aboutsummaryrefslogtreecommitdiff
path: root/gcc/doc/passes.texi
blob: b109f2db896b7c65b648a5a73305bba1b9a29ee8 (plain)
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
@c markers: CROSSREF BUG TODO

@c Copyright (C) 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
@c 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009 Free Software
@c Foundation, Inc.
@c This is part of the GCC manual.
@c For copying conditions, see the file gcc.texi.

@node Passes
@chapter Passes and Files of the Compiler
@cindex passes and files of the compiler
@cindex files and passes of the compiler
@cindex compiler passes and files

This chapter is dedicated to giving an overview of the optimization and
code generation passes of the compiler.  In the process, it describes
some of the language front end interface, though this description is no
where near complete.

@menu
* Parsing pass::         The language front end turns text into bits.
* Gimplification pass::  The bits are turned into something we can optimize.
* Pass manager::         Sequencing the optimization passes.
* Tree SSA passes::      Optimizations on a high-level representation.
* RTL passes::           Optimizations on a low-level representation.
@end menu

@node Parsing pass
@section Parsing pass
@cindex GENERIC
@findex lang_hooks.parse_file
The language front end is invoked only once, via
@code{lang_hooks.parse_file}, to parse the entire input.  The language
front end may use any intermediate language representation deemed
appropriate.  The C front end uses GENERIC trees (CROSSREF), plus
a double handful of language specific tree codes defined in
@file{c-common.def}.  The Fortran front end uses a completely different
private representation.

@cindex GIMPLE
@cindex gimplification
@cindex gimplifier
@cindex language-independent intermediate representation
@cindex intermediate representation lowering
@cindex lowering, language-dependent intermediate representation
At some point the front end must translate the representation used in the
front end to a representation understood by the language-independent
portions of the compiler.  Current practice takes one of two forms.
The C front end manually invokes the gimplifier (CROSSREF) on each function,
and uses the gimplifier callbacks to convert the language-specific tree
nodes directly to GIMPLE (CROSSREF) before passing the function off to
be compiled.
The Fortran front end converts from a private representation to GENERIC,
which is later lowered to GIMPLE when the function is compiled.  Which
route to choose probably depends on how well GENERIC (plus extensions)
can be made to match up with the source language and necessary parsing
data structures.

BUG: Gimplification must occur before nested function lowering,
and nested function lowering must be done by the front end before
passing the data off to cgraph.

TODO: Cgraph should control nested function lowering.  It would
only be invoked when it is certain that the outer-most function
is used.

TODO: Cgraph needs a gimplify_function callback.  It should be
invoked when (1) it is certain that the function is used, (2)
warning flags specified by the user require some amount of
compilation in order to honor, (3) the language indicates that
semantic analysis is not complete until gimplification occurs.
Hum@dots{} this sounds overly complicated.  Perhaps we should just
have the front end gimplify always; in most cases it's only one
function call.

The front end needs to pass all function definitions and top level
declarations off to the middle-end so that they can be compiled and
emitted to the object file.  For a simple procedural language, it is
usually most convenient to do this as each top level declaration or
definition is seen.  There is also a distinction to be made between
generating functional code and generating complete debug information.
The only thing that is absolutely required for functional code is that
function and data @emph{definitions} be passed to the middle-end.  For
complete debug information, function, data and type declarations
should all be passed as well.

@findex rest_of_decl_compilation
@findex rest_of_type_compilation
@findex cgraph_finalize_function
In any case, the front end needs each complete top-level function or
data declaration, and each data definition should be passed to
@code{rest_of_decl_compilation}.  Each complete type definition should
be passed to @code{rest_of_type_compilation}.  Each function definition
should be passed to @code{cgraph_finalize_function}.

TODO: I know rest_of_compilation currently has all sorts of
RTL generation semantics.  I plan to move all code generation
bits (both Tree and RTL) to compile_function.  Should we hide
cgraph from the front ends and move back to rest_of_compilation
as the official interface?  Possibly we should rename all three
interfaces such that the names match in some meaningful way and
that is more descriptive than "rest_of".

The middle-end will, at its option, emit the function and data
definitions immediately or queue them for later processing.

@node Gimplification pass
@section Gimplification pass

@cindex gimplification
@cindex GIMPLE
@dfn{Gimplification} is a whimsical term for the process of converting
the intermediate representation of a function into the GIMPLE language
(CROSSREF).  The term stuck, and so words like ``gimplification'',
``gimplify'', ``gimplifier'' and the like are sprinkled throughout this
section of code.

