-
Notifications
You must be signed in to change notification settings - Fork 1
/
PE.v
1610 lines (1431 loc) · 64.4 KB
/
PE.v
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
(** * PE: Partial Evaluation *)
(* $Date: 2013-07-17 16:19:11 -0400 (Wed, 17 Jul 2013) $ *)
(* Chapter author/maintainer: Chung-chieh Shan *)
(** Equiv.v introduced constant folding as an example of a program
transformation and proved that it preserves the meaning of the
program. Constant folding operates on manifest constants such
as [ANum] expressions. For example, it simplifies the command
[Y ::= APlus (ANum 3) (ANum 1)] to the command [Y ::= ANum 4].
However, it does not propagate known constants along data flow.
For example, it does not simplify the sequence
X ::= ANum 3;; Y ::= APlus (AId X) (ANum 1)
to
X ::= ANum 3;; Y ::= ANum 4
because it forgets that [X] is [3] by the time it gets to [Y].
We naturally want to enhance constant folding so that it
propagates known constants and uses them to simplify programs.
Doing so constitutes a rudimentary form of _partial evaluation_.
As we will see, partial evaluation is so called because it is
like running a program, except only part of the program can be
evaluated because only part of the input to the program is known.
For example, we can only simplify the program
X ::= ANum 3;; Y ::= AMinus (APlus (AId X) (ANum 1)) (AId Y)
to
X ::= ANum 3;; Y ::= AMinus (ANum 4) (AId Y)
without knowing the initial value of [Y]. *)
Require Export Imp.
Require Import FunctionalExtensionality.
(* ####################################################### *)
(** * Generalizing Constant Folding *)
(** The starting point of partial evaluation is to represent our
partial knowledge about the state. For example, between the two
assignments above, the partial evaluator may know only that [X] is
[3] and nothing about any other variable. *)
(** ** Partial States *)
(** Conceptually speaking, we can think of such partial states as the
type [id -> option nat] (as opposed to the type [id -> nat] of
concrete, full states). However, in addition to looking up and
updating the values of individual variables in a partial state, we
may also want to compare two partial states to see if and where
they differ, to handle conditional control flow. It is not possible
to compare two arbitrary functions in this way, so we represent
partial states in a more concrete format: as a list of [id * nat]
pairs. *)
Definition pe_state := list (id * nat).
(** The idea is that a variable [id] appears in the list if and only
if we know its current [nat] value. The [pe_lookup] function thus
interprets this concrete representation. (If the same variable
[id] appears multiple times in the list, the first occurrence
wins, but we will define our partial evaluator to never construct
such a [pe_state].) *)
Fixpoint pe_lookup (pe_st : pe_state) (V:id) : option nat :=
match pe_st with
| [] => None
| (V',n')::pe_st => if eq_id_dec V V' then Some n'
else pe_lookup pe_st V
end.
(** For example, [empty_pe_state] represents complete ignorance about
every variable -- the function that maps every [id] to [None]. *)
Definition empty_pe_state : pe_state := [].
(** More generally, if the [list] representing a [pe_state] does not
contain some [id], then that [pe_state] must map that [id] to
[None]. Before we prove this fact, we first define a useful
tactic for reasoning with [id] equality. The tactic
compare V V' SCase
means to reason by cases over [eq_id_dec V V'].
In the case where [V = V'], the tactic
substitutes [V] for [V'] throughout. *)
Tactic Notation "compare" ident(i) ident(j) ident(c) :=
let H := fresh "Heq" i j in
destruct (eq_id_dec i j);
[ Case_aux c "equal"; subst j
| Case_aux c "not equal" ].
Theorem pe_domain: forall pe_st V n,
pe_lookup pe_st V = Some n ->
In V (map (@fst _ _) pe_st).
Proof. intros pe_st V n H. induction pe_st as [| [V' n'] pe_st].
Case "[]". inversion H.
Case "::". simpl in H. simpl. compare V V' SCase; auto. Qed.
(** *** Aside on [In].
We will make heavy use of the [In] predicate from the standard library.
[In] is equivalent to the [appears_in] predicate introduced in Logic.v, but
defined using a [Fixpoint] rather than an [Inductive]. *)
Print In.
(* ===> Fixpoint In {A:Type} (a: A) (l:list A) : Prop :=
match l with
| [] => False
| b :: m => b = a \/ In a m
end
: forall A : Type, A -> list A -> Prop *)
(** [In] comes with various useful lemmas. *)
Check in_or_app.
(* ===> in_or_app: forall (A : Type) (l m : list A) (a : A),
In a l \/ In a m -> In a (l ++ m) *)
Check filter_In.
(* ===> filter_In : forall (A : Type) (f : A -> bool) (x : A) (l : list A),
In x (filter f l) <-> In x l /\ f x = true *)
Check in_dec.
(* ===> in_dec : forall A : Type,
(forall x y : A, {x = y} + {x <> y}) ->
forall (a : A) (l : list A), {In a l} + {~ In a l}] *)
(** Note that we can compute with [in_dec], just as with [eq_id_dec]. *)
(** ** Arithmetic Expressions *)
(** Partial evaluation of [aexp] is straightforward -- it is basically
the same as constant folding, [fold_constants_aexp], except that
sometimes the partial state tells us the current value of a
variable and we can replace it by a constant expression. *)
Fixpoint pe_aexp (pe_st : pe_state) (a : aexp) : aexp :=
match a with
| ANum n => ANum n
| AId i => match pe_lookup pe_st i with (* <----- NEW *)
| Some n => ANum n
| None => AId i
end
| APlus a1 a2 =>
match (pe_aexp pe_st a1, pe_aexp pe_st a2) with
| (ANum n1, ANum n2) => ANum (n1 + n2)
| (a1', a2') => APlus a1' a2'
end
| AMinus a1 a2 =>
match (pe_aexp pe_st a1, pe_aexp pe_st a2) with
| (ANum n1, ANum n2) => ANum (n1 - n2)
| (a1', a2') => AMinus a1' a2'
end
| AMult a1 a2 =>
match (pe_aexp pe_st a1, pe_aexp pe_st a2) with
| (ANum n1, ANum n2) => ANum (n1 * n2)
| (a1', a2') => AMult a1' a2'
end
end.
