-
Notifications
You must be signed in to change notification settings - Fork 2.9k
/
beam_ssa_opt.erl
1650 lines (1484 loc) · 55.5 KB
/
beam_ssa_opt.erl
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
%%
%% %CopyrightBegin%
%%
%% Copyright Ericsson AB 2018. All Rights Reserved.
%%
%% Licensed under the Apache License, Version 2.0 (the "License");
%% you may not use this file except in compliance with the License.
%% You may obtain a copy of the License at
%%
%% http://www.apache.org/licenses/LICENSE-2.0
%%
%% Unless required by applicable law or agreed to in writing, software
%% distributed under the License is distributed on an "AS IS" BASIS,
%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
%% See the License for the specific language governing permissions and
%% limitations under the License.
%%
%% %CopyrightEnd%
%%
-module(beam_ssa_opt).
-export([module/2]).
-include("beam_ssa.hrl").
-import(lists, [all/2,append/1,foldl/3,keyfind/3,member/2,reverse/1,reverse/2,
splitwith/2,takewhile/2,unzip/1]).
-spec module(beam_ssa:b_module(), [compile:option()]) ->
{'ok',beam_ssa:b_module()}.
module(#b_module{body=Fs0}=Module, Opts) ->
Ps = passes(Opts),
Fs = functions(Fs0, Ps),
{ok,Module#b_module{body=Fs}}.
functions([F|Fs], Ps) ->
[function(F, Ps)|functions(Fs, Ps)];
functions([], _Ps) -> [].
-type b_blk() :: beam_ssa:b_blk().
-type b_var() :: beam_ssa:b_var().
-type label() :: beam_ssa:label().
-record(st, {ssa :: beam_ssa:block_map() | [{label(),b_blk()}],
args :: [b_var()],
cnt :: label()}).
-define(PASS(N), {N,fun N/1}).
passes(Opts0) ->
Ps = [?PASS(ssa_opt_split_blocks),
?PASS(ssa_opt_coalesce_phis),
?PASS(ssa_opt_element),
?PASS(ssa_opt_linearize),
?PASS(ssa_opt_record),
%% Run ssa_opt_cse twice, because it will help ssa_opt_dead,
%% and ssa_opt_dead will help ssa_opt_cse. Run ssa_opt_live
%% twice, because it will help ssa_opt_dead and ssa_opt_dead
%% will help ssa_opt_live.
?PASS(ssa_opt_cse),
?PASS(ssa_opt_type),
?PASS(ssa_opt_live),
?PASS(ssa_opt_dead),
?PASS(ssa_opt_cse), %Second time.
?PASS(ssa_opt_float),
?PASS(ssa_opt_live), %Second time.
?PASS(ssa_opt_bsm),
?PASS(ssa_opt_bsm_shortcut),
?PASS(ssa_opt_misc),
?PASS(ssa_opt_tuple_size),
?PASS(ssa_opt_sw),
?PASS(ssa_opt_blockify),
?PASS(ssa_opt_sink),
?PASS(ssa_opt_merge_blocks)],
Negations = [{list_to_atom("no_"++atom_to_list(N)),N} ||
{N,_} <- Ps],
Opts = proplists:substitute_negations(Negations, Opts0),
[case proplists:get_value(Name, Opts, true) of
true ->
P;
false ->
{NoName,Name} = keyfind(Name, 2, Negations),
{NoName,fun(S) -> S end}
end || {Name,_}=P <- Ps].
function(#b_function{anno=Anno,bs=Blocks0,args=Args,cnt=Count0}=F, Ps) ->
try
St = #st{ssa=Blocks0,args=Args,cnt=Count0},
#st{ssa=Blocks,cnt=Count} = compile:run_sub_passes(Ps, St),
F#b_function{bs=Blocks,cnt=Count}
catch
Class:Error:Stack ->
#{func_info:={_,Name,Arity}} = Anno,
io:fwrite("Function: ~w/~w\n", [Name,Arity]),
erlang:raise(Class, Error, Stack)
end.
%%%
%%% Trivial sub passes.
%%%
ssa_opt_dead(#st{ssa=Linear}=St) ->
St#st{ssa=beam_ssa_dead:opt(Linear)}.
ssa_opt_linearize(#st{ssa=Blocks}=St) ->
St#st{ssa=beam_ssa:linearize(Blocks)}.
ssa_opt_type(#st{ssa=Linear,args=Args}=St) ->
St#st{ssa=beam_ssa_type:opt(Linear, Args)}.
ssa_opt_blockify(#st{ssa=Linear}=St) ->
St#st{ssa=maps:from_list(Linear)}.
%%%
%%% Split blocks before certain instructions to enable more optimizations.
%%%
%%% Splitting before element/2 enables the optimization that swaps
%%% element/2 instructions.
%%%
%%% Splitting before call and make_fun instructions gives more opportunities
%%% for sinking get_tuple_element instructions.
%%%
ssa_opt_split_blocks(#st{ssa=Blocks0,cnt=Count0}=St) ->
P = fun(#b_set{op={bif,element}}) -> true;
(#b_set{op=call}) -> true;
(#b_set{op=make_fun}) -> true;
(_) -> false
end,
{Blocks,Count} = beam_ssa:split_blocks(P, Blocks0, Count0),
St#st{ssa=Blocks,cnt=Count}.
