1 /* Natural loop functions
2 Copyright (C) 1987-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
23 #include "double-int.h"
29 /* Structure to hold decision about unrolling/peeling. */
40 struct GTY (()) lpt_decision
{
41 enum lpt_dec decision
;
45 /* The type of extend applied to an IV. */
53 /* The structure describing a bound on number of iterations of a loop. */
55 struct GTY ((chain_next ("%h.next"))) nb_iter_bound
{
56 /* The statement STMT is executed at most ... */
59 /* ... BOUND + 1 times (BOUND must be an unsigned constant).
60 The + 1 is added for the following reasons:
62 a) 0 would otherwise be unused, while we would need to care more about
63 overflows (as MAX + 1 is sometimes produced as the estimate on number
64 of executions of STMT).
65 b) it is consistent with the result of number_of_iterations_exit. */
68 /* True if the statement will cause the loop to be leaved the (at most)
69 BOUND + 1-st time it is executed, that is, all the statements after it
70 are executed at most BOUND times. */
73 /* The next bound in the list. */
74 struct nb_iter_bound
*next
;
77 /* Description of the loop exit. */
79 struct GTY ((for_user
)) loop_exit
{
83 /* Previous and next exit in the list of the exits of the loop. */
84 struct loop_exit
*prev
;
85 struct loop_exit
*next
;
87 /* Next element in the list of loops from that E exits. */
88 struct loop_exit
*next_e
;
91 struct loop_exit_hasher
: ggc_hasher
<loop_exit
*>
93 typedef edge compare_type
;
95 static hashval_t
hash (loop_exit
*);
96 static bool equal (loop_exit
*, edge
);
97 static void remove (loop_exit
*);
100 typedef struct loop
*loop_p
;
102 /* An integer estimation of the number of iterations. Estimate_state
103 describes what is the state of the estimation. */
106 /* Estimate was not computed yet. */
108 /* Estimate is ready. */
113 /* Structure to hold information for each natural loop. */
114 struct GTY ((chain_next ("%h.next"))) loop
{
115 /* Index into loops array. */
118 /* Number of loop insns. */
121 /* Basic block of loop header. */
124 /* Basic block of loop latch. */
127 /* For loop unrolling/peeling decision. */
128 struct lpt_decision lpt_decision
;
130 /* Average number of executed insns per iteration. */
133 /* Number of blocks contained within the loop. */
136 /* Superloops of the loop, starting with the outermost loop. */
137 vec
<loop_p
, va_gc
> *superloops
;
139 /* The first inner (child) loop or NULL if innermost loop. */
142 /* Link to the next (sibling) loop. */
145 /* Auxiliary info specific to a pass. */
146 PTR
GTY ((skip (""))) aux
;
148 /* The number of times the latch of the loop is executed. This can be an
149 INTEGER_CST, or a symbolic expression representing the number of
150 iterations like "N - 1", or a COND_EXPR containing the runtime
151 conditions under which the number of iterations is non zero.
153 Don't access this field directly: number_of_latch_executions
154 computes and caches the computed information in this field. */
157 /* An integer guaranteed to be greater or equal to nb_iterations. Only
158 valid if any_upper_bound is true. */
159 widest_int nb_iterations_upper_bound
;
161 /* An integer giving an estimate on nb_iterations. Unlike
162 nb_iterations_upper_bound, there is no guarantee that it is at least
164 widest_int nb_iterations_estimate
;
166 bool any_upper_bound
;
169 /* True if the loop can be parallel. */
170 bool can_be_parallel
;
172 /* True if -Waggressive-loop-optimizations warned about this loop
174 bool warned_aggressive_loop_optimizations
;
176 /* An integer estimation of the number of iterations. Estimate_state
177 describes what is the state of the estimation. */
178 enum loop_estimation estimate_state
;
180 /* If > 0, an integer, where the user asserted that for any
181 I in [ 0, nb_iterations ) and for any J in
182 [ I, min ( I + safelen, nb_iterations ) ), the Ith and Jth iterations
183 of the loop can be safely evaluated concurrently. */
186 /* True if this loop should never be vectorized. */
189 /* True if we should try harder to vectorize this loop. */
190 bool force_vectorize