(** This partial evaluator folds constants but does not apply the
associativity of addition. *)
Example test_pe_aexp1:
pe_aexp [(X,3)] (APlus (APlus (AId X) (ANum 1)) (AId Y))
= APlus (ANum 4) (AId Y).
Proof. reflexivity. Qed.
Example text_pe_aexp2:
pe_aexp [(Y,3)] (APlus (APlus (AId X) (ANum 1)) (AId Y))
= APlus (APlus (AId X) (ANum 1)) (ANum 3).
Proof. reflexivity. Qed.
(** Now, in what sense is [pe_aexp] correct? It is reasonable to
define the correctness of [pe_aexp] as follows: whenever a full
state [st:state] is _consistent_ with a partial state
[pe_st:pe_state] (in other words, every variable to which [pe_st]
assigns a value is assigned the same value by [st]), evaluating
[a] and evaluating [pe_aexp pe_st a] in [st] yields the same
result. This statement is indeed true. *)
Definition pe_consistent (st:state) (pe_st:pe_state) :=
forall V n, Some n = pe_lookup pe_st V -> st V = n.
Theorem pe_aexp_correct_weak: forall st pe_st, pe_consistent st pe_st ->
forall a, aeval st a = aeval st (pe_aexp pe_st a).
Proof. unfold pe_consistent. intros st pe_st H a.
aexp_cases (induction a) Case; simpl;
try reflexivity;
try (destruct (pe_aexp pe_st a1);
destruct (pe_aexp pe_st a2);
rewrite IHa1; rewrite IHa2; reflexivity).
(* Compared to fold_constants_aexp_sound,
the only interesting case is AId *)
Case "AId".
remember (pe_lookup pe_st i) as l. destruct l.
SCase "Some". rewrite H with (n:=n) by apply Heql. reflexivity.
SCase "None". reflexivity.
Qed.
(** However, we will soon want our partial evaluator to remove
assignments. For example, it will simplify
X ::= ANum 3;; Y ::= AMinus (AId X) (AId Y);; X ::= ANum 4
to just
Y ::= AMinus (ANum 3) (AId Y);; X ::= ANum 4
by delaying the assignment to [X] until the end. To accomplish
this simplification, we need the result of partial evaluating
pe_aexp [(X,3)] (AMinus (AId X) (AId Y))
to be equal to [AMinus (ANum 3) (AId Y)] and _not_ the original
expression [AMinus (AId X) (AId Y)]. After all, it would be
incorrect, not just inefficient, to transform
X ::= ANum 3;; Y ::= AMinus (AId X) (AId Y);; X ::= ANum 4
to
Y ::= AMinus (AId X) (AId Y);; X ::= ANum 4
even though the output expressions [AMinus (ANum 3) (AId Y)] and
[AMinus (AId X) (AId Y)] both satisfy the correctness criterion
that we just proved. Indeed, if we were to just define [pe_aexp
pe_st a = a] then the theorem [pe_aexp_correct'] would already
trivially hold.
Instead, we want to prove that the [pe_aexp] is correct in a
stronger sense: evaluating the expression produced by partial
evaluation ([aeval st (pe_aexp pe_st a)]) must not depend on those
parts of the full state [st] that are already specified in the
partial state [pe_st]. To be more precise, let us define a
function [pe_override], which updates [st] with the contents of
[pe_st]. In other words, [pe_override] carries out the
assignments listed in [pe_st] on top of [st]. *)
Fixpoint pe_override (st:state) (pe_st:pe_state) : state :=
match pe_st with
| [] => st
| (V,n)::pe_st => update (pe_override st pe_st) V n
end.
Example test_pe_override:
pe_override (update empty_state Y 1) [(X,3);(Z,2)]
= update (update (update empty_state Y 1) Z 2) X 3.
Proof. reflexivity. Qed.
(** Although [pe_override] operates on a concrete [list] representing
a [pe_state], its behavior is defined entirely by the [pe_lookup]
interpretation of the [pe_state]. *)
Theorem pe_override_correct: forall st pe_st V0,
pe_override st pe_st V0 =
match pe_lookup pe_st V0 with
| Some n => n
| None => st V0
end.
Proof. intros. induction pe_st as [| [V n] pe_st]. reflexivity.
simpl in *. unfold update.
compare V0 V Case; auto. rewrite eq_id; auto. rewrite neq_id; auto. Qed.
(** We can relate [pe_consistent] to [pe_override] in two ways.
First, overriding a state with a partial state always gives a
state that is consistent with the partial state. Second, if a
state is already consistent with a partial state, then overriding
the state with the partial state gives the same state. *)
Theorem pe_override_consistent: forall st pe_st,
pe_consistent (pe_override st pe_st) pe_st.
Proof. intros st pe_st V n H. rewrite pe_override_correct.
destruct (pe_lookup pe_st V); inversion H. reflexivity. Qed.
Theorem pe_consistent_override: forall st pe_st,
pe_consistent st pe_st -> forall V, st V = pe_override st pe_st V.
Proof. intros st pe_st H V. rewrite pe_override_correct.
remember (pe_lookup pe_st V) as l. destruct l; auto. Qed.
(** Now we can state and prove that [pe_aexp] is correct in the
stronger sense that will help us define the rest of the partial
evaluator.
