurrence
info for the free vars of <rhs>
----------------------------------------------
For example, it's tempting to look at trivial binding like
x = y
and inline it unconditionally. But suppose x is used many times,
but this is the unique occurrence of y. Then inlining x would change
y's occurrence info, which breaks the invariant. It matters: y
might have a BIG rhs, which will now be dup'd at every occurrenc of x.
Evne RHSs labelled InlineMe aren't caught here, because there might be
no benefit from inlining at the call site.
[Sept 01] Don't unconditionally inline a top-level thing, because that
can simply make a static thing into something built dynamically. E.g.
x = (a,b)
main = \s -> h x
[Remember that we treat \s as a one-shot lambda.] No point in
inlining x unless there is something interesting about the call site.
But watch out: if you aren't careful, some useful foldr/build fusion
can be lost (most notably in spectral/hartel/parstof) because the
foldr didn't see the build. Doing the dynamic allocation isn't a big
deal, in fact, but losing the fusion can be. But the right thing here
seems to be to do a callSiteInline based on the fact that there is
something interesting about the call site (it's strict). Hmm. That
seems a bit fragile.
Conclusion: inline top level things gaily until Phase 0 (the last
phase), at which point don't.
\begin{code}
preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
preInlineUnconditionally env top_lvl bndr rhs
| not active = False
| opt_SimplNoPreInlining = False
| otherwise = case idOccInfo bndr of
IAmDead -> True -- Happens in ((\x.1) v)
OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
other -> False
where
phase = getMode env
active = case phase of
SimplGently -> isAlwaysActive prag
SimplPhase n -> isActive n prag
prag = idInlinePragma bndr
try_once in_lam int_cxt -- There's one textual occurrence
| not in_lam = isNotTopLevel top_lvl || early_phase
| otherwise = int_cxt && canInlineInLam rhs
-- Be very careful before inlining inside a lambda, becuase (a) we must not
-- invalidate occurrence information, and (b) we want to avoid pushing a
-- single allocation (here) into multiple allocations (inside lambda).
-- Inlining a *function* with a single *saturated* call would be ok, mind you.
-- || (if is_cheap && not (canInlineInLam rhs) then pprTrace "preinline" (ppr bndr <+> ppr rhs) ok else ok)
-- where
-- is_cheap = exprIsCheap rhs
-- ok = is_cheap && int_cxt
-- int_cxt The context isn't totally boring
-- E.g. let f = \ab.BIG in \y. map f xs
-- Don't want to substitute for f, because then we allocate
-- its closure every time the \y is called
-- But: let f = \ab.BIG in \y. map (f y) xs
-- Now we do want to substitute for f, even though it's not
-- saturated, because we're going to allocate a closure for
-- (f y) every time round the loop anyhow.
-- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
-- so substituting rhs inside a lambda doesn't change the occ info.
-- Sadly, not quite the same as exprIsHNF.
canInlineInLam (Lit l) = True
canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
canInlineInLam (Note _ e) = canInlineInLam e
canInlineInLam _ = False
early_phase = case phase of
SimplPhase 0 -> False
other -> True
-- If we don't have this early_phase test, consider
-- x = length [1,2,3]
-- The full laziness pass carefully floats all the cons cells to
-- top level, and preInlineUnconditionally floats them all back in.
-- Result is (a) static allocation replaced by dynamic allocation
-- (b) many simplifier iterations because this tickles
-- a related problem; only one inlining per pass
--
-- On the other hand, I have seen cases where top-level fusion is
-- lost if we don't inline top level thing (e.g. string constants)
-- Hence the test for phase zero (which is the phase for all the final
-- simplifications). Until phase zero we take no special notice of
-- top level things, but then we become more leery about inlining
-- them.
\end{code}
postInlineUnconditionally
~~~~~~~~~~~~~~~~~~~~~~~~~
@postInlineUnconditionally@ decides whether to unconditionally inline
a thing based on the form of its RHS; in particular if it has a
trivial RHS. If so, we can inline and discard the binding altogether.
NB: a loop breaker has must_keep_binding = True and non-loop-breakers
only have *forward* references Hence, it's safe to discard the binding
NOTE: This isn't our last opportunity to inline. We're at the binding
site right now, and we'll get another opportunity when we get to the
ocurrence(s)
Note that we do this unconditional inlining only for trival RHSs.
Don't inline even WHNFs inside lambdas; doing so may simply increase
allocation when the function is called. This isn't the last chance; see
NOTE above.
NB: Even inline pragmas (e.g. IMustBeINLINEd) are ignored here Why?