%%%
%%% Coalesce phi nodes.
%%%
%%% Nested cases can led to code such as this:
%%%
%%% 10:
%%% _1 = phi {literal value1, label 8}, {Var, label 9}
%%% br 11
%%%
%%% 11:
%%% _2 = phi {_1, label 10}, {literal false, label 3}
%%%
%%% The phi nodes can be coalesced like this:
%%%
%%% 11:
%%% _2 = phi {literal value1, label 8}, {Var, label 9}, {literal false, label 3}
%%%
%%% Coalescing can help other optimizations, and can in some cases reduce register
%%% shuffling (if the phi variables for two phi nodes happens to be allocated to
%%% different registers).
%%%
ssa_opt_coalesce_phis(#st{ssa=Blocks0}=St) ->
Ls = beam_ssa:rpo(Blocks0),
Blocks = c_phis_1(Ls, Blocks0),
St#st{ssa=Blocks}.
c_phis_1([L|Ls], Blocks0) ->
case maps:get(L, Blocks0) of
#b_blk{is=[#b_set{op=phi}|_]}=Blk ->
Blocks = c_phis_2(L, Blk, Blocks0),
c_phis_1(Ls, Blocks);
#b_blk{} ->
c_phis_1(Ls, Blocks0)
end;
c_phis_1([], Blocks) -> Blocks.
c_phis_2(L, #b_blk{is=Is0}=Blk0, Blocks0) ->
case c_phis_args(Is0, Blocks0) of
none ->
Blocks0;
{_,_,Preds}=Info ->
Is = c_rewrite_phis(Is0, Info),
Blk = Blk0#b_blk{is=Is},
Blocks = Blocks0#{L:=Blk},
c_fix_branches(Preds, L, Blocks)
end.
c_phis_args([#b_set{op=phi,args=Args0}|Is], Blocks) ->
case c_phis_args_1(Args0, Blocks) of
none ->
c_phis_args(Is, Blocks);
Res ->
Res
end;
c_phis_args(_, _Blocks) -> none.
c_phis_args_1([{Var,Pred}|As], Blocks) ->
case c_get_pred_vars(Var, Pred, Blocks) of
none ->
c_phis_args_1(As, Blocks);
Result ->
Result
end;
c_phis_args_1([], _Blocks) -> none.
c_get_pred_vars(Var, Pred, Blocks) ->
case maps:get(Pred, Blocks) of
#b_blk{is=[#b_set{op=phi,dst=Var,args=Args}]} ->
{Var,Pred,Args};
#b_blk{} ->
none
end.
c_rewrite_phis([#b_set{op=phi,args=Args0}=I|Is], Info) ->
Args = c_rewrite_phi(Args0, Info),
[I#b_set{args=Args}|c_rewrite_phis(Is, Info)];
c_rewrite_phis(Is, _Info) -> Is.
c_rewrite_phi([{Var,Pred}|As], {Var,Pred,Values}) ->
Values ++ As;
c_rewrite_phi([{Value,Pred}|As], {_,Pred,Values}) ->
[{Value,P} || {_,P} <- Values] ++ As;
c_rewrite_phi([A|As], Info) ->
[A|c_rewrite_phi(As, Info)];
c_rewrite_phi([], _Info) -> [].
c_fix_branches([{_,Pred}|As], L, Blocks0) ->
#b_blk{last=Last0} = Blk0 = maps:get(Pred, Blocks0),
#b_br{bool=#b_literal{val=true}} = Last0, %Assertion.
Last = Last0#b_br{bool=#b_literal{val=true},succ=L,fail=L},
Blk = Blk0#b_blk{last=Last},
Blocks = Blocks0#{Pred:=Blk},
c_fix_branches(As, L, Blocks);
c_fix_branches([], _, Blocks) -> Blocks.
%%%
%%% Order element/2 calls.
%%%
%%% Order an unbroken chain of element/2 calls for the same tuple
%%% with the same failure label so that the highest element is
%%% retrieved first. That will allow the other element/2 calls to
%%% be replaced with get_tuple_element/3 instructions.
%%%
ssa_opt_element(#st{ssa=Blocks}=St) ->
%% Collect the information about element instructions in this
%% function.
GetEls = collect_element_calls(beam_ssa:linearize(Blocks)),
%% Collect the element instructions into chains. The
%% element calls in each chain are ordered in reverse
%% execution order.
Chains = collect_chains(GetEls, []),
%% For each chain, swap the first element call with the
%% element call with the highest index.
St#st{ssa=swap_element_calls(Chains, Blocks)}.
collect_element_calls([{L,#b_blk{is=Is0,last=Last}}|Bs]) ->
case {Is0,Last} of
{[#b_set{op={bif,element},dst=Element,
args=[#b_literal{val=N},#b_var{}=Tuple]},
#b_set{op=succeeded,dst=Bool,args=[Element]}],
#b_br{bool=Bool,succ=Succ,fail=Fail}} ->
Info = {L,Succ,{Tuple,Fail},N},
[Info|collect_element_calls(Bs)];
{_,_} ->
collect_element_calls(Bs)
end;
collect_element_calls([]) -> [].
collect_chains([{This,_,V,_}=El|Els], [{_,This,V,_}|_]=Chain) ->
%% Add to the previous chain.
collect_chains(Els, [El|Chain]);
collect_chains([El|Els], [_,_|_]=Chain) ->
%% Save the previous chain and start a new chain.