Intuitively, running a program using partial evaluation is a
two-stage process. In the first, _static_ stage, we partially
evaluate the given program with respect to some partial state to
get a _residual_ program. In the second, _dynamic_ stage, we
evaluate the residual program with respect to the rest of the
state. This dynamic state provides values for those variables
that are unknown in the static (partial) state. Thus, the
residual program should be equivalent to _prepending_ the
assignments listed in the partial state to the original program. *)
Theorem pe_aexp_correct: forall (pe_st:pe_state) (a:aexp) (st:state),
aeval (pe_override st pe_st) a = aeval st (pe_aexp pe_st a).
Proof.
intros pe_st a st.
aexp_cases (induction a) Case; simpl;
try reflexivity;
try (destruct (pe_aexp pe_st a1);
destruct (pe_aexp pe_st a2);
rewrite IHa1; rewrite IHa2; reflexivity).
(* Compared to fold_constants_aexp_sound, the only
interesting case is AId. *)
rewrite pe_override_correct. destruct (pe_lookup pe_st i); reflexivity.
Qed.
(** ** Boolean Expressions *)
(** The partial evaluation of boolean expressions is similar. In
fact, it is entirely analogous to the constant folding of boolean
expressions, because our language has no boolean variables. *)
Fixpoint pe_bexp (pe_st : pe_state) (b : bexp) : bexp :=
match b with
| BTrue => BTrue
| BFalse => BFalse
| BEq a1 a2 =>
match (pe_aexp pe_st a1, pe_aexp pe_st a2) with
| (ANum n1, ANum n2) => if beq_nat n1 n2 then BTrue else BFalse
| (a1', a2') => BEq a1' a2'
end
| BLe a1 a2 =>
match (pe_aexp pe_st a1, pe_aexp pe_st a2) with
| (ANum n1, ANum n2) => if ble_nat n1 n2 then BTrue else BFalse
| (a1', a2') => BLe a1' a2'
end
| BNot b1 =>
match (pe_bexp pe_st b1) with
| BTrue => BFalse
| BFalse => BTrue
| b1' => BNot b1'
end
| BAnd b1 b2 =>
match (pe_bexp pe_st b1, pe_bexp pe_st b2) with
| (BTrue, BTrue) => BTrue
| (BTrue, BFalse) => BFalse
| (BFalse, BTrue) => BFalse
| (BFalse, BFalse) => BFalse
| (b1', b2') => BAnd b1' b2'
end
end.
Example test_pe_bexp1:
pe_bexp [(X,3)] (BNot (BLe (AId X) (ANum 3)))
= BFalse.
Proof. reflexivity. Qed.
Example test_pe_bexp2: forall b,
b = BNot (BLe (AId X) (APlus (AId X) (ANum 1))) ->
pe_bexp [] b = b.
Proof. intros b H. rewrite -> H. reflexivity. Qed.
(** The correctness of [pe_bexp] is analogous to the correctness of
[pe_aexp] above. *)
Theorem pe_bexp_correct: forall (pe_st:pe_state) (b:bexp) (st:state),
beval (pe_override st pe_st) b = beval st (pe_bexp pe_st b).
Proof.
intros pe_st b st.
bexp_cases (induction b) Case; simpl;
try reflexivity;
try (remember (pe_aexp pe_st a) as a';
remember (pe_aexp pe_st a0) as a0';
assert (Ha: aeval (pe_override st pe_st) a = aeval st a');
assert (Ha0: aeval (pe_override st pe_st) a0 = aeval st a0');
try (subst; apply pe_aexp_correct);
destruct a'; destruct a0'; rewrite Ha; rewrite Ha0;
simpl; try destruct (beq_nat n n0); try destruct (ble_nat n n0);
reflexivity);
try (destruct (pe_bexp pe_st b); rewrite IHb; reflexivity);
try (destruct (pe_bexp pe_st b1);
destruct (pe_bexp pe_st b2);
rewrite IHb1; rewrite IHb2; reflexivity).
Qed.
(* ####################################################### *)
(** * Partial Evaluation of Commands, Without Loops *)
(** What about the partial evaluation of commands? The analogy
between partial evaluation and full evaluation continues: Just as
full evaluation of a command turns an initial state into a final
state, partial evaluation of a command turns an initial partial
state into a final partial state. The difference is that, because
the state is partial, some parts of the command may not be
executable at the static stage. Therefore, just as [pe_aexp]
returns a residual [aexp] and [pe_bexp] returns a residual [bexp]
above, partially evaluating a command yields a residual command.
Another way in which our partial evaluator is similar to a full
evaluator is that it does not terminate on all commands. It is
not hard to build a partial evaluator that terminates on all
commands; what is hard is building a partial evaluator that
terminates on all commands yet automatically performs desired
optimizations such as unrolling loops. Often a partial evaluator
can be coaxed into terminating more often and performing more
optimizations by writing the source program differently so that
the separation between static and dynamic information becomes more
apparent. Such coaxing is the art of _binding-time improvement_.
The binding time of a variable tells when its value is known --
either "static", or "dynamic."