@cindex GENERIC
While a front end may certainly choose to generate GIMPLE directly if
it chooses, this can be a moderately complex process unless the
intermediate language used by the front end is already fairly simple.
Usually it is easier to generate GENERIC trees plus extensions
and let the language-independent gimplifier do most of the work.

@findex gimplify_function_tree
@findex gimplify_expr
@findex lang_hooks.gimplify_expr
The main entry point to this pass is @code{gimplify_function_tree}
located in @file{gimplify.c}.  From here we process the entire
function gimplifying each statement in turn.  The main workhorse
for this pass is @code{gimplify_expr}.  Approximately everything
passes through here at least once, and it is from here that we
invoke the @code{lang_hooks.gimplify_expr} callback.

The callback should examine the expression in question and return
@code{GS_UNHANDLED} if the expression is not a language specific
construct that requires attention.  Otherwise it should alter the
expression in some way to such that forward progress is made toward
producing valid GIMPLE@.  If the callback is certain that the
transformation is complete and the expression is valid GIMPLE, it
should return @code{GS_ALL_DONE}.  Otherwise it should return
@code{GS_OK}, which will cause the expression to be processed again.
If the callback encounters an error during the transformation (because
the front end is relying on the gimplification process to finish
semantic checks), it should return @code{GS_ERROR}.

@node Pass manager
@section Pass manager

The pass manager is located in @file{passes.c}, @file{tree-optimize.c}
and @file{tree-pass.h}.
Its job is to run all of the individual passes in the correct order,
and take care of standard bookkeeping that applies to every pass.

The theory of operation is that each pass defines a structure that
represents everything we need to know about that pass---when it
should be run, how it should be run, what intermediate language
form or on-the-side data structures it needs.  We register the pass
to be run in some particular order, and the pass manager arranges
for everything to happen in the correct order.

The actuality doesn't completely live up to the theory at present.
Command-line switches and @code{timevar_id_t} enumerations must still
be defined elsewhere.  The pass manager validates constraints but does
not attempt to (re-)generate data structures or lower intermediate
language form based on the requirements of the next pass.  Nevertheless,
what is present is useful, and a far sight better than nothing at all.

Each pass should have a unique name.
Each pass may have its own dump file (for GCC debugging purposes).
Passes with a name starting with a star do not dump anything.
Sometimes passes are supposed to share a dump file / option name.
To still give these unique names, you can use a prefix that is delimited
by a space from the part that is used for the dump file / option name.
E.g. When the pass name is "ud dce", the name used for dump file/options
is "dce".

TODO: describe the global variables set up by the pass manager,
and a brief description of how a new pass should use it.
I need to look at what info RTL passes use first@enddots{}

@node Tree SSA passes
@section Tree SSA passes

The following briefly describes the Tree optimization passes that are
run after gimplification and what source files they are located in.

@itemize @bullet
@item Remove useless statements

This pass is an extremely simple sweep across the gimple code in which
we identify obviously dead code and remove it.  Here we do things like
simplify @code{if} statements with constant conditions, remove
exception handling constructs surrounding code that obviously cannot
throw, remove lexical bindings that contain no variables, and other
assorted simplistic cleanups.  The idea is to get rid of the obvious
stuff quickly rather than wait until later when it's more work to get
rid of it.  This pass is located in @file{tree-cfg.c} and described by
@code{pass_remove_useless_stmts}.

@item Mudflap declaration registration

If mudflap (@pxref{Optimize Options,,-fmudflap -fmudflapth
-fmudflapir,gcc,Using the GNU Compiler Collection (GCC)}) is
enabled, we generate code to register some variable declarations with
the mudflap runtime.  Specifically, the runtime tracks the lifetimes of
those variable declarations that have their addresses taken, or whose
bounds are unknown at compile time (@code{extern}).  This pass generates
new exception handling constructs (@code{try}/@code{finally}), and so
must run before those are lowered.  In addition, the pass enqueues
declarations of static variables whose lifetimes extend to the entire
program.  The pass is located in @file{tree-mudflap.c} and is described
by @code{pass_mudflap_1}.

@item OpenMP lowering

If OpenMP generation (@option{-fopenmp}) is enabled, this pass lowers
OpenMP constructs into GIMPLE.

Lowering of OpenMP constructs involves creating replacement
expressions for local variables that have been mapped using data
sharing clauses, exposing the control flow of most synchronization
directives and adding region markers to facilitate the creation of the
control flow graph.  The pass is located in @file{omp-low.c} and is
described by @code{pass_lower_omp}.

@item OpenMP expansion

If OpenMP generation (@option{-fopenmp}) is enabled, this pass expands
parallel regions into their own functions to be invoked by the thread
library.  The pass is located in @file{omp-low.c} and is described by
@code{pass_expand_omp}.