[Chain|collect_chains(Els, [El])];
collect_chains([El|Els], _Chain) ->
%% The previous chain is too short; discard it and start a new.
collect_chains(Els, [El]);
collect_chains([], [_,_|_]=Chain) ->
%% Save the last chain.
[Chain];
collect_chains([], _) -> [].
swap_element_calls([[{L,_,_,N}|_]=Chain|Chains], Blocks0) ->
Blocks = swap_element_calls_1(Chain, {N,L}, Blocks0),
swap_element_calls(Chains, Blocks);
swap_element_calls([], Blocks) -> Blocks.
swap_element_calls_1([{L1,_,_,N1}], {N2,L2}, Blocks) when N2 > N1 ->
%% We have reached the end of the chain, and the first
%% element instrution to be executed. Its index is lower
%% than the maximum index found while traversing the chain,
%% so we will need to swap the instructions.
#{L1:=Blk1,L2:=Blk2} = Blocks,
[#b_set{dst=Dst1}=GetEl1,Succ1] = Blk1#b_blk.is,
[#b_set{dst=Dst2}=GetEl2,Succ2] = Blk2#b_blk.is,
Is1 = [GetEl2,Succ1#b_set{args=[Dst2]}],
Is2 = [GetEl1,Succ2#b_set{args=[Dst1]}],
Blocks#{L1:=Blk1#b_blk{is=Is1},L2:=Blk2#b_blk{is=Is2}};
swap_element_calls_1([{L,_,_,N1}|Els], {N2,_}, Blocks) when N1 > N2 ->
swap_element_calls_1(Els, {N2,L}, Blocks);
swap_element_calls_1([_|Els], Highest, Blocks) ->
swap_element_calls_1(Els, Highest, Blocks);
swap_element_calls_1([], _, Blocks) ->
%% Nothing to do. The element call with highest index
%% is already the first one to be executed.
Blocks.
%%%
%%% Record optimization.
%%%
%%% Replace tuple matching with an is_tagged_tuple instruction
%%% when applicable.
%%%
ssa_opt_record(#st{ssa=Linear}=St) ->
Blocks = maps:from_list(Linear),
St#st{ssa=record_opt(Linear, Blocks)}.
record_opt([{L,#b_blk{is=Is0,last=Last}=Blk0}|Bs], Blocks) ->
Is = record_opt_is(Is0, Last, Blocks),
Blk = Blk0#b_blk{is=Is},
[{L,Blk}|record_opt(Bs, Blocks)];
record_opt([], _Blocks) -> [].
record_opt_is([#b_set{op={bif,is_tuple},dst=Bool,args=[Tuple]}=Set],
Last, Blocks) ->
case is_tagged_tuple(Tuple, Bool, Last, Blocks) of
{yes,Size,Tag} ->
Args = [Tuple,Size,Tag],
[Set#b_set{op=is_tagged_tuple,args=Args}];
no ->
[Set]
end;
record_opt_is([I|Is], Last, Blocks) ->
[I|record_opt_is(Is, Last, Blocks)];
record_opt_is([], _Last, _Blocks) -> [].
is_tagged_tuple(#b_var{}=Tuple, Bool,
#b_br{bool=Bool,succ=Succ,fail=Fail},
Blocks) ->
SuccBlk = maps:get(Succ, Blocks),
is_tagged_tuple_1(SuccBlk, Tuple, Fail, Blocks);
is_tagged_tuple(_, _, _, _) -> no.
is_tagged_tuple_1(#b_blk{is=Is,last=Last}, Tuple, Fail, Blocks) ->
case Is of
[#b_set{op={bif,tuple_size},dst=ArityVar,
args=[#b_var{}=Tuple]},
#b_set{op={bif,'=:='},
dst=Bool,
args=[ArityVar, #b_literal{val=ArityVal}=Arity]}]
when is_integer(ArityVal) ->
case Last of
#b_br{bool=Bool,succ=Succ,fail=Fail} ->
SuccBlk = maps:get(Succ, Blocks),
case is_tagged_tuple_2(SuccBlk, Tuple, Fail) of
no ->
no;
{yes,Tag} ->
{yes,Arity,Tag}
end;
_ ->
no
end;
_ ->
no
end.
is_tagged_tuple_2(#b_blk{is=Is,
last=#b_br{bool=#b_var{}=Bool,fail=Fail}},
Tuple, Fail) ->
is_tagged_tuple_3(Is, Bool, Tuple);
is_tagged_tuple_2(#b_blk{}, _, _) -> no.
is_tagged_tuple_3([#b_set{op=get_tuple_element,
dst=TagVar,
args=[#b_var{}=Tuple,#b_literal{val=0}]}|Is],
Bool, Tuple) ->
is_tagged_tuple_4(Is, Bool, TagVar);
is_tagged_tuple_3([_|Is], Bool, Tuple) ->
is_tagged_tuple_3(Is, Bool, Tuple);
is_tagged_tuple_3([], _, _) -> no.
is_tagged_tuple_4([#b_set{op={bif,'=:='},dst=Bool,
args=[#b_var{}=TagVar,
#b_literal{val=TagVal}=Tag]}],
Bool, TagVar) when is_atom(TagVal) ->
{yes,Tag};
is_tagged_tuple_4([_|Is], Bool, TagVar) ->
is_tagged_tuple_4(Is, Bool, TagVar);
is_tagged_tuple_4([], _, _) -> no.