Anyway, for now we will just live with the fact that our partial
evaluator is not a total function from the source command and the
initial partial state to the residual command and the final
partial state. To model this non-termination, just as with the
full evaluation of commands, we use an inductively defined
relation. We write
c1 / st || c1' / st'
to mean that partially evaluating the source command [c1] in the
initial partial state [st] yields the residual command [c1'] and
the final partial state [st']. For example, we want something like
(X ::= ANum 3 ;; Y ::= AMult (AId Z) (APlus (AId X) (AId X)))
/ [] || (Y ::= AMult (AId Z) (ANum 6)) / [(X,3)]
to hold. The assignment to [X] appears in the final partial state,
not the residual command. *)
(** ** Assignment *)
(** Let's start by considering how to partially evaluate an
assignment. The two assignments in the source program above needs
to be treated differently. The first assignment [X ::= ANum 3],
is _static_: its right-hand-side is a constant (more generally,
simplifies to a constant), so we should update our partial state
at [X] to [3] and produce no residual code. (Actually, we produce
a residual [SKIP].) The second assignment [Y ::= AMult (AId Z)
(APlus (AId X) (AId X))] is _dynamic_: its right-hand-side does
not simplify to a constant, so we should leave it in the residual
code and remove [Y], if present, from our partial state. To
implement these two cases, we define the functions [pe_add] and
[pe_remove]. Like [pe_override] above, these functions operate on
a concrete [list] representing a [pe_state], but the theorems
[pe_add_correct] and [pe_remove_correct] specify their behavior by
the [pe_lookup] interpretation of the [pe_state]. *)
Fixpoint pe_remove (pe_st:pe_state) (V:id) : pe_state :=
match pe_st with
| [] => []
| (V',n')::pe_st => if eq_id_dec V V' then pe_remove pe_st V
else (V',n') :: pe_remove pe_st V
end.
Theorem pe_remove_correct: forall pe_st V V0,
pe_lookup (pe_remove pe_st V) V0
= if eq_id_dec V V0 then None else pe_lookup pe_st V0.
Proof. intros pe_st V V0. induction pe_st as [| [V' n'] pe_st].
Case "[]". destruct (eq_id_dec V V0); reflexivity.
Case "::". simpl. compare V V' SCase.
SCase "equal". rewrite IHpe_st.
destruct (eq_id_dec V V0). reflexivity. rewrite neq_id; auto.
SCase "not equal". simpl. compare V0 V' SSCase.
SSCase "equal". rewrite neq_id; auto.
SSCase "not equal". rewrite IHpe_st. reflexivity.
Qed.
Definition pe_add (pe_st:pe_state) (V:id) (n:nat) : pe_state :=
(V,n) :: pe_remove pe_st V.
Theorem pe_add_correct: forall pe_st V n V0,
pe_lookup (pe_add pe_st V n) V0
= if eq_id_dec V V0 then Some n else pe_lookup pe_st V0.
Proof. intros pe_st V n V0. unfold pe_add. simpl.
compare V V0 Case.
Case "equal". rewrite eq_id; auto.
Case "not equal". rewrite pe_remove_correct. repeat rewrite neq_id; auto.
Qed.
(** We will use the two theorems below to show that our partial
evaluator correctly deals with dynamic assignments and static
assignments, respectively. *)
Theorem pe_override_update_remove: forall st pe_st V n,
update (pe_override st pe_st) V n =
pe_override (update st V n) (pe_remove pe_st V).
Proof. intros st pe_st V n. apply functional_extensionality. intros V0.
unfold update. rewrite !pe_override_correct. rewrite pe_remove_correct.
destruct (eq_id_dec V V0); reflexivity. Qed.
Theorem pe_override_update_add: forall st pe_st V n,
update (pe_override st pe_st) V n =
pe_override st (pe_add pe_st V n).
Proof. intros st pe_st V n. apply functional_extensionality. intros V0.
unfold update. rewrite !pe_override_correct. rewrite pe_add_correct.
destruct (eq_id_dec V V0); reflexivity. Qed.
(** ** Conditional *)
(** Trickier than assignments to partially evaluate is the
conditional, [IFB b1 THEN c1 ELSE c2 FI]. If [b1] simplifies to
[BTrue] or [BFalse] then it's easy: we know which branch will be
taken, so just take that branch. If [b1] does not simplify to a
constant, then we need to take both branches, and the final
partial state may differ between the two branches!
The following program illustrates the difficulty:
X ::= ANum 3;;
IFB BLe (AId Y) (ANum 4) THEN
Y ::= ANum 4;;
IFB BEq (AId X) (AId Y) THEN Y ::= ANum 999 ELSE SKIP FI
ELSE SKIP FI
Suppose the initial partial state is empty. We don't know
statically how [Y] compares to [4], so we must partially evaluate
both branches of the (outer) conditional. On the [THEN] branch,
we know that [Y] is set to [4] and can even use that knowledge to
simplify the code somewhat. On the [ELSE] branch, we still don't
know the exact value of [Y] at the end. What should the final
partial state and residual program be?
One way to handle such a dynamic conditional is to take the
intersection of the final partial states of the two branches. In
this example, we take the intersection of [(Y,4),(X,3)] and
[(X,3)], so the overall final partial state is [(X,3)]. To
compensate for forgetting that [Y] is [4], we need to add an
assignment [Y ::= ANum 4] to the end of the [THEN] branch. So,
the residual program will be something like
SKIP;;
IFB BLe (AId Y) (ANum 4) THEN
SKIP;;
SKIP;;
Y ::= ANum 4
ELSE SKIP FI
Programming this case in Coq calls for several auxiliary
functions: we need to compute the intersection of two [pe_state]s
and turn their difference into sequences of assignments.
First, we show how to compute whether two [pe_state]s to disagree
at a given variable. In the theorem [pe_disagree_domain], we
prove that two [pe_state]s can only disagree at variables that
appear in at least one of them. *)
Definition pe_disagree_at (pe_st1 pe_st2 : pe_state) (V:id) : bool :=
match pe_lookup pe_st1 V, pe_lookup pe_st2 V with
| Some x, Some y => negb (beq_nat x y)
| None, None => false
| _, _ => true
end.
Theorem pe_disagree_domain: forall (pe_st1 pe_st2 : pe_state) (V:id),
true = pe_disagree_at pe_st1 pe_st2 V ->
In V (map (@fst _ _) pe_st1 ++ map (@fst _ _) pe_st2).
Proof. unfold pe_disagree_at. intros pe_st1 pe_st2 V H.
apply in_or_app.
remember (pe_lookup pe_st1 V) as lookup1.
destruct lookup1 as [n1|]. left. apply pe_domain with n1. auto.
remember (pe_lookup pe_st2 V) as lookup2.
destruct lookup2 as [n2|]. right. apply pe_domain with n2. auto.
inversion H. Qed.