@item Lower control flow

This pass flattens @code{if} statements (@code{COND_EXPR})
and moves lexical bindings (@code{BIND_EXPR}) out of line.  After
this pass, all @code{if} statements will have exactly two @code{goto}
statements in its @code{then} and @code{else} arms.  Lexical binding
information for each statement will be found in @code{TREE_BLOCK} rather
than being inferred from its position under a @code{BIND_EXPR}.  This
pass is found in @file{gimple-low.c} and is described by
@code{pass_lower_cf}.

@item Lower exception handling control flow

This pass decomposes high-level exception handling constructs
(@code{TRY_FINALLY_EXPR} and @code{TRY_CATCH_EXPR}) into a form
that explicitly represents the control flow involved.  After this
pass, @code{lookup_stmt_eh_region} will return a non-negative
number for any statement that may have EH control flow semantics;
examine @code{tree_can_throw_internal} or @code{tree_can_throw_external}
for exact semantics.  Exact control flow may be extracted from
@code{foreach_reachable_handler}.  The EH region nesting tree is defined
in @file{except.h} and built in @file{except.c}.  The lowering pass
itself is in @file{tree-eh.c} and is described by @code{pass_lower_eh}.

@item Build the control flow graph

This pass decomposes a function into basic blocks and creates all of
the edges that connect them.  It is located in @file{tree-cfg.c} and
is described by @code{pass_build_cfg}.

@item Find all referenced variables

This pass walks the entire function and collects an array of all
variables referenced in the function, @code{referenced_vars}.  The
index at which a variable is found in the array is used as a UID
for the variable within this function.  This data is needed by the
SSA rewriting routines.  The pass is located in @file{tree-dfa.c}
and is described by @code{pass_referenced_vars}.

@item Enter static single assignment form

This pass rewrites the function such that it is in SSA form.  After
this pass, all @code{is_gimple_reg} variables will be referenced by
@code{SSA_NAME}, and all occurrences of other variables will be
annotated with @code{VDEFS} and @code{VUSES}; PHI nodes will have
been inserted as necessary for each basic block.  This pass is
located in @file{tree-ssa.c} and is described by @code{pass_build_ssa}.

@item Warn for uninitialized variables

This pass scans the function for uses of @code{SSA_NAME}s that
are fed by default definition.  For non-parameter variables, such
uses are uninitialized.  The pass is run twice, before and after
optimization (if turned on).  In the first pass we only warn for uses that are
positively uninitialized; in the second pass we warn for uses that
are possibly uninitialized.  The pass is located in @file{tree-ssa.c}
and is defined by @code{pass_early_warn_uninitialized} and
@code{pass_late_warn_uninitialized}.

@item Dead code elimination

This pass scans the function for statements without side effects whose
result is unused.  It does not do memory life analysis, so any value
that is stored in memory is considered used.  The pass is run multiple
times throughout the optimization process.  It is located in
@file{tree-ssa-dce.c} and is described by @code{pass_dce}.

@item Dominator optimizations

This pass performs trivial dominator-based copy and constant propagation,
expression simplification, and jump threading.  It is run multiple times
throughout the optimization process.  It is located in @file{tree-ssa-dom.c}
and is described by @code{pass_dominator}.

@item Forward propagation of single-use variables

This pass attempts to remove redundant computation by substituting
variables that are used once into the expression that uses them and
seeing if the result can be simplified.  It is located in
@file{tree-ssa-forwprop.c} and is described by @code{pass_forwprop}.

@item Copy Renaming

This pass attempts to change the name of compiler temporaries involved in
copy operations such that SSA->normal can coalesce the copy away.  When compiler
temporaries are copies of user variables, it also renames the compiler
temporary to the user variable resulting in better use of user symbols.  It is
located in @file{tree-ssa-copyrename.c} and is described by
@code{pass_copyrename}.

@item PHI node optimizations

This pass recognizes forms of PHI inputs that can be represented as
conditional expressions and rewrites them into straight line code.
It is located in @file{tree-ssa-phiopt.c} and is described by
@code{pass_phiopt}.

@item May-alias optimization

This pass performs a flow sensitive SSA-based points-to analysis.
The resulting may-alias, must-alias, and escape analysis information
is used to promote variables from in-memory addressable objects to
non-aliased variables that can be renamed into SSA form.  We also
update the @code{VDEF}/@code{VUSE} memory tags for non-renameable
aggregates so that we get fewer false kills.  The pass is located
in @file{tree-ssa-alias.c} and is described by @code{pass_may_alias}.

Interprocedural points-to information is located in
@file{tree-ssa-structalias.c} and described by @code{pass_ipa_pta}.