Because we don't even want to inline them into the RHS of constructor
arguments. See NOTE above
NB: At one time even NOINLINE was ignored here: if the rhs is trivial
it's best to inline it anyway. We often get a=E; b=a from desugaring,
with both a and b marked NOINLINE. But that seems incompatible with
our new view that inlining is like a RULE, so I'm sticking to the 'active'
story for now.
\begin{code}
postInlineUnconditionally
:: SimplEnv -> TopLevelFlag
-> InId -- The binder (an OutId would be fine too)
-> OccInfo -- From the InId
-> OutExpr
-> Unfolding
-> Bool
postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
| not active = False
| isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, dont' inline
-- because it might be referred to "earlier"
| isExportedId bndr = False
| exprIsTrivial rhs = True
| otherwise
= case occ_info of
-- The point of examining occ_info here is that for *non-values*
-- that occur outside a lambda, the call-site inliner won't have
-- a chance (becuase it doesn't know that the thing
-- only occurs once). The pre-inliner won't have gotten
-- it either, if the thing occurs in more than one branch
-- So the main target is things like
-- let x = f y in
-- case v of
-- True -> case x of ...
-- False -> case x of ...
-- I'm not sure how important this is in practice
OneOcc in_lam one_br int_cxt -- OneOcc => no code-duplication issue
-> smallEnoughToInline unfolding -- Small enough to dup
-- ToDo: consider discount on smallEnoughToInline if int_cxt is true
--
-- NB: Do NOT inline arbitrarily big things, even if one_br is True
-- Reason: doing so risks exponential behaviour. We simplify a big
-- expression, inline it, and simplify it again. But if the
-- very same thing happens in the big expression, we get
-- exponential cost!
-- PRINCIPLE: when we've already simplified an expression once,
-- make sure that we only inline it if it's reasonably small.
&& ((isNotTopLevel top_lvl && not in_lam) ||
-- But outside a lambda, we want to be reasonably aggressive
-- about inlining into multiple branches of case
-- e.g. let x = <non-value>
-- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
-- Inlining can be a big win if C3 is the hot-spot, even if
-- the uses in C1, C2 are not 'interesting'
-- An example that gets worse if you add int_cxt here is 'clausify'
(isCheapUnfolding unfolding && int_cxt))
-- isCheap => acceptable work duplication; in_lam may be true
-- int_cxt to prevent us inlining inside a lambda without some
-- good reason. See the notes on int_cxt in preInlineUnconditionally
IAmDead -> True -- This happens; for example, the case_bndr during case of
-- known constructor: case (a,b) of x { (p,q) -> ... }
-- Here x isn't mentioned in the RHS, so we don't want to
-- create the (dead) let-binding let x = (a,b) in ...
other -> False
-- Here's an example that we don't handle well:
-- let f = if b then Left (\x.BIG) else Right (\y.BIG)
-- in \y. ....case f of {...} ....
-- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
-- But
-- * We can't preInlineUnconditionally because that woud invalidate
-- the occ info for b.
-- * We can't postInlineUnconditionally because the RHS is big, and
-- that risks exponential behaviour
-- * We can't call-site inline, because the rhs is big
-- Alas!
where
active = case getMode env of
SimplGently -> isAlwaysActive prag
SimplPhase n -> isActive n prag
prag = idInlinePragma bndr
activeInline :: SimplEnv -> OutId -> Bool
activeInline env id
= case getMode env of
SimplGently -> False
-- No inlining at all when doing gentle stuff,
-- except for local things that occur once
-- The reason is that too little clean-up happens if you
-- don't inline use-once things. Also a bit of inlining is *good* for
-- full laziness; it can expose constant sub-expressions.
-- Example in spectral/mandel/Mandel.hs, where the mandelset
-- function gets a useful let-float if you inline windowToViewport
-- NB: we used to have a second exception, for data con wrappers.
-- On the grounds that we use gentle mode for rule LHSs, and
-- they match better when data con wrappers are inlined.
-- But that only really applies to the trivial wrappers (like (:)),
-- and they are now constructed as Compulsory unfoldings (in MkId)
-- so they'll happen anyway.
SimplPhase n -> isActive n prag
where
prag = idInlinePragma id
activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
-- Nothing => No rules at all
activeRule dflags env
| not (dopt Opt_RewriteRules dflags)
= Nothing -- Rewriting is off
| otherwise
= case getMode env of
SimplGently -> Just isAlwaysActive
-- Used to be Nothing (no rules in gentle mode)
-- Main motivation for changing is that I wanted
-- lift String ===> ...
-- to work in Template Haskell when simplifying
-- splices, so we get simpler code for literal strings
SimplPhase n -> Just (isActive n)
\end{code}
%************************************************************************
%* *
Rebuilding a lambda
%* *
%************************************************************************
\begin{code}
mkLam :: [OutBndr] -> OutExpr -> SimplM OutExpr
-- mkLam tries three things
-- a) eta reduction, if that gives a trivial expression
-- b) eta expansion [only if there are some value lambdas]
mkLam bndrs body
= do { dflags <- getDOptsSmpl
; mkLam' dflags bndrs body }
where
mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
mkLam' dflags bndrs (Cast body@(Lam _ _) co)
-- Note [Casts and lambdas]
= do { lam <- mkLam' dflags (bndrs ++ bndrs') body'
; return (mkCoerce (mkPiTypes bndrs co) lam) }
where
(bndrs',body') = collectBinders body
mkLam' dflags bndrs body
| dopt Opt_DoEtaReduction dflags,
Just etad_lam <- tryEtaReduce bndrs body
= do { tick (EtaReduction (head bndrs))
; return etad_lam }
| dopt Opt_DoLambdaEtaExpansion dflags,
any isRuntimeVar bndrs
= do { body' <- tryEtaExpansion dflags body
; return (mkLams bndrs body') }
| otherwise
= returnSmpl (mkLams bndrs body)
\end{code}
Note [Casts and lambdas]
~~~~~~~~~~~~~~~~~~~~~~~~
Consider
(\x. (\y. e) `cast` g1) `cast` g2
There is a danger here that the two lambdas look separated, and the
full laziness pass might float an expression to between the two.