(** We define the [pe_compare] function to list the variables where
two given [pe_state]s disagree. This list is exact, according to
the theorem [pe_compare_correct]: a variable appears on the list
if and only if the two given [pe_state]s disagree at that
variable. Furthermore, we use the [pe_unique] function to
eliminate duplicates from the list. *)
Fixpoint pe_unique (l : list id) : list id :=
match l with
| [] => []
| x::l => x :: filter (fun y => if eq_id_dec x y then false else true) (pe_unique l)
end.
Theorem pe_unique_correct: forall l x,
In x l <-> In x (pe_unique l).
Proof. intros l x. induction l as [| h t]. reflexivity.
simpl in *. split.
Case "->".
intros. inversion H; clear H.
left. assumption.
destruct (eq_id_dec h x).
left. assumption.
right. apply filter_In. split.
apply IHt. assumption.
rewrite neq_id; auto.
Case "<-".
intros. inversion H; clear H.
left. assumption.
apply filter_In in H0. inversion H0. right. apply IHt. assumption.
Qed.
Definition pe_compare (pe_st1 pe_st2 : pe_state) : list id :=
pe_unique (filter (pe_disagree_at pe_st1 pe_st2)
(map (@fst _ _) pe_st1 ++ map (@fst _ _) pe_st2)).
Theorem pe_compare_correct: forall pe_st1 pe_st2 V,
pe_lookup pe_st1 V = pe_lookup pe_st2 V <->
~ In V (pe_compare pe_st1 pe_st2).
Proof. intros pe_st1 pe_st2 V.
unfold pe_compare. rewrite <- pe_unique_correct. rewrite filter_In.
split; intros Heq.
Case "->".
intro. destruct H. unfold pe_disagree_at in H0. rewrite Heq in H0.
destruct (pe_lookup pe_st2 V).
rewrite <- beq_nat_refl in H0. inversion H0.
inversion H0.
Case "<-".
assert (Hagree: pe_disagree_at pe_st1 pe_st2 V = false).
SCase "Proof of assertion".
remember (pe_disagree_at pe_st1 pe_st2 V) as disagree.
destruct disagree; [| reflexivity].
apply pe_disagree_domain in Heqdisagree.
apply ex_falso_quodlibet. apply Heq. split. assumption. reflexivity.
unfold pe_disagree_at in Hagree.
destruct (pe_lookup pe_st1 V) as [n1|];
destruct (pe_lookup pe_st2 V) as [n2|];
try reflexivity; try solve by inversion.
rewrite negb_false_iff in Hagree.
apply beq_nat_true in Hagree. subst. reflexivity. Qed.
(** The intersection of two partial states is the result of removing
from one of them all the variables where the two disagree. We
define the function [pe_removes], in terms of [pe_remove] above,
to perform such a removal of a whole list of variables at once.
The theorem [pe_compare_removes] testifies that the [pe_lookup]
interpretation of the result of this intersection operation is the
same no matter which of the two partial states we remove the
variables from. Because [pe_override] only depends on the
[pe_lookup] interpretation of partial states, [pe_override] also
does not care which of the two partial states we remove the
variables from; that theorem [pe_compare_override] is used in the
correctness proof shortly. *)
Fixpoint pe_removes (pe_st:pe_state) (ids : list id) : pe_state :=
match ids with
| [] => pe_st
| V::ids => pe_remove (pe_removes pe_st ids) V
end.
Theorem pe_removes_correct: forall pe_st ids V,
pe_lookup (pe_removes pe_st ids) V =
if in_dec eq_id_dec V ids then None else pe_lookup pe_st V.
Proof. intros pe_st ids V. induction ids as [| V' ids]. reflexivity.
simpl. rewrite pe_remove_correct. rewrite IHids.
compare V' V Case.
reflexivity.
destruct (in_dec eq_id_dec V ids);
reflexivity.
Qed.
Theorem pe_compare_removes: forall pe_st1 pe_st2 V,
pe_lookup (pe_removes pe_st1 (pe_compare pe_st1 pe_st2)) V =
pe_lookup (pe_removes pe_st2 (pe_compare pe_st1 pe_st2)) V.
Proof. intros pe_st1 pe_st2 V. rewrite !pe_removes_correct.
destruct (in_dec eq_id_dec V (pe_compare pe_st1 pe_st2)).
reflexivity.
apply pe_compare_correct. auto. Qed.
Theorem pe_compare_override: forall pe_st1 pe_st2 st,
pe_override st (pe_removes pe_st1 (pe_compare pe_st1 pe_st2)) =
pe_override st (pe_removes pe_st2 (pe_compare pe_st1 pe_st2)).
Proof. intros. apply functional_extensionality. intros V.
rewrite !pe_override_correct. rewrite pe_compare_removes. reflexivity.
Qed.
(** Finally, we define an [assign] function to turn the difference
between two partial states into a sequence of assignment commands.
More precisely, [assign pe_st ids] generates an assignment command
for each variable listed in [ids]. *)
Fixpoint assign (pe_st : pe_state) (ids : list id) : com :=
match ids with
| [] => SKIP
| V::ids => match pe_lookup pe_st V with
| Some n => (assign pe_st ids;; V ::= ANum n)
| None => assign pe_st ids
end
end.
(** The command generated by [assign] always terminates, because it is
just a sequence of assignments. The (total) function [assigned]
below computes the effect of the command on the (dynamic state).
The theorem [assign_removes] then confirms that the generated
assignments perfectly compensate for removing the variables from
the partial state. *)
Definition assigned (pe_st:pe_state) (ids : list id) (st:state) : state :=
fun V => if in_dec eq_id_dec V ids then
match pe_lookup pe_st V with
| Some n => n
| None => st V
end
else st V.