@item Profiling

This pass rewrites the function in order to collect runtime block
and value profiling data.  Such data may be fed back into the compiler
on a subsequent run so as to allow optimization based on expected
execution frequencies.  The pass is located in @file{predict.c} and
is described by @code{pass_profile}.

@item Lower complex arithmetic

This pass rewrites complex arithmetic operations into their component
scalar arithmetic operations.  The pass is located in @file{tree-complex.c}
and is described by @code{pass_lower_complex}.

@item Scalar replacement of aggregates

This pass rewrites suitable non-aliased local aggregate variables into
a set of scalar variables.  The resulting scalar variables are
rewritten into SSA form, which allows subsequent optimization passes
to do a significantly better job with them.  The pass is located in
@file{tree-sra.c} and is described by @code{pass_sra}.

@item Dead store elimination

This pass eliminates stores to memory that are subsequently overwritten
by another store, without any intervening loads.  The pass is located
in @file{tree-ssa-dse.c} and is described by @code{pass_dse}.

@item Tail recursion elimination

This pass transforms tail recursion into a loop.  It is located in
@file{tree-tailcall.c} and is described by @code{pass_tail_recursion}.

@item Forward store motion

This pass sinks stores and assignments down the flowgraph closer to their
use point.  The pass is located in @file{tree-ssa-sink.c} and is
described by @code{pass_sink_code}.

@item Partial redundancy elimination

This pass eliminates partially redundant computations, as well as
performing load motion.  The pass is located in @file{tree-ssa-pre.c}
and is described by @code{pass_pre}.

Just before partial redundancy elimination, if
@option{-funsafe-math-optimizations} is on, GCC tries to convert
divisions to multiplications by the reciprocal.  The pass is located
in @file{tree-ssa-math-opts.c} and is described by
@code{pass_cse_reciprocal}.

@item Full redundancy elimination

This is a simpler form of PRE that only eliminates redundancies that
occur an all paths.  It is located in @file{tree-ssa-pre.c} and
described by @code{pass_fre}.

@item Loop optimization

The main driver of the pass is placed in @file{tree-ssa-loop.c}
and described by @code{pass_loop}.

The optimizations performed by this pass are:

Loop invariant motion.  This pass moves only invariants that
would be hard to handle on RTL level (function calls, operations that expand to
nontrivial sequences of insns).  With @option{-funswitch-loops} it also moves
operands of conditions that are invariant out of the loop, so that we can use
just trivial invariantness analysis in loop unswitching.  The pass also includes
store motion.  The pass is implemented in @file{tree-ssa-loop-im.c}.

Canonical induction variable creation.  This pass creates a simple counter
for number of iterations of the loop and replaces the exit condition of the
loop using it, in case when a complicated analysis is necessary to determine
the number of iterations.  Later optimizations then may determine the number
easily.  The pass is implemented in @file{tree-ssa-loop-ivcanon.c}.

Induction variable optimizations.  This pass performs standard induction
variable optimizations, including strength reduction, induction variable
merging and induction variable elimination.  The pass is implemented in
@file{tree-ssa-loop-ivopts.c}.

Loop unswitching.  This pass moves the conditional jumps that are invariant
out of the loops.  To achieve this, a duplicate of the loop is created for
each possible outcome of conditional jump(s).  The pass is implemented in
@file{tree-ssa-loop-unswitch.c}.  This pass should eventually replace the
RTL level loop unswitching in @file{loop-unswitch.c}, but currently
the RTL level pass is not completely redundant yet due to deficiencies
in tree level alias analysis.

The optimizations also use various utility functions contained in
@file{tree-ssa-loop-manip.c}, @file{cfgloop.c}, @file{cfgloopanal.c} and
@file{cfgloopmanip.c}.

Vectorization.  This pass transforms loops to operate on vector types
instead of scalar types.  Data parallelism across loop iterations is exploited
to group data elements from consecutive iterations into a vector and operate 
on them in parallel.  Depending on available target support the loop is 
conceptually unrolled by a factor @code{VF} (vectorization factor), which is
the number of elements operated upon in parallel in each iteration, and the 
@code{VF} copies of each scalar operation are fused to form a vector operation.
Additional loop transformations such as peeling and versioning may take place
to align the number of iterations, and to align the memory accesses in the 
loop.
The pass is implemented in @file{tree-vectorizer.c} (the main driver),
@file{tree-vect-loop.c} and @file{tree-vect-loop-manip.c} (loop specific parts 
and general loop utilities), @file{tree-vect-slp} (loop-aware SLP 
functionality), @file{tree-vect-stmts.c} and @file{tree-vect-data-refs.c}.
Analysis of data references is in @file{tree-data-ref.c}.