%%%
%%% Common subexpression elimination (CSE).
%%%
%%% Eliminate repeated evaluation of identical expressions. To avoid
%%% increasing the size of the stack frame, we don't eliminate
%%% subexpressions across instructions that clobber the X registers.
%%%
ssa_opt_cse(#st{ssa=Linear}=St) ->
M = #{0=>#{}},
St#st{ssa=cse(Linear, #{}, M)}.
cse([{L,#b_blk{is=Is0,last=Last0}=Blk}|Bs], Sub0, M0) ->
Es0 = maps:get(L, M0),
{Is1,Es,Sub} = cse_is(Is0, Es0, Sub0, []),
Last = sub(Last0, Sub),
M = cse_successors(Is1, Blk, Es, M0),
Is = reverse(Is1),
[{L,Blk#b_blk{is=Is,last=Last}}|cse(Bs, Sub, M)];
cse([], _, _) -> [].
cse_successors([#b_set{op=succeeded,args=[Src]},Bif|_], Blk, EsSucc, M0) ->
case cse_suitable(Bif) of
true ->
%% The previous instruction only has a valid value at the success branch.
%% We must remove the substitution for Src from the failure branch.
#b_blk{last=#b_br{succ=Succ,fail=Fail}} = Blk,
M = cse_successors_1([Succ], EsSucc, M0),
EsFail = maps:filter(fun(_, Val) -> Val =/= Src end, EsSucc),
cse_successors_1([Fail], EsFail, M);
false ->
%% There can't be any replacement for Src in EsSucc. No need for
%% any special handling.
cse_successors_1(beam_ssa:successors(Blk), EsSucc, M0)
end;
cse_successors(_Is, Blk, Es, M) ->
cse_successors_1(beam_ssa:successors(Blk), Es, M).
cse_successors_1([L|Ls], Es0, M) ->
case M of
#{L:=Es1} when map_size(Es1) =:= 0 ->
%% The map is already empty. No need to do anything
%% since the intersection will be empty.
cse_successors_1(Ls, Es0, M);
#{L:=Es1} ->
%% Calculate the intersection of the two maps.
%% Both keys and values must match.
Es = maps:filter(fun(Key, Value) ->
case Es1 of
#{Key:=Value} -> true;
#{} -> false
end
end, Es0),
cse_successors_1(Ls, Es0, M#{L:=Es});
#{} ->
cse_successors_1(Ls, Es0, M#{L=>Es0})
end;
cse_successors_1([], _, M) -> M.
cse_is([#b_set{op=succeeded,dst=Bool,args=[Src]}=I0|Is], Es, Sub0, Acc) ->
I = sub(I0, Sub0),
case I of
#b_set{args=[Src]} ->
cse_is(Is, Es, Sub0, [I|Acc]);
#b_set{} ->
%% The previous instruction has been eliminated. Eliminate the
%% 'succeeded' instruction too.
Sub = Sub0#{Bool=>#b_literal{val=true}},
cse_is(Is, Es, Sub, Acc)
end;
cse_is([#b_set{dst=Dst}=I0|Is], Es0, Sub0, Acc) ->
I = sub(I0, Sub0),
case beam_ssa:clobbers_xregs(I) of
true ->
%% Retaining the expressions map across calls and other
%% clobbering instructions would work, but it would cause
%% the common subexpressions to be saved to Y registers,
%% which would probably increase the size of the stack
%% frame.
cse_is(Is, #{}, Sub0, [I|Acc]);
false ->
case cse_expr(I) of
none ->
%% Not suitable for CSE.
cse_is(Is, Es0, Sub0, [I|Acc]);
{ok,ExprKey} ->
case Es0 of
#{ExprKey:=Src} ->
Sub = Sub0#{Dst=>Src},
cse_is(Is, Es0, Sub, Acc);
#{} ->
Es = Es0#{ExprKey=>Dst},
cse_is(Is, Es, Sub0, [I|Acc])
end
end
end;
cse_is([], Es, Sub, Acc) ->
{Acc,Es,Sub}.
cse_expr(#b_set{op=Op,args=Args}=I) ->
case cse_suitable(I) of
true -> {ok,{Op,Args}};
false -> none
end.
cse_suitable(#b_set{op=get_hd}) -> true;
cse_suitable(#b_set{op=get_tl}) -> true;
cse_suitable(#b_set{op=put_list}) -> true;
cse_suitable(#b_set{op=put_tuple}) -> true;
cse_suitable(#b_set{op={bif,tuple_size}}) ->
%% Doing CSE for tuple_size/1 can prevent the
%% creation of test_arity and select_tuple_arity
%% instructions. That could decrease performance
%% and beam_validator could fail to understand
%% that tuple operations that follow are safe.
false;
cse_suitable(#b_set{anno=Anno,op={bif,Name},args=Args}) ->
%% Doing CSE for floating point operators is unsafe.
%% Doing CSE for comparison operators would prevent
%% creation of 'test' instructions.