;
192 /* For SIMD loops, this is a unique identifier of the loop, referenced
193 by IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LANE and IFN_GOMP_SIMD_LAST_LANE
197 /* Upper bound on number of iterations of a loop. */
198 struct nb_iter_bound
*bounds
;
200 /* Head of the cyclic list of the exits of the loop. */
201 struct loop_exit
*exits
;
203 /* Number of iteration analysis data for RTL. */
204 struct niter_desc
*simple_loop_desc
;
206 /* For sanity checking during loop fixup we record here the former
207 loop header for loops marked for removal. Note that this prevents
208 the basic-block from being collected but its index can still be
210 basic_block former_header
;
213 /* Flags for state of loop structure. */
216 LOOPS_HAVE_PREHEADERS
= 1,
217 LOOPS_HAVE_SIMPLE_LATCHES
= 2,
218 LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
= 4,
219 LOOPS_HAVE_RECORDED_EXITS
= 8,
220 LOOPS_MAY_HAVE_MULTIPLE_LATCHES
= 16,
221 LOOP_CLOSED_SSA
= 32,
222 LOOPS_NEED_FIXUP
= 64,
223 LOOPS_HAVE_FALLTHRU_PREHEADERS
= 128
226 #define LOOPS_NORMAL (LOOPS_HAVE_PREHEADERS | LOOPS_HAVE_SIMPLE_LATCHES \
227 | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)
228 #define AVOID_CFG_MODIFICATIONS (LOOPS_MAY_HAVE_MULTIPLE_LATCHES)
230 /* Structure to hold CFG information about natural loops within a function. */
231 struct GTY (()) loops
{
232 /* State of loops. */
235 /* Array of the loops. */
236 vec
<loop_p
, va_gc
> *larray
;
238 /* Maps edges to the list of their descriptions as loop exits. Edges
239 whose sources or destinations have loop_father == NULL (which may
240 happen during the cfg manipulations) should not appear in EXITS. */
241 hash_table
<loop_exit_hasher
> *GTY(()) exits
;
243 /* Pointer to root of loop hierarchy tree. */
244 struct loop
*tree_root
;
247 /* Loop recognition. */
248 bool bb_loop_header_p (basic_block
);
249 void init_loops_structure (struct function
*, struct loops
*, unsigned);
250 extern struct loops
*flow_loops_find (struct loops
*);
251 extern void disambiguate_loops_with_multiple_latches (void);
252 extern void flow_loops_free (struct loops
*);
253 extern void flow_loops_dump (FILE *,
254 void (*)(const struct loop
*, FILE *, int), int);
255 extern void flow_loop_dump (const struct loop
*, FILE *,
256 void (*)(const struct loop
*, FILE *, int), int);
257 struct loop
*alloc_loop (void);
258 extern void flow_loop_free (struct loop
*);
259 int flow_loop_nodes_find (basic_block
, struct loop
*);
260 unsigned fix_loop_structure (bitmap changed_bbs
);
261 bool mark_irreducible_loops (void);
262 void release_recorded_exits (void);
263 void record_loop_exits (void);
264 void rescan_loop_exit (edge
, bool, bool);
266 /* Loop data structure manipulation/querying. */
267 extern void flow_loop_tree_node_add (struct loop
*, struct loop
*);
268 extern void flow_loop_tree_node_remove (struct loop
*);
269 extern void place_new_loop (struct function
*, struct loop
*);
270 extern void add_loop (struct loop
*, struct loop
*);
271 extern bool flow_loop_nested_p (const struct loop
*, const struct loop
*);
272 extern bool flow_bb_inside_loop_p (const struct loop
*, const_basic_block
);
273 extern struct loop
* find_common_loop (struct loop
*, struct loop
*);
274 struct loop
*superloop_at_depth (struct loop
*, unsigned);
275 struct eni_weights_d
;
276 extern int num_loop_insns (const struct loop
*);
277 extern int average_num_loop_insns (const struct loop
*);
278 extern unsigned get_loop_level (const struct loop
*);
279 extern bool loop_exit_edge_p (const struct loop
*, const_edge
);
280 extern bool loop_exits_to_bb_p (struct loop
*, basic_block
);
281 extern bool loop_exits_from_bb_p (struct loop
*, basic_block
);
282 extern void mark_loop_exit_edges (void);
283 extern location_t
get_loop_location (struct loop
*loop
);
285 /* Loops & cfg manipulation. */
286 extern basic_block
*get_loop_body (const struct loop
*);
287 extern unsigned get_loop_body_with_size (const struct loop
*, basic_block
*,
289 extern basic_block
*get_loop_body_in_dom_order (const struct loop
*);
290 extern basic_block