SLP Vectorization.  This pass performs vectorization of straight-line code. The
pass is implemented in @file{tree-vectorizer.c} (the main driver),
@file{tree-vect-slp.c}, @file{tree-vect-stmts.c} and 
@file{tree-vect-data-refs.c}.

Autoparallelization.  This pass splits the loop iteration space to run
into several threads.  The pass is implemented in @file{tree-parloops.c}.

Graphite is a loop transformation framework based on the polyhedral
model.  Graphite stands for Gimple Represented as Polyhedra.  The
internals of this infrastructure are documented in
@w{@uref{http://gcc.gnu.org/wiki/Graphite}}.  The passes working on
this representation are implemented in the various @file{graphite-*}
files.

@item Tree level if-conversion for vectorizer

This pass applies if-conversion to simple loops to help vectorizer.
We identify if convertible loops, if-convert statements and merge
basic blocks in one big block.  The idea is to present loop in such
form so that vectorizer can have one to one mapping between statements
and available vector operations.  This patch re-introduces COND_EXPR
at GIMPLE level.  This pass is located in @file{tree-if-conv.c} and is
described by @code{pass_if_conversion}.

@item Conditional constant propagation

This pass relaxes a lattice of values in order to identify those
that must be constant even in the presence of conditional branches.
The pass is located in @file{tree-ssa-ccp.c} and is described
by @code{pass_ccp}.

A related pass that works on memory loads and stores, and not just
register values, is located in @file{tree-ssa-ccp.c} and described by
@code{pass_store_ccp}.

@item Conditional copy propagation

This is similar to constant propagation but the lattice of values is
the ``copy-of'' relation.  It eliminates redundant copies from the
code.  The pass is located in @file{tree-ssa-copy.c} and described by
@code{pass_copy_prop}.

A related pass that works on memory copies, and not just register
copies, is located in @file{tree-ssa-copy.c} and described by
@code{pass_store_copy_prop}.

@item Value range propagation

This transformation is similar to constant propagation but
instead of propagating single constant values, it propagates
known value ranges.  The implementation is based on Patterson's
range propagation algorithm (Accurate Static Branch Prediction by
Value Range Propagation, J. R. C. Patterson, PLDI '95).  In
contrast to Patterson's algorithm, this implementation does not
propagate branch probabilities nor it uses more than a single
range per SSA name. This means that the current implementation
cannot be used for branch prediction (though adapting it would
not be difficult).  The pass is located in @file{tree-vrp.c} and is
described by @code{pass_vrp}.

@item Folding built-in functions

This pass simplifies built-in functions, as applicable, with constant
arguments or with inferable string lengths.  It is located in
@file{tree-ssa-ccp.c} and is described by @code{pass_fold_builtins}.

@item Split critical edges

This pass identifies critical edges and inserts empty basic blocks
such that the edge is no longer critical.  The pass is located in
@file{tree-cfg.c} and is described by @code{pass_split_crit_edges}.

@item Control dependence dead code elimination

This pass is a stronger form of dead code elimination that can
eliminate unnecessary control flow statements.   It is located
in @file{tree-ssa-dce.c} and is described by @code{pass_cd_dce}.

@item Tail call elimination

This pass identifies function calls that may be rewritten into
jumps.  No code transformation is actually applied here, but the
data and control flow problem is solved.  The code transformation
requires target support, and so is delayed until RTL@.  In the
meantime @code{CALL_EXPR_TAILCALL} is set indicating the possibility.
The pass is located in @file{tree-tailcall.c} and is described by
@code{pass_tail_calls}.  The RTL transformation is handled by
@code{fixup_tail_calls} in @file{calls.c}.

@item Warn for function return without value

For non-void functions, this pass locates return statements that do
not specify a value and issues a warning.  Such a statement may have
been injected by falling off the end of the function.  This pass is
run last so that we have as much time as possible to prove that the
statement is not reachable.  It is located in @file{tree-cfg.c} and
is described by @code{pass_warn_function_return}.

@item Mudflap statement annotation

If mudflap is enabled, we rewrite some memory accesses with code to
validate that the memory access is correct.  In particular, expressions
involving pointer dereferences (@code{INDIRECT_REF}, @code{ARRAY_REF},
etc.) are replaced by code that checks the selected address range
against the mudflap runtime's database of valid regions.  This check
includes an inline lookup into a direct-mapped cache, based on
shift/mask operations of the pointer value, with a fallback function
call into the runtime.  The pass is located in @file{tree-mudflap.c} and
is described by @code{pass_mudflap_2}.