Arity = length(Args),
not (is_map_key(float_op, Anno) orelse
erl_internal:new_type_test(Name, Arity) orelse
erl_internal:comp_op(Name, Arity) orelse
erl_internal:bool_op(Name, Arity));
cse_suitable(#b_set{}) -> false.
%%%
%%% Using floating point instructions.
%%%
%%% Use the special floating points version of arithmetic
%%% instructions, if the operands are known to be floats or the result
%%% of the operation will be a float.
%%%
%%% The float instructions were never used in guards before, so we
%%% will take special care to keep not using them in guards. Using
%%% them in guards would require a new version of the 'fconv'
%%% instruction that would take a failure label. Since it is unlikely
%%% that using float instructions in guards would be benefical, why
%%% bother implementing a new instruction? Also, implementing float
%%% instructions in guards in HiPE could turn out to be a lot of work.
%%%
-record(fs,
{s=undefined :: 'undefined' | 'cleared',
regs=#{} :: #{beam_ssa:b_var():=beam_ssa:b_var()},
fail=none :: 'none' | beam_ssa:label(),
non_guards :: gb_sets:set(beam_ssa:label()),
bs :: beam_ssa:block_map()
}).
ssa_opt_float(#st{ssa=Linear0,cnt=Count0}=St) ->
NonGuards0 = float_non_guards(Linear0),
NonGuards = gb_sets:from_list(NonGuards0),
Blocks = maps:from_list(Linear0),
Fs = #fs{non_guards=NonGuards,bs=Blocks},
{Linear,Count} = float_opt(Linear0, Count0, Fs),
St#st{ssa=Linear,cnt=Count}.
float_non_guards([{L,#b_blk{is=Is}}|Bs]) ->
case Is of
[#b_set{op=landingpad}|_] ->
[L|float_non_guards(Bs)];
_ ->
float_non_guards(Bs)
end;
float_non_guards([]) -> [?BADARG_BLOCK].
float_opt([{L,#b_blk{last=#b_br{fail=F}}=Blk}|Bs0],
Count0, #fs{non_guards=NonGuards}=Fs) ->
case gb_sets:is_member(F, NonGuards) of
true ->
%% This block is not inside a guard.
%% We can do the optimization.
float_opt_1(L, Blk, Bs0, Count0, Fs);
false ->
%% This block is inside a guard. Don't do
%% any floating point optimizations.
{Bs,Count} = float_opt(Bs0, Count0, Fs),
{[{L,Blk}|Bs],Count}
end;
float_opt([{L,Blk}|Bs], Count, Fs) ->
float_opt_1(L, Blk, Bs, Count, Fs);
float_opt([], Count, _Fs) ->
{[],Count}.
float_opt_1(L, #b_blk{is=Is0}=Blk0, Bs0, Count0, Fs0) ->
case float_opt_is(Is0, Fs0, Count0, []) of
{Is1,Fs1,Count1} ->
Fs2 = float_fail_label(Blk0, Fs1),
Fail = Fs2#fs.fail,
{Flush,Blk,Fs,Count2} = float_maybe_flush(Blk0, Fs2, Count1),
Split = float_split_conv(Is1, Blk),
{Blks0,Count3} = float_number(Split, L, Count2),
{Blks,Count4} = float_conv(Blks0, Fail, Count3),
{Bs,Count} = float_opt(Bs0, Count4, Fs),
{Blks++Flush++Bs,Count};
none ->
{Bs,Count} = float_opt(Bs0, Count0, Fs0),
{[{L,Blk0}|Bs],Count}
end.
%% Split {float,convert} instructions into individual blocks.
float_split_conv(Is0, Blk) ->
Br = #b_br{bool=#b_literal{val=true},succ=0,fail=0},
case splitwith(fun(#b_set{op=Op}) ->
Op =/= {float,convert}
end, Is0) of
{Is,[]} ->
[Blk#b_blk{is=Is}];
{[_|_]=Is1,[#b_set{op={float,convert}}=Conv|Is2]} ->
[#b_blk{is=Is1,last=Br},
#b_blk{is=[Conv],last=Br}|float_split_conv(Is2, Blk)];
{[],[#b_set{op={float,convert}}=Conv|Is1]} ->
[#b_blk{is=[Conv],last=Br}|float_split_conv(Is1, Blk)]
end.
%% Number the blocks that were split.
float_number([B|Bs0], FirstL, Count0) ->
{Bs,Count} = float_number(Bs0, Count0),
{[{FirstL,B}|Bs],Count}.
float_number([B|Bs0], Count0) ->
{Bs,Count} = float_number(Bs0, Count0+1),
{[{Count0,B}|Bs],Count};
float_number([], Count) ->
{[],Count}.