*get_loop_body_in_bfs_order (const struct loop
*);
291 extern basic_block
*get_loop_body_in_custom_order (const struct loop
*,
292 int (*) (const void *, const void *));
294 extern vec
<edge
> get_loop_exit_edges (const struct loop
*);
295 extern edge
single_exit (const struct loop
*);
296 extern edge
single_likely_exit (struct loop
*loop
);
297 extern unsigned num_loop_branches (const struct loop
*);
299 extern edge
loop_preheader_edge (const struct loop
*);
300 extern edge
loop_latch_edge (const struct loop
*);
302 extern void add_bb_to_loop (basic_block
, struct loop
*);
303 extern void remove_bb_from_loops (basic_block
);
305 extern void cancel_loop_tree (struct loop
*);
306 extern void delete_loop (struct loop
*);
310 CP_SIMPLE_PREHEADERS
= 1,
311 CP_FALLTHRU_PREHEADERS
= 2
314 basic_block
create_preheader (struct loop
*, int);
315 extern void create_preheaders (int);
316 extern void force_single_succ_latches (void);
318 extern void verify_loop_structure (void);
321 extern bool just_once_each_iteration_p (const struct loop
*, const_basic_block
);
322 gcov_type
expected_loop_iterations_unbounded (const struct loop
*);
323 extern unsigned expected_loop_iterations (const struct loop
*);
324 extern rtx
doloop_condition_get (rtx
);
327 /* Loop manipulation. */
328 extern bool can_duplicate_loop_p (const struct loop
*loop
);
330 #define DLTHE_FLAG_UPDATE_FREQ 1 /* Update frequencies in
331 duplicate_loop_to_header_edge. */
332 #define DLTHE_RECORD_COPY_NUMBER 2 /* Record copy number in the aux
333 field of newly create BB. */
334 #define DLTHE_FLAG_COMPLETTE_PEEL 4 /* Update frequencies expecting
335 a complete peeling. */
337 extern edge
create_empty_if_region_on_edge (edge
, tree
);
338 extern struct loop
*create_empty_loop_on_edge (edge
, tree
, tree
, tree
, tree
,
339 tree
*, tree
*, struct loop
*);
340 extern struct loop
* duplicate_loop (struct loop
*, struct loop
*);
341 extern void copy_loop_info (struct loop
*loop
, struct loop
*target
);
342 extern void duplicate_subloops (struct loop
*, struct loop
*);
343 extern bool duplicate_loop_to_header_edge (struct loop
*, edge
,
344 unsigned, sbitmap
, edge
,
346 extern struct loop
*loopify (edge
, edge
,
347 basic_block
, edge
, edge
, bool,
349 struct loop
* loop_version (struct loop
*, void *,
350 basic_block
*, unsigned, unsigned, unsigned, bool);
351 extern bool remove_path (edge
);
352 extern void unloop (struct loop
*, bool *, bitmap
);
353 extern void scale_loop_frequencies (struct loop
*, int, int);
354 void mark_loop_for_removal (loop_p
);
357 /* Induction variable analysis. */
359 /* The description of induction variable. The things are a bit complicated
360 due to need to handle subregs and extends. The value of the object described
361 by it can be obtained as follows (all computations are done in extend_mode):
363 Value in i-th iteration is
364 delta + mult * extend_{extend_mode} (subreg_{mode} (base + i * step)).
366 If first_special is true, the value in the first iteration is
369 If extend = UNKNOWN, first_special must be false, delta 0, mult 1 and value is
370 subreg_{mode} (base + i * step)
372 The get_iv_value function can be used to obtain these expressions.
374 ??? Add a third mode field that would specify the mode in that inner
375 computation is done, which would enable it to be different from the
380 /* Its base and step (mode of base and step is supposed to be extend_mode,
381 see the description above). */
384 /* The type of extend applied to it (IV_SIGN_EXTEND, IV_ZERO_EXTEND,
385 or IV_UNKNOWN_EXTEND). */
386 enum iv_extend_code extend
;
388 /* Operations applied in the extended mode. */
391 /* The mode it is extended to. */
392 enum machine_mode extend_mode
;
394 /* The mode the variable iterates in. */
395 enum machine_mode mode
;
397 /* Whether the first iteration needs to be handled specially. */
398 unsigned first_special
: 1;
401 /* The description of an exit from the loop and of the number of iterations
402 till we take the exit. */
404 struct GTY(()) niter_desc
406 /* The edge out of the loop. */
409 /* The other edge leading from the condition. */