@item Leave static single assignment form

This pass rewrites the function such that it is in normal form.  At
the same time, we eliminate as many single-use temporaries as possible,
so the intermediate language is no longer GIMPLE, but GENERIC@.  The
pass is located in @file{tree-outof-ssa.c} and is described by
@code{pass_del_ssa}.

@item Merge PHI nodes that feed into one another

This is part of the CFG cleanup passes.  It attempts to join PHI nodes
from a forwarder CFG block into another block with PHI nodes.  The
pass is located in @file{tree-cfgcleanup.c} and is described by
@code{pass_merge_phi}.

@item Return value optimization

If a function always returns the same local variable, and that local
variable is an aggregate type, then the variable is replaced with the
return value for the function (i.e., the function's DECL_RESULT).  This
is equivalent to the C++ named return value optimization applied to
GIMPLE@.  The pass is located in @file{tree-nrv.c} and is described by
@code{pass_nrv}.

@item Return slot optimization

If a function returns a memory object and is called as @code{var =
foo()}, this pass tries to change the call so that the address of
@code{var} is sent to the caller to avoid an extra memory copy.  This
pass is located in @code{tree-nrv.c} and is described by
@code{pass_return_slot}.

@item Optimize calls to @code{__builtin_object_size}

This is a propagation pass similar to CCP that tries to remove calls
to @code{__builtin_object_size} when the size of the object can be
computed at compile-time.  This pass is located in
@file{tree-object-size.c} and is described by
@code{pass_object_sizes}.

@item Loop invariant motion

This pass removes expensive loop-invariant computations out of loops.
The pass is located in @file{tree-ssa-loop.c} and described by
@code{pass_lim}.

@item Loop nest optimizations

This is a family of loop transformations that works on loop nests.  It
includes loop interchange, scaling, skewing and reversal and they are
all geared to the optimization of data locality in array traversals
and the removal of dependencies that hamper optimizations such as loop
parallelization and vectorization.  The pass is located in
@file{tree-loop-linear.c} and described by
@code{pass_linear_transform}.

@item Removal of empty loops

This pass removes loops with no code in them.  The pass is located in
@file{tree-ssa-loop-ivcanon.c} and described by
@code{pass_empty_loop}.

@item Unrolling of small loops

This pass completely unrolls loops with few iterations.  The pass
is located in @file{tree-ssa-loop-ivcanon.c} and described by
@code{pass_complete_unroll}.

@item Predictive commoning

This pass makes the code reuse the computations from the previous
iterations of the loops, especially loads and stores to memory.
It does so by storing the values of these computations to a bank
of temporary variables that are rotated at the end of loop.  To avoid
the need for this rotation, the loop is then unrolled and the copies
of the loop body are rewritten to use the appropriate version of
the temporary variable.  This pass is located in @file{tree-predcom.c}
and described by @code{pass_predcom}.

@item Array prefetching

This pass issues prefetch instructions for array references inside
loops.  The pass is located in @file{tree-ssa-loop-prefetch.c} and
described by @code{pass_loop_prefetch}.

@item Reassociation

This pass rewrites arithmetic expressions to enable optimizations that
operate on them, like redundancy elimination and vectorization.  The
pass is located in @file{tree-ssa-reassoc.c} and described by
@code{pass_reassoc}.

@item Optimization of @code{stdarg} functions

This pass tries to avoid the saving of register arguments into the
stack on entry to @code{stdarg} functions.  If the function doesn't
use any @code{va_start} macros, no registers need to be saved.  If
@code{va_start} macros are used, the @code{va_list} variables don't
escape the function, it is only necessary to save registers that will
be used in @code{va_arg} macros.  For instance, if @code{va_arg} is
only used with integral types in the function, floating point
registers don't need to be saved.  This pass is located in
@code{tree-stdarg.c} and described by @code{pass_stdarg}.

@end itemize

@node RTL passes
@section RTL passes

The following briefly describes the RTL generation and optimization
passes that are run after the Tree optimization passes.

@itemize @bullet
@item RTL generation

@c Avoiding overfull is tricky here.
The source files for RTL generation include
@file{stmt.c},
@file{calls.c},
@file{expr.c},
@file{explow.c},
@file{expmed.c},
@file{function.c},
@file{optabs.c}
and @file{emit-rtl.c}.
Also, the file
@file{insn-emit.c}, generated from the machine description by the
program @code{genemit}, is used in this pass.  The header file
@file{expr.h} is used for communication within this pass.

@findex genflags
@findex gencodes
The header files @file{insn-flags.h} and @file{insn-codes.h},
generated from the machine description by the programs @code{genflags}
and @code{gencodes}, tell this pass which standard names are available
for use and which patterns correspond to them.