%% Insert 'succeeded' instructions after each {float,convert}
%% instruction.
float_conv([{L,#b_blk{is=Is0}=Blk0}|Bs0], Fail, Count0) ->
case Is0 of
[#b_set{op={float,convert}}=Conv] ->
{Bool0,Count1} = new_reg('@ssa_bool', Count0),
Bool = #b_var{name=Bool0},
Succeeded = #b_set{op=succeeded,dst=Bool,
args=[Conv#b_set.dst]},
Is = [Conv,Succeeded],
[{NextL,_}|_] = Bs0,
Br = #b_br{bool=Bool,succ=NextL,fail=Fail},
Blk = Blk0#b_blk{is=Is,last=Br},
{Bs,Count} = float_conv(Bs0, Fail, Count1),
{[{L,Blk}|Bs],Count};
[_|_] ->
case Bs0 of
[{NextL,_}|_] ->
Br = #b_br{bool=#b_literal{val=true},
succ=NextL,fail=NextL},
Blk = Blk0#b_blk{last=Br},
{Bs,Count} = float_conv(Bs0, Fail, Count0),
{[{L,Blk}|Bs],Count};
[] ->
{[{L,Blk0}],Count0}
end
end.
float_maybe_flush(Blk0, #fs{s=cleared,fail=Fail,bs=Blocks}=Fs0, Count0) ->
#b_blk{last=#b_br{bool=#b_var{},succ=Succ}=Br} = Blk0,
#b_blk{is=Is} = maps:get(Succ, Blocks),
case Is of
[#b_set{anno=#{float_op:=_}}|_] ->
%% The next operation is also a floating point operation.
%% No flush needed.
{[],Blk0,Fs0,Count0};
_ ->
%% Flush needed.
{Bool0,Count1} = new_reg('@ssa_bool', Count0),
Bool = #b_var{name=Bool0},
%% Allocate block numbers.
CheckL = Count1, %For checkerror.
FlushL = Count1 + 1, %For flushing of float regs.
Count = Count1 + 2,
Blk = Blk0#b_blk{last=Br#b_br{succ=CheckL}},
%% Build the block with the checkerror instruction.
CheckIs = [#b_set{op={float,checkerror},dst=Bool}],
CheckBr = #b_br{bool=Bool,succ=FlushL,fail=Fail},
CheckBlk = #b_blk{is=CheckIs,last=CheckBr},
%% Build the block that flushes all registers.
FlushIs = float_flush_regs(Fs0),
FlushBr = #b_br{bool=#b_literal{val=true},succ=Succ,fail=Succ},
FlushBlk = #b_blk{is=FlushIs,last=FlushBr},
%% Update state and blocks.
Fs = Fs0#fs{s=undefined,regs=#{},fail=none},
FlushBs = [{CheckL,CheckBlk},{FlushL,FlushBlk}],
{FlushBs,Blk,Fs,Count}
end;
float_maybe_flush(Blk, Fs, Count) ->
{[],Blk,Fs,Count}.
float_opt_is([#b_set{op=succeeded,args=[Src]}=I0],
#fs{regs=Rs}=Fs, Count, Acc) ->
case Rs of
#{Src:=Fr} ->
I = I0#b_set{args=[Fr]},
{reverse(Acc, [I]),Fs,Count};
#{} ->
{reverse(Acc, [I0]),Fs,Count}
end;
float_opt_is([#b_set{anno=Anno0}=I0|Is0], Fs0, Count0, Acc) ->
case Anno0 of
#{float_op:=FTypes} ->
Anno = maps:remove(float_op, Anno0),
I1 = I0#b_set{anno=Anno},
{Is,Fs,Count} = float_make_op(I1, FTypes, Fs0, Count0),
float_opt_is(Is0, Fs, Count, reverse(Is, Acc));
#{} ->
float_opt_is(Is0, Fs0#fs{regs=#{}}, Count0, [I0|Acc])
end;
float_opt_is([], Fs, _Count, _Acc) ->
#fs{s=undefined} = Fs, %Assertion.
none.
float_make_op(#b_set{op={bif,Op},dst=Dst,args=As0}=I0,
Ts, #fs{s=S,regs=Rs0}=Fs, Count0) ->
{As1,Rs1,Count1} = float_load(As0, Ts, Rs0, Count0, []),
{As,Is0} = unzip(As1),
{Fr,Count2} = new_reg('@fr', Count1),
FrDst = #b_var{name=Fr},
I = I0#b_set{op={float,Op},dst=FrDst,args=As},
Rs = Rs1#{Dst=>FrDst},
Is = append(Is0) ++ [I],
case S of
undefined ->
{Ignore,Count} = new_reg('@ssa_ignore', Count2),
C = #b_set{op={float,clearerror},dst=#b_var{name=Ignore}},
{[C|Is],Fs#fs{s=cleared,regs=Rs},Count};
cleared ->
{Is,Fs#fs{regs=Rs},Count2}
end.
float_load([A|As], [T|Ts], Rs0, Count0, Acc) ->
{Load,Rs,Count} = float_reg_arg(A, T, Rs0, Count0),
float_load(As, Ts, Rs, Count, [Load|Acc]);
float_load([], [], Rs, Count, Acc) ->
{reverse(Acc),Rs,Count}.
float_reg_arg(A, T, Rs, Count0) ->
case Rs of
#{A:=Fr} ->
{{Fr,[]},Rs,Count0};
#{} ->
{Fr,Count} = new_float_copy_reg(Count0),
Dst = #b_var{name=Fr},
I = float_load_reg(T, A, Dst),
{{Dst,[I]},Rs#{A=>Dst},Count}
end.
float_load_reg(convert, #b_var{}=Src, Dst) ->
#b_set{op={float,convert},dst=Dst,args=[Src]};
float_load_reg(convert, #b_literal{val=Val}=Src, Dst) ->
try float(Val) of
F ->
#b_set{op={float,put},dst=Dst,args=[#b_literal{val=F}]}
catch
error:_ ->
%% Let the exception happen at runtime.