Theorem assign_removes: forall pe_st ids st,
pe_override st pe_st =
pe_override (assigned pe_st ids st) (pe_removes pe_st ids).
Proof. intros pe_st ids st. apply functional_extensionality. intros V.
rewrite !pe_override_correct. rewrite pe_removes_correct. unfold assigned.
destruct (in_dec eq_id_dec V ids); destruct (pe_lookup pe_st V); reflexivity.
Qed.
Lemma ceval_extensionality: forall c st st1 st2,
c / st || st1 -> (forall V, st1 V = st2 V) -> c / st || st2.
Proof. intros c st st1 st2 H Heq.
apply functional_extensionality in Heq. rewrite <- Heq. apply H. Qed.
Theorem eval_assign: forall pe_st ids st,
assign pe_st ids / st || assigned pe_st ids st.
Proof. intros pe_st ids st. induction ids as [| V ids]; simpl.
Case "[]". eapply ceval_extensionality. apply E_Skip. reflexivity.
Case "V::ids".
remember (pe_lookup pe_st V) as lookup. destruct lookup.
SCase "Some". eapply E_Seq. apply IHids. unfold assigned. simpl.
eapply ceval_extensionality. apply E_Ass. simpl. reflexivity.
intros V0. unfold update. compare V V0 SSCase.
SSCase "equal". rewrite <- Heqlookup. reflexivity.
SSCase "not equal". destruct (in_dec eq_id_dec V0 ids); auto.
SCase "None". eapply ceval_extensionality. apply IHids.
unfold assigned. intros V0. simpl. compare V V0 SSCase.
SSCase "equal". rewrite <- Heqlookup.
destruct (in_dec eq_id_dec V ids); reflexivity.
SSCase "not equal". destruct (in_dec eq_id_dec V0 ids); reflexivity. Qed.
(** ** The Partial Evaluation Relation *)
(** At long last, we can define a partial evaluator for commands
without loops, as an inductive relation! The inequality
conditions in [PE_AssDynamic] and [PE_If] are just to keep the
partial evaluator deterministic; they are not required for
correctness. *)
Reserved Notation "c1 '/' st '||' c1' '/' st'"
(at level 40, st at level 39, c1' at level 39).
Inductive pe_com : com -> pe_state -> com -> pe_state -> Prop :=
| PE_Skip : forall pe_st,
SKIP / pe_st || SKIP / pe_st
| PE_AssStatic : forall pe_st a1 n1 l,
pe_aexp pe_st a1 = ANum n1 ->
(l ::= a1) / pe_st || SKIP / pe_add pe_st l n1
| PE_AssDynamic : forall pe_st a1 a1' l,
pe_aexp pe_st a1 = a1' ->
(forall n, a1' <> ANum n) ->
(l ::= a1) / pe_st || (l ::= a1') / pe_remove pe_st l
| PE_Seq : forall pe_st pe_st' pe_st'' c1 c2 c1' c2',
c1 / pe_st || c1' / pe_st' ->
c2 / pe_st' || c2' / pe_st'' ->
(c1 ;; c2) / pe_st || (c1' ;; c2') / pe_st''
| PE_IfTrue : forall pe_st pe_st' b1 c1 c2 c1',
pe_bexp pe_st b1 = BTrue ->
c1 / pe_st || c1' / pe_st' ->
(IFB b1 THEN c1 ELSE c2 FI) / pe_st || c1' / pe_st'
| PE_IfFalse : forall pe_st pe_st' b1 c1 c2 c2',
pe_bexp pe_st b1 = BFalse ->
c2 / pe_st || c2' / pe_st' ->
(IFB b1 THEN c1 ELSE c2 FI) / pe_st || c2' / pe_st'
| PE_If : forall pe_st pe_st1 pe_st2 b1 c1 c2 c1' c2',
pe_bexp pe_st b1 <> BTrue ->
pe_bexp pe_st b1 <> BFalse ->
c1 / pe_st || c1' / pe_st1 ->
c2 / pe_st || c2' / pe_st2 ->
(IFB b1 THEN c1 ELSE c2 FI) / pe_st
|| (IFB pe_bexp pe_st b1
THEN c1' ;; assign pe_st1 (pe_compare pe_st1 pe_st2)
ELSE c2' ;; assign pe_st2 (pe_compare pe_st1 pe_st2) FI)
/ pe_removes pe_st1 (pe_compare pe_st1 pe_st2)
where "c1 '/' st '||' c1' '/' st'" := (pe_com c1 st c1' st').
Tactic Notation "pe_com_cases" tactic(first) ident(c) :=
first;
[ Case_aux c "PE_Skip"
| Case_aux c "PE_AssStatic" | Case_aux c "PE_AssDynamic"
| Case_aux c "PE_Seq"
| Case_aux c "PE_IfTrue" | Case_aux c "PE_IfFalse" | Case_aux c "PE_If" ].
Hint Constructors pe_com.
Hint Constructors ceval.
(** ** Examples *)
(** Below are some examples of using the partial evaluator. To make
the [pe_com] relation actually usable for automatic partial
evaluation, we would need to define more automation tactics in
Coq. That is not hard to do, but it is not needed here. *)
Example pe_example1:
(X ::= ANum 3 ;; Y ::= AMult (AId Z) (APlus (AId X) (AId X)))
/ [] || (SKIP;; Y ::= AMult (AId Z) (ANum 6)) / [(X,3)].
Proof. eapply PE_Seq. eapply PE_AssStatic. reflexivity.
eapply PE_AssDynamic. reflexivity. intros n H. inversion H. Qed.
Example pe_example2:
(X ::= ANum 3 ;; IFB BLe (AId X) (ANum 4) THEN X ::= ANum 4 ELSE SKIP FI)
/ [] || (SKIP;; SKIP) / [(X,4)].