So this equation in mkLam' floats the g1 out, thus:
(\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
where x:tx.
In general, this floats casts outside lambdas, where (I hope) they might meet
and cancel with some other cast.
-- c) floating lets out through big lambdas
-- [only if all tyvar lambdas, and only if this lambda
-- is the RHS of a let]
{- Sept 01: I'm experimenting with getting the
full laziness pass to float out past big lambdsa
| all isTyVar bndrs, -- Only for big lambdas
contIsRhs cont -- Only try the rhs type-lambda floating
-- if this is indeed a right-hand side; otherwise
-- we end up floating the thing out, only for float-in
-- to float it right back in again!
= tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
returnSmpl (floats, mkLams bndrs body')
-}
%************************************************************************
%* *
\subsection{Eta expansion and reduction}
%* *
%************************************************************************
We try for eta reduction here, but *only* if we get all the
way to an exprIsTrivial expression.
We don't want to remove extra lambdas unless we are going
to avoid allocating this thing altogether
\begin{code}
tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
tryEtaReduce bndrs body
-- We don't use CoreUtils.etaReduce, because we can be more
-- efficient here:
-- (a) we already have the binders
-- (b) we can do the triviality test before computing the free vars
= go (reverse bndrs) body
where
go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
go [] fun | ok_fun fun = Just fun -- Success!
go _ _ = Nothing -- Failure!
ok_fun fun = exprIsTrivial fun
&& not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
&& (exprIsHNF fun || all ok_lam bndrs)
ok_lam v = isTyVar v || isDictId v
-- The exprIsHNF is because eta reduction is not
-- valid in general: \x. bot /= bot
-- So we need to be sure that the "fun" is a value.
--
-- However, we always want to reduce (/\a -> f a) to f
-- This came up in a RULE: foldr (build (/\a -> g a))
-- did not match foldr (build (/\b -> ...something complex...))
-- The type checker can insert these eta-expanded versions,
-- with both type and dictionary lambdas; hence the slightly
-- ad-hoc isDictTy
ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
\end{code}
Try eta expansion for RHSs
We go for:
f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
(n >= 0)
where (in both cases)
* The xi can include type variables
* The yi are all value variables
* N is a NORMAL FORM (i.e. no redexes anywhere)
wanting a suitable number of extra args.
We may have to sandwich some coerces between the lambdas
to make the types work. exprEtaExpandArity looks through coerces
when computing arity; and etaExpand adds the coerces as necessary when
actually computing the expansion.
\begin{code}
tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
-- There is at least one runtime binder in the binders
tryEtaExpansion dflags body
= getUniquesSmpl `thenSmpl` \ us ->
returnSmpl (etaExpand fun_arity us body (exprType body))
where
fun_arity = exprEtaExpandArity dflags body
\end{code}
%************************************************************************
%* *
\subsection{Floating lets out of big lambdas}
%* *
%************************************************************************
tryRhsTyLam tries this transformation, when the big lambda appears as
the RHS of a let(rec) binding:
/\abc -> let(rec) x = e in b
==>
let(rec) x' = /\abc -> let x = x' a b c in e
in
/\abc -> let x = x' a b c in b
This is good because it can turn things like:
let f = /\a -> letrec g = ... g ... in g
into
letrec g' = /\a -> ... g' a ...
in
let f = /\ a -> g' a
which is better. In effect, it means that big lambdas don't impede
let-floating.
This optimisation is CRUCIAL in eliminating the junk introduced by
desugaring mutually recursive definitions. Don't eliminate it lightly!
So far as the implementation is concerned:
Invariant: go F e = /\tvs -> F e
Equalities:
go F (Let x=e in b)
= Let x' = /\tvs -> F e
in
go G b
where
G = F . Let x = x' tvs
go F (Letrec xi=ei in b)
= Letrec {xi' = /\tvs -> G ei}
in
go G b
where
G = F . Let {xi = xi' tvs}
[May 1999] If we do this transformation *regardless* then we can
end up with some pretty silly stuff. For example,
let
st = /\ s -> let { x1=r1 ; x2=r2 } in ...
in ..
becomes
let y1 = /\s -> r1
y2 = /\s -> r2
st = /\s -> ...[y1 s/x1, y2 s/x2]
in ..
Unless the "..." is a WHNF there is really no point in doing this.
Indeed it can make things worse. Suppose x1 is used strictly,
and is of the form