#b_set{op={float,convert},dst=Dst,args=[Src]}
end;
float_load_reg(float, Src, Dst) ->
#b_set{op={float,put},dst=Dst,args=[Src]}.
new_float_copy_reg(Count) ->
new_reg('@fr_copy', Count).
new_reg(Base, Count) ->
Fr = {Base,Count},
{Fr,Count+1}.
float_fail_label(#b_blk{last=Last}, Fs) ->
case Last of
#b_br{bool=#b_var{},fail=Fail} ->
Fs#fs{fail=Fail};
_ ->
Fs
end.
float_flush_regs(#fs{regs=Rs}) ->
maps:fold(fun(_, #b_var{name={'@fr_copy',_}}, Acc) ->
Acc;
(Dst, Fr, Acc) ->
[#b_set{op={float,get},dst=Dst,args=[Fr]}|Acc]
end, [], Rs).
%%%
%%% Live optimization.
%%%
%%% Optimize instructions whose values are not used. They could be
%%% removed if they have no side effects, or in a few cases replaced
%%% with a cheaper instructions
%%%
ssa_opt_live(#st{ssa=Linear}=St) ->
St#st{ssa=live_opt(reverse(Linear), #{}, [])}.
live_opt([{L,Blk0}|Bs], LiveMap0, Acc) ->
Successors = beam_ssa:successors(Blk0),
Live0 = live_opt_succ(Successors, L, LiveMap0),
{Blk,Live} = live_opt_blk(Blk0, Live0),
LiveMap = live_opt_phis(Blk#b_blk.is, L, Live, LiveMap0),
live_opt(Bs, LiveMap, [{L,Blk}|Acc]);
live_opt([], _, Acc) -> Acc.
live_opt_succ([S|Ss], L, LiveMap) ->
Live0 = live_opt_succ(Ss, L, LiveMap),
Key = {S,L},
case LiveMap of
#{Key:=Live} ->
gb_sets:union(Live, Live0);
#{S:=Live} ->
gb_sets:union(Live, Live0);
#{} ->
Live0
end;
live_opt_succ([], _, _) ->
gb_sets:empty().
live_opt_phis(Is, L, Live0, LiveMap0) ->
LiveMap = LiveMap0#{L=>Live0},
Phis = takewhile(fun(#b_set{op=Op}) -> Op =:= phi end, Is),
case Phis of
[] ->
LiveMap;
[_|_] ->
PhiArgs = append([Args || #b_set{args=Args} <- Phis]),
case [{P,V} || {#b_var{}=V,P} <- PhiArgs] of
[_|_]=PhiVars ->
PhiLive0 = rel2fam(PhiVars),
PhiLive = [{{L,P},gb_sets:union(gb_sets:from_list(Vs), Live0)} ||
{P,Vs} <- PhiLive0],
maps:merge(LiveMap, maps:from_list(PhiLive));
[] ->
%% There were only literals in the phi node(s).
LiveMap
end
end.
live_opt_blk(#b_blk{is=Is0,last=Last}=Blk, Live0) ->
Live1 = gb_sets:union(Live0, gb_sets:from_ordset(beam_ssa:used(Last))),
{Is,Live} = live_opt_is(reverse(Is0), Live1, []),
{Blk#b_blk{is=Is},Live}.
live_opt_is([#b_set{op=phi,dst=Dst}=I|Is], Live, Acc) ->
case gb_sets:is_member(Dst, Live) of
true ->
live_opt_is(Is, Live, [I|Acc]);
false ->
live_opt_is(Is, Live, Acc)
end;
live_opt_is([#b_set{op=succeeded,dst=SuccDst=SuccDstVar,
args=[Dst]}=SuccI,
#b_set{dst=Dst}=I|Is], Live0, Acc) ->
case gb_sets:is_member(Dst, Live0) of
true ->
case gb_sets:is_member(SuccDst, Live0) of
true ->
Live1 = gb_sets:add(Dst, Live0),
Live = gb_sets:delete_any(SuccDst, Live1),
live_opt_is([I|Is], Live, [SuccI|Acc]);
false ->
live_opt_is([I|Is], Live0, Acc)
end;
false ->
case live_opt_unused(I) of
{replace,NewI0} ->
NewI = NewI0#b_set{dst=SuccDstVar},
live_opt_is([NewI|Is], Live0, Acc);
keep ->
case gb_sets:is_member(SuccDst, Live0) of
true ->
Live1 = gb_sets:add(Dst, Live0),
Live = gb_sets:delete_any(SuccDst, Live1),
live_opt_is([I|Is], Live, [SuccI|Acc]);
false ->
live_opt_is([I|Is], Live0, Acc)
end
end
end;
live_opt_is([#b_set{dst=Dst}=I|Is], Live0, Acc) ->
case gb_sets:is_member(Dst, Live0) of
true ->
Live1 = gb_sets:union(Live0, gb_sets:from_ordset(beam_ssa:used(I))),
Live = gb_sets:delete_any(Dst, Live1),
live_opt_is(Is, Live, [I|Acc]);
false ->
case beam_ssa:no_side_effect(I) of
true ->
live_opt_is(Is, Live0, Acc);
false ->
Live = gb_sets:union(Live0, gb_sets:from_ordset(beam_ssa:used(I))),
live_opt_is(Is, Live, [I|Acc])
end
end;
live_opt_is([], Live, Acc) ->
{Acc,Live}.
live_opt_unused(#b_set{op=get_map_element}=Set) ->
{replace,Set#b_set{op=has_map_field}};
live_opt_unused(_) -> keep.