Proof. eapply PE_Seq. eapply PE_AssStatic. reflexivity.
eapply PE_IfTrue. reflexivity.
eapply PE_AssStatic. reflexivity. Qed.
Example pe_example3:
(X ::= ANum 3;;
IFB BLe (AId Y) (ANum 4) THEN
Y ::= ANum 4;;
IFB BEq (AId X) (AId Y) THEN Y ::= ANum 999 ELSE SKIP FI
ELSE SKIP FI) / []
|| (SKIP;;
IFB BLe (AId Y) (ANum 4) THEN
(SKIP;; SKIP);; (SKIP;; Y ::= ANum 4)
ELSE SKIP;; SKIP FI)
/ [(X,3)].
Proof. erewrite f_equal2 with (f := fun c st => _ / _ || c / st).
eapply PE_Seq. eapply PE_AssStatic. reflexivity.
eapply PE_If; intuition eauto; try solve by inversion.
econstructor. eapply PE_AssStatic. reflexivity.
eapply PE_IfFalse. reflexivity. econstructor.
reflexivity. reflexivity. Qed.
(** ** Correctness of Partial Evaluation *)
(** Finally let's prove that this partial evaluator is correct! *)
Reserved Notation "c' '/' pe_st' '/' st '||' st''"
(at level 40, pe_st' at level 39, st at level 39).
Inductive pe_ceval
(c':com) (pe_st':pe_state) (st:state) (st'':state) : Prop :=
| pe_ceval_intro : forall st',
c' / st || st' ->
pe_override st' pe_st' = st'' ->
c' / pe_st' / st || st''
where "c' '/' pe_st' '/' st '||' st''" := (pe_ceval c' pe_st' st st'').
Hint Constructors pe_ceval.
Theorem pe_com_complete:
forall c pe_st pe_st' c', c / pe_st || c' / pe_st' ->
forall st st'',
(c / pe_override st pe_st || st'') ->
(c' / pe_st' / st || st'').
Proof. intros c pe_st pe_st' c' Hpe.
pe_com_cases (induction Hpe) Case; intros st st'' Heval;
try (inversion Heval; subst;
try (rewrite -> pe_bexp_correct, -> H in *; solve by inversion);
[]);
eauto.
Case "PE_AssStatic". econstructor. econstructor.
rewrite -> pe_aexp_correct. rewrite <- pe_override_update_add.
rewrite -> H. reflexivity.
Case "PE_AssDynamic". econstructor. econstructor. reflexivity.
rewrite -> pe_aexp_correct. rewrite <- pe_override_update_remove.
reflexivity.
Case "PE_Seq".
edestruct IHHpe1. eassumption. subst.
edestruct IHHpe2. eassumption.
eauto.
Case "PE_If". inversion Heval; subst.
SCase "E'IfTrue". edestruct IHHpe1. eassumption.
econstructor. apply E_IfTrue. rewrite <- pe_bexp_correct. assumption.
eapply E_Seq. eassumption. apply eval_assign.
rewrite <- assign_removes. eassumption.
SCase "E_IfFalse". edestruct IHHpe2. eassumption.
econstructor. apply E_IfFalse. rewrite <- pe_bexp_correct. assumption.
eapply E_Seq. eassumption. apply eval_assign.
rewrite -> pe_compare_override.
rewrite <- assign_removes. eassumption.
Qed.
Theorem pe_com_sound:
forall c pe_st pe_st' c', c / pe_st || c' / pe_st' ->
forall st st'',
(c' / pe_st' / st || st'') ->
(c / pe_override st pe_st || st'').
Proof. intros c pe_st pe_st' c' Hpe.
pe_com_cases (induction Hpe) Case;
intros st st'' [st' Heval Heq];
try (inversion Heval; []; subst); auto.
Case "PE_AssStatic". rewrite <- pe_override_update_add. apply E_Ass.
rewrite -> pe_aexp_correct. rewrite -> H. reflexivity.
Case "PE_AssDynamic". rewrite <- pe_override_update_remove. apply E_Ass.
rewrite <- pe_aexp_correct. reflexivity.
Case "PE_Seq". eapply E_Seq; eauto.
Case "PE_IfTrue". apply E_IfTrue.
rewrite -> pe_bexp_correct. rewrite -> H. reflexivity. eauto.
Case "PE_IfFalse". apply E_IfFalse.
rewrite -> pe_bexp_correct. rewrite -> H. reflexivity. eauto.
Case "PE_If".
inversion Heval; subst; inversion H7;
(eapply ceval_deterministic in H8; [| apply eval_assign]); subst.
SCase "E_IfTrue".
apply E_IfTrue. rewrite -> pe_bexp_correct. assumption.
rewrite <- assign_removes. eauto.
SCase "E_IfFalse".
rewrite -> pe_compare_override.
apply E_IfFalse. rewrite -> pe_bexp_correct. assumption.
rewrite <- assign_removes. eauto.
Qed.
(** The main theorem. Thanks to David Menendez for this formulation! *)
Corollary pe_com_correct:
forall c pe_st pe_st' c', c / pe_st || c' / pe_st' ->
forall st st'',
(c / pe_override st pe_st || st'') <->
(c' / pe_st' / st || st'').
Proof. intros c pe_st pe_st' c' H st st''. split.
Case "->". apply pe_com_complete. apply H.
Case "<-". apply pe_com_sound. apply H.
Qed.