@item Generation of exception landing pads

This pass generates the glue that handles communication between the
exception handling library routines and the exception handlers within
the function.  Entry points in the function that are invoked by the
exception handling library are called @dfn{landing pads}.  The code
for this pass is located in @file{except.c}.

@item Control flow graph cleanup

This pass removes unreachable code, simplifies jumps to next, jumps to
jump, jumps across jumps, etc.  The pass is run multiple times.
For historical reasons, it is occasionally referred to as the ``jump
optimization pass''.  The bulk of the code for this pass is in
@file{cfgcleanup.c}, and there are support routines in @file{cfgrtl.c}
and @file{jump.c}.

@item Forward propagation of single-def values

This pass attempts to remove redundant computation by substituting
variables that come from a single definition, and
seeing if the result can be simplified.  It performs copy propagation
and addressing mode selection.  The pass is run twice, with values
being propagated into loops only on the second run.  The code is
located in @file{fwprop.c}.

@item Common subexpression elimination

This pass removes redundant computation within basic blocks, and
optimizes addressing modes based on cost.  The pass is run twice.
The code for this pass is located in @file{cse.c}.

@item Global common subexpression elimination

This pass performs two
different types of GCSE  depending on whether you are optimizing for
size or not (LCM based GCSE tends to increase code size for a gain in
speed, while Morel-Renvoise based GCSE does not).
When optimizing for size, GCSE is done using Morel-Renvoise Partial
Redundancy Elimination, with the exception that it does not try to move
invariants out of loops---that is left to  the loop optimization pass.
If MR PRE GCSE is done, code hoisting (aka unification) is also done, as
well as load motion.
If you are optimizing for speed, LCM (lazy code motion) based GCSE is
done.  LCM is based on the work of Knoop, Ruthing, and Steffen.  LCM
based GCSE also does loop invariant code motion.  We also perform load
and store motion when optimizing for speed.
Regardless of which type of GCSE is used, the GCSE pass also performs
global constant and  copy propagation.
The source file for this pass is @file{gcse.c}, and the LCM routines
are in @file{lcm.c}.

@item Loop optimization

This pass performs several loop related optimizations.
The source files @file{cfgloopanal.c} and @file{cfgloopmanip.c} contain
generic loop analysis and manipulation code.  Initialization and finalization
of loop structures is handled by @file{loop-init.c}.
A loop invariant motion pass is implemented in @file{loop-invariant.c}.
Basic block level optimizations---unrolling, peeling and unswitching loops---
are implemented in @file{loop-unswitch.c} and @file{loop-unroll.c}.
Replacing of the exit condition of loops by special machine-dependent
instructions is handled by @file{loop-doloop.c}.

@item Jump bypassing

This pass is an aggressive form of GCSE that transforms the control
flow graph of a function by propagating constants into conditional
branch instructions.  The source file for this pass is @file{gcse.c}.

@item If conversion

This pass attempts to replace conditional branches and surrounding
assignments with arithmetic, boolean value producing comparison
instructions, and conditional move instructions.  In the very last
invocation after reload, it will generate predicated instructions
when supported by the target.  The code is located in @file{ifcvt.c}.

@item Web construction

This pass splits independent uses of each pseudo-register.  This can
improve effect of the other transformation, such as CSE or register
allocation.  The code for this pass is located in @file{web.c}.

@item Instruction combination

This pass attempts to combine groups of two or three instructions that
are related by data flow into single instructions.  It combines the
RTL expressions for the instructions by substitution, simplifies the
result using algebra, and then attempts to match the result against
the machine description.  The code is located in @file{combine.c}.

@item Register movement

This pass looks for cases where matching constraints would force an
instruction to need a reload, and this reload would be a
register-to-register move.  It then attempts to change the registers
used by the instruction to avoid the move instruction.  The code is
located in @file{regmove.c}.

@item Mode switching optimization

This pass looks for instructions that require the processor to be in a
specific ``mode'' and minimizes the number of mode changes required to
satisfy all users.  What these modes are, and what they apply to are
completely target-specific.  The code for this pass is located in
@file{mode-switching.c}.

@cindex modulo scheduling
@cindex sms, swing, software pipelining
@item Modulo scheduling

This pass looks at innermost loops and reorders their instructions
by overlapping different iterations.  Modulo scheduling is performed
immediately before instruction scheduling.  The code for this pass is
located in @file{modulo-sched.c}.