%%%
%%% Optimize binary matching instructions.
%%%
ssa_opt_bsm(#st{ssa=Linear}=St) ->
Extracted0 = bsm_extracted(Linear),
Extracted = cerl_sets:from_list(Extracted0),
St#st{ssa=bsm_skip(Linear, Extracted)}.
bsm_skip([{L,#b_blk{is=Is0}=Blk}|Bs], Extracted) ->
Is = bsm_skip_is(Is0, Extracted),
[{L,Blk#b_blk{is=Is}}|bsm_skip(Bs, Extracted)];
bsm_skip([], _) -> [].
bsm_skip_is([I0|Is], Extracted) ->
case I0 of
#b_set{op=bs_match,args=[#b_literal{val=string}|_]} ->
[I0|bsm_skip_is(Is, Extracted)];
#b_set{op=bs_match,dst=Ctx,args=[Type,PrevCtx|Args0]} ->
I = case cerl_sets:is_element(Ctx, Extracted) of
true ->
I0;
false ->
%% The value is never extracted.
Args = [#b_literal{val=skip},PrevCtx,Type|Args0],
I0#b_set{args=Args}
end,
[I|Is];
#b_set{} ->
[I0|bsm_skip_is(Is, Extracted)]
end;
bsm_skip_is([], _) -> [].
bsm_extracted([{_,#b_blk{is=Is}}|Bs]) ->
case Is of
[#b_set{op=bs_extract,args=[Ctx]}|_] ->
[Ctx|bsm_extracted(Bs)];
_ ->
bsm_extracted(Bs)
end;
bsm_extracted([]) -> [].
%%%
%%% Short-cutting binary matching instructions.
%%%
ssa_opt_bsm_shortcut(#st{ssa=Linear}=St) ->
Positions = bsm_positions(Linear, #{}),
case map_size(Positions) of
0 ->
%% No binary matching instructions.
St;
_ ->
St#st{ssa=bsm_shortcut(Linear, Positions)}
end.
bsm_positions([{L,#b_blk{is=Is,last=Last}}|Bs], PosMap0) ->
PosMap = bsm_positions_is(Is, PosMap0),
case {Is,Last} of
{[#b_set{op=bs_test_tail,dst=Bool,args=[Ctx,#b_literal{val=Bits0}]}],
#b_br{bool=Bool,fail=Fail}} ->
Bits = Bits0 + maps:get(Ctx, PosMap0),
bsm_positions(Bs, PosMap#{L=>{Bits,Fail}});
{_,_} ->
bsm_positions(Bs, PosMap)
end;
bsm_positions([], PosMap) -> PosMap.
bsm_positions_is([#b_set{op=bs_start_match,dst=New}|Is], PosMap0) ->
PosMap = PosMap0#{New=>0},
bsm_positions_is(Is, PosMap);
bsm_positions_is([#b_set{op=bs_match,dst=New,args=Args}|Is], PosMap0) ->
[_,Old|_] = Args,
#{Old:=Bits0} = PosMap0,
Bits = bsm_update_bits(Args, Bits0),
PosMap = PosMap0#{New=>Bits},
bsm_positions_is(Is, PosMap);
bsm_positions_is([_|Is], PosMap) ->
bsm_positions_is(Is, PosMap);
bsm_positions_is([], PosMap) -> PosMap.
bsm_update_bits([#b_literal{val=string},_,#b_literal{val=String}], Bits) ->
Bits + bit_size(String);
bsm_update_bits([#b_literal{val=utf8}|_], Bits) ->
Bits + 8;
bsm_update_bits([#b_literal{val=utf16}|_], Bits) ->
Bits + 16;
bsm_update_bits([#b_literal{val=utf32}|_], Bits) ->
Bits + 32;
bsm_update_bits([_,_,_,#b_literal{val=Sz},#b_literal{val=U}], Bits)
when is_integer(Sz) ->
Bits + Sz*U;
bsm_update_bits(_, Bits) -> Bits.
bsm_shortcut([{L,#b_blk{is=Is,last=Last0}=Blk}|Bs], PosMap) ->
case {Is,Last0} of
{[#b_set{op=bs_match,dst=New,args=[_,Old|_]},
#b_set{op=succeeded,dst=Bool,args=[New]}],
#b_br{bool=Bool,fail=Fail}} ->
case PosMap of
#{Old:=Bits,Fail:={TailBits,NextFail}} when Bits > TailBits ->
Last = Last0#b_br{fail=NextFail},
[{L,Blk#b_blk{last=Last}}|bsm_shortcut(Bs, PosMap)];
#{} ->
[{L,Blk}|bsm_shortcut(Bs, PosMap)]
end;
{_,_} ->
[{L,Blk}|bsm_shortcut(Bs, PosMap)]