412 /* True if we are able to say anything about number of iterations of the
416 /* True if the loop iterates the constant number of times. */
419 /* Number of iterations if constant. */
422 /* Assumptions under that the rest of the information is valid. */
425 /* Assumptions under that the loop ends before reaching the latch,
426 even if value of niter_expr says otherwise. */
427 rtx noloop_assumptions
;
429 /* Condition under that the loop is infinite. */
432 /* Whether the comparison is signed. */
435 /* The mode in that niter_expr should be computed. */
436 enum machine_mode mode
;
438 /* The number of iterations of the loop. */
442 extern void iv_analysis_loop_init (struct loop
*);
443 extern bool iv_analyze (rtx_insn
*, rtx
, struct rtx_iv
*);
444 extern bool iv_analyze_result (rtx_insn
*, rtx
, struct rtx_iv
*);
445 extern bool iv_analyze_expr (rtx_insn
*, rtx
, enum machine_mode
,
447 extern rtx
get_iv_value (struct rtx_iv
*, rtx
);
448 extern bool biv_p (rtx_insn
*, rtx
);
449 extern void find_simple_exit (struct loop
*, struct niter_desc
*);
450 extern void iv_analysis_done (void);
452 extern struct niter_desc
*get_simple_loop_desc (struct loop
*loop
);
453 extern void free_simple_loop_desc (struct loop
*loop
);
455 static inline struct niter_desc
*
456 simple_loop_desc (struct loop
*loop
)
458 return loop
->simple_loop_desc
;
461 /* Accessors for the loop structures. */
463 /* Returns the loop with index NUM from FNs loop tree. */
465 static inline struct loop
*
466 get_loop (struct function
*fn
, unsigned num
)
468 return (*loops_for_fn (fn
)->larray
)[num
];
471 /* Returns the number of superloops of LOOP. */
473 static inline unsigned
474 loop_depth (const struct loop
*loop
)
476 return vec_safe_length (loop
->superloops
);
479 /* Returns the immediate superloop of LOOP, or NULL if LOOP is the outermost
482 static inline struct loop
*
483 loop_outer (const struct loop
*loop
)
485 unsigned n
= vec_safe_length (loop
->superloops
);
490 return (*loop
->superloops
)[n
- 1];
493 /* Returns true if LOOP has at least one exit edge. */
496 loop_has_exit_edges (const struct loop
*loop
)
498 return loop
->exits
->next
->e
!= NULL
;
501 /* Returns the list of loops in FN. */
503 inline vec
<loop_p
, va_gc
> *
504 get_loops (struct function
*fn
)
506 struct loops
*loops
= loops_for_fn (fn
);
510 return loops
->larray
;
513 /* Returns the number of loops in FN (including the removed
514 ones and the fake loop that forms the root of the loop tree). */
516 static inline unsigned
517 number_of_loops (struct function
*fn
)
519 struct loops
*loops
= loops_for_fn (fn
);
523 return vec_safe_length (loops
->larray
);
526 /* Returns true if state of the loops satisfies all properties
527 described by FLAGS. */
530 loops_state_satisfies_p (unsigned flags
)
532 return (current_loops
->state
& flags
) == flags
;
535 /* Sets FLAGS to the loops state. */
538 loops_state_set (unsigned flags
)
540 current_loops
->state
|= flags
;
543 /* Clears FLAGS from the loops state. */
546 loops_state_clear (unsigned flags
)
550 current_loops
->state
&= ~flags
;
553 /* Loop iterators. */
555 /* Flags for loop iteration. */
559 LI_INCLUDE_ROOT
= 1, /* Include the fake root of the loop tree. */
560 LI_FROM_INNERMOST
= 2, /* Iterate over the loops in the reverse order,
561 starting from innermost ones. */
562 LI_ONLY_INNERMOST
= 4 /* Iterate only over innermost loops. */
565 /* The iterator for loops. */
569 loop_iterator (loop_p
*loop
, unsigned flags
);
572 inline loop_p
next ();
574 /* The list of loops to visit. */
577 /* The index of the actual loop. */
582 loop_iterator::next ()
586 while (this->to_visit
.iterate (this->idx
, &anum
))
589 loop_p loop
= get_loop (cfun
, anum
);
598 loop_iterator::loop_iterator (loop_p
*loop
, unsigned flags
)
607 this->to_visit
.create (0);
612 this->to_visit
.create (number_of_loops (cfun
));
613 mn
= (flags
& LI_INCLUDE_ROOT
) ? 0 : 1;
615 if (flags
& LI_ONLY_INNERMOST
)
617 for (i
= 0; vec_safe_iterate (current_loops
->larray
, i
, &aloop
); i
++)
619 && aloop
->inner
== NULL
621 this->to_visit