(* ####################################################### *)
(** * Partial Evaluation of Loops *)
(** It may seem straightforward at first glance to extend the partial
evaluation relation [pe_com] above to loops. Indeed, many loops
are easy to deal with. Considered this repeated-squaring loop,
for example:
WHILE BLe (ANum 1) (AId X) DO
Y ::= AMult (AId Y) (AId Y);;
X ::= AMinus (AId X) (ANum 1)
END
If we know neither [X] nor [Y] statically, then the entire loop is
dynamic and the residual command should be the same. If we know
[X] but not [Y], then the loop can be unrolled all the way and the
residual command should be
Y ::= AMult (AId Y) (AId Y);;
Y ::= AMult (AId Y) (AId Y);;
Y ::= AMult (AId Y) (AId Y)
if [X] is initially [3] (and finally [0]). In general, a loop is
easy to partially evaluate if the final partial state of the loop
body is equal to the initial state, or if its guard condition is
static.
But there are other loops for which it is hard to express the
residual program we want in Imp. For example, take this program
for checking if [Y] is even or odd:
X ::= ANum 0;;
WHILE BLe (ANum 1) (AId Y) DO
Y ::= AMinus (AId Y) (ANum 1);;
X ::= AMinus (ANum 1) (AId X)
END
The value of [X] alternates between [0] and [1] during the loop.
Ideally, we would like to unroll this loop, not all the way but
_two-fold_, into something like
WHILE BLe (ANum 1) (AId Y) DO
Y ::= AMinus (AId Y) (ANum 1);;
IF BLe (ANum 1) (AId Y) THEN
Y ::= AMinus (AId Y) (ANum 1)
ELSE
X ::= ANum 1;; EXIT
FI
END;;
X ::= ANum 0
Unfortunately, there is no [EXIT] command in Imp. Without
extending the range of control structures available in our
language, the best we can do is to repeat loop-guard tests or add
flag variables. Neither option is terribly attractive.
Still, as a digression, below is an attempt at performing partial
evaluation on Imp commands. We add one more command argument
[c''] to the [pe_com] relation, which keeps track of a loop to
roll up. *)
Module Loop.
Reserved Notation "c1 '/' st '||' c1' '/' st' '/' c''"
(at level 40, st at level 39, c1' at level 39, st' at level 39).
Inductive pe_com : com -> pe_state -> com -> pe_state -> com -> Prop :=
| PE_Skip : forall pe_st,
SKIP / pe_st || SKIP / pe_st / SKIP
| PE_AssStatic : forall pe_st a1 n1 l,
pe_aexp pe_st a1 = ANum n1 ->
(l ::= a1) / pe_st || SKIP / pe_add pe_st l n1 / SKIP
| PE_AssDynamic : forall pe_st a1 a1' l,
pe_aexp pe_st a1 = a1' ->
(forall n, a1' <> ANum n) ->
(l ::= a1) / pe_st || (l ::= a1') / pe_remove pe_st l / SKIP
| PE_Seq : forall pe_st pe_st' pe_st'' c1 c2 c1' c2' c'',
c1 / pe_st || c1' / pe_st' / SKIP ->
c2 / pe_st' || c2' / pe_st'' / c'' ->
(c1 ;; c2) / pe_st || (c1' ;; c2') / pe_st'' / c''
| PE_IfTrue : forall pe_st pe_st' b1 c1 c2 c1' c'',
pe_bexp pe_st b1 = BTrue ->
c1 / pe_st || c1' / pe_st' / c'' ->
(IFB b1 THEN c1 ELSE c2 FI) / pe_st || c1' / pe_st' / c''
| PE_IfFalse : forall pe_st pe_st' b1 c1 c2 c2' c'',
pe_bexp pe_st b1 = BFalse ->
c2 / pe_st || c2' / pe_st' / c'' ->
(IFB b1 THEN c1 ELSE c2 FI) / pe_st || c2' / pe_st' / c''
| PE_If : forall pe_st pe_st1 pe_st2 b1 c1 c2 c1' c2' c'',
pe_bexp pe_st b1 <> BTrue ->
pe_bexp pe_st b1 <> BFalse ->
c1 / pe_st || c1' / pe_st1 / c'' ->
c2 / pe_st || c2' / pe_st2 / c'' ->
(IFB b1 THEN c1 ELSE c2 FI) / pe_st
|| (IFB pe_bexp pe_st b1
THEN c1' ;; assign pe_st1 (pe_compare pe_st1 pe_st2)
ELSE c2' ;; assign pe_st2 (pe_compare pe_st1 pe_st2) FI)
/ pe_removes pe_st1 (pe_compare pe_st1 pe_st2)
/ c''
| PE_WhileEnd : forall pe_st b1 c1,
pe_bexp pe_st b1 = BFalse ->
(WHILE b1 DO c1 END) / pe_st || SKIP / pe_st / SKIP
| PE_WhileLoop : forall pe_st pe_st' pe_st'' b1 c1 c1' c2' c2'',
pe_bexp pe_st b1 = BTrue ->
c1 / pe_st || c1' / pe_st' / SKIP ->
(WHILE b1 DO c1 END) / pe_st' || c2' / pe_st'' / c2'' ->
pe_compare pe_st pe_st'' <> [] ->
(WHILE b1 DO c1 END) / pe_st || (c1';;c2') / pe_st'' / c2''
| PE_While : forall pe_st pe_st' pe_st'' b1 c1 c1' c2' c2'',
pe_bexp pe_st b1 <> BFalse ->
pe_bexp pe_st b1 <> BTrue ->
c1 / pe_st || c1' / pe_st' / SKIP ->
(WHILE b1 DO c1 END) / pe_st' || c2' / pe_st'' / c2'' ->
pe_compare pe_st pe_st'' <> [] ->
(c2'' = SKIP \/ c2'' = WHILE b1 DO c1 END) ->
(WHILE b1 DO c1 END) / pe_st
|| (IFB pe_bexp pe_st b1
THEN c1';; c2';; assign pe_st'' (pe_compare pe_st pe_st'')
ELSE assign pe_st (pe_compare pe_st pe_st'') FI)