@item Instruction scheduling

This pass looks for instructions whose output will not be available by
the time that it is used in subsequent instructions.  Memory loads and
floating point instructions often have this behavior on RISC machines.
It re-orders instructions within a basic block to try to separate the
definition and use of items that otherwise would cause pipeline
stalls.  This pass is performed twice, before and after register
allocation.  The code for this pass is located in @file{haifa-sched.c},
@file{sched-deps.c}, @file{sched-ebb.c}, @file{sched-rgn.c} and
@file{sched-vis.c}.

@item Register allocation

These passes make sure that all occurrences of pseudo registers are
eliminated, either by allocating them to a hard register, replacing
them by an equivalent expression (e.g.@: a constant) or by placing
them on the stack.  This is done in several subpasses:

@itemize @bullet
@item
Register move optimizations.  This pass makes some simple RTL code
transformations which improve the subsequent register allocation.  The
source file is @file{regmove.c}.

@item
The integrated register allocator (@acronym{IRA}).  It is called
integrated because coalescing, register live range splitting, and hard
register preferencing are done on-the-fly during coloring.  It also
has better integration with the reload pass.  Pseudo-registers spilled
by the allocator or the reload have still a chance to get
hard-registers if the reload evicts some pseudo-registers from
hard-registers.  The allocator helps to choose better pseudos for
spilling based on their live ranges and to coalesce stack slots
allocated for the spilled pseudo-registers.  IRA is a regional
register allocator which is transformed into Chaitin-Briggs allocator
if there is one region.  By default, IRA chooses regions using
register pressure but the user can force it to use one region or
regions corresponding to all loops.

Source files of the allocator are @file{ira.c}, @file{ira-build.c},
@file{ira-costs.c}, @file{ira-conflicts.c}, @file{ira-color.c},
@file{ira-emit.c}, @file{ira-lives}, plus header files @file{ira.h}
and @file{ira-int.h} used for the communication between the allocator
and the rest of the compiler and between the IRA files.

@cindex reloading
@item
Reloading.  This pass renumbers pseudo registers with the hardware
registers numbers they were allocated.  Pseudo registers that did not
get hard registers are replaced with stack slots.  Then it finds
instructions that are invalid because a value has failed to end up in
a register, or has ended up in a register of the wrong kind.  It fixes
up these instructions by reloading the problematical values
temporarily into registers.  Additional instructions are generated to
do the copying.

The reload pass also optionally eliminates the frame pointer and inserts
instructions to save and restore call-clobbered registers around calls.

Source files are @file{reload.c} and @file{reload1.c}, plus the header
@file{reload.h} used for communication between them.
@end itemize

@item Basic block reordering

This pass implements profile guided code positioning.  If profile
information is not available, various types of static analysis are
performed to make the predictions normally coming from the profile
feedback (IE execution frequency, branch probability, etc).  It is
implemented in the file @file{bb-reorder.c}, and the various
prediction routines are in @file{predict.c}.

@item Variable tracking

This pass computes where the variables are stored at each
position in code and generates notes describing the variable locations
to RTL code.  The location lists are then generated according to these
notes to debug information if the debugging information format supports
location lists.  The code is located in @file{var-tracking.c}.

@item Delayed branch scheduling

This optional pass attempts to find instructions that can go into the
delay slots of other instructions, usually jumps and calls.  The code
for this pass is located in @file{reorg.c}.

@item Branch shortening

On many RISC machines, branch instructions have a limited range.
Thus, longer sequences of instructions must be used for long branches.
In this pass, the compiler figures out what how far each instruction
will be from each other instruction, and therefore whether the usual
instructions, or the longer sequences, must be used for each branch.
The code for this pass is located in @file{final.c}.

@item Register-to-stack conversion

Conversion from usage of some hard registers to usage of a register
stack may be done at this point.  Currently, this is supported only
for the floating-point registers of the Intel 80387 coprocessor.  The
code for this pass is located in @file{reg-stack.c}.

@item Final

This pass outputs the assembler code for the function.  The source files
are @file{final.c} plus @file{insn-output.c}; the latter is generated
automatically from the machine description by the tool @file{genoutput}.
The header file @file{conditions.h} is used for communication between
these files.  If mudflap is enabled, the queue of deferred declarations
and any addressed constants (e.g., string literals) is processed by
@code{mudflap_finish_file} into a synthetic constructor function
containing calls into the mudflap runtime.

@item Debugging information output

This is run after final because it must output the stack slot offsets
for pseudo registers that did not get hard registers.  Source files
are @file{dbxout.c} for DBX symbol table format, @file{sdbout.c} for
SDB symbol table format, @file{dwarfout.c} for DWARF symbol table
format, files @file{dwarf2out.c} and @file{dwarf2asm.c} for DWARF2
symbol table format, and @file{vmsdbgout.c} for VMS debug symbol table
format.

@end itemize