.quick_push (aloop
->num
);
623 else if (flags
& LI_FROM_INNERMOST
)
625 /* Push the loops to LI->TO_VISIT in postorder. */
626 for (aloop
= current_loops
->tree_root
;
627 aloop
->inner
!= NULL
;
628 aloop
= aloop
->inner
)
633 if (aloop
->num
>= mn
)
634 this->to_visit
.quick_push (aloop
->num
);
638 for (aloop
= aloop
->next
;
639 aloop
->inner
!= NULL
;
640 aloop
= aloop
->inner
)
643 else if (!loop_outer (aloop
))
646 aloop
= loop_outer (aloop
);
651 /* Push the loops to LI->TO_VISIT in preorder. */
652 aloop
= current_loops
->tree_root
;
655 if (aloop
->num
>= mn
)
656 this->to_visit
.quick_push (aloop
->num
);
658 if (aloop
->inner
!= NULL
)
659 aloop
= aloop
->inner
;
662 while (aloop
!= NULL
&& aloop
->next
== NULL
)
663 aloop
= loop_outer (aloop
);
671 *loop
= this->next ();
675 loop_iterator::~loop_iterator ()
677 this->to_visit
.release ();
680 #define FOR_EACH_LOOP(LOOP, FLAGS) \
681 for (loop_iterator li(&(LOOP), FLAGS); \
685 /* The properties of the target. */
686 struct target_cfgloop
{
687 /* Number of available registers. */
688 unsigned x_target_avail_regs
;
690 /* Number of available registers that are call-clobbered. */
691 unsigned x_target_clobbered_regs
;
693 /* Number of registers reserved for temporary expressions. */
694 unsigned x_target_res_regs
;
696 /* The cost for register when there still is some reserve, but we are
697 approaching the number of available registers. */
698 unsigned x_target_reg_cost
[2];
700 /* The cost for register when we need to spill. */
701 unsigned x_target_spill_cost
[2];
704 extern struct target_cfgloop default_target_cfgloop
;
705 #if SWITCHABLE_TARGET
706 extern struct target_cfgloop
*this_target_cfgloop
;
708 #define this_target_cfgloop (&default_target_cfgloop)
711 #define target_avail_regs \
712 (this_target_cfgloop->x_target_avail_regs)
713 #define target_clobbered_regs \
714 (this_target_cfgloop->x_target_clobbered_regs)
715 #define target_res_regs \
716 (this_target_cfgloop->x_target_res_regs)
717 #define target_reg_cost \
718 (this_target_cfgloop->x_target_reg_cost)
719 #define target_spill_cost \
720 (this_target_cfgloop->x_target_spill_cost)
722 /* Register pressure estimation for induction variable optimizations & loop
724 extern unsigned estimate_reg_pressure_cost (unsigned, unsigned, bool, bool);
725 extern void init_set_costs (void);
727 /* Loop optimizer initialization. */
728 extern void loop_optimizer_init (unsigned);
729 extern void loop_optimizer_finalize (void);
731 /* Optimization passes. */
734 UAP_PEEL
= 1, /* Enables loop peeling. */
735 UAP_UNROLL
= 2, /* Enables unrolling of loops if it seems profitable. */
736 UAP_UNROLL_ALL
= 4 /* Enables unrolling of all loops. */
739 extern void unroll_and_peel_loops (int);
740 extern void doloop_optimize_loops (void);
741 extern void move_loop_invariants (void);
742 extern void scale_loop_profile (struct loop
*loop
, int scale
, gcov_type iteration_bound
);
743 extern vec
<basic_block
> get_loop_hot_path (const struct loop
*loop
);
745 /* Returns the outermost loop of the loop nest that contains LOOP.*/
746 static inline struct loop
*
747 loop_outermost (struct loop
*loop
)
749 unsigned n
= vec_safe_length (loop
->superloops
);
754 return (*loop
->superloops
)[1];
757 extern void record_niter_bound (struct loop
*, const widest_int
&, bool, bool);
758 extern HOST_WIDE_INT
get_estimated_loop_iterations_int (struct loop
*);
759 extern HOST_WIDE_INT
get_max_loop_iterations_int (struct loop
*);
760 extern bool get_estimated_loop_iterations (struct loop
*loop
, widest_int
*nit
);
761 extern bool get_max_loop_iterations (struct loop
*loop
, widest_int
*nit
);
762 extern int bb_loop_depth (const_basic_block
);
764 /* Converts VAL to widest_int. */
766 static inline widest_int
767 gcov_type_to_wide_int (gcov_type val
)
771 a
[0] = (unsigned HOST_WIDE_INT
) val
;
772 /* If HOST_BITS_PER_WIDE_INT == HOST_BITS_PER_WIDEST_INT, avoid shifting by
774 val
>>= HOST_BITS_PER_WIDE_INT
- 1;
776 a
[1] = (unsigned HOST_WIDE_INT
) val
;
778 return widest_int::from_array (a
, 2);
780 #endif /* GCC_CFGLOOP_H */