1 /* Natural loop functions
2 Copyright (C) 1987-2016 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 "cfgloopmanip.h"
25 /* Structure to hold decision about unrolling/peeling. */
34 struct GTY (()) lpt_decision
{
35 enum lpt_dec decision
;
39 /* The type of extend applied to an IV. */
47 /* The structure describing a bound on number of iterations of a loop. */
49 struct GTY ((chain_next ("%h.next"))) nb_iter_bound
{
50 /* The statement STMT is executed at most ... */
53 /* ... BOUND + 1 times (BOUND must be an unsigned constant).
54 The + 1 is added for the following reasons:
56 a) 0 would otherwise be unused, while we would need to care more about
57 overflows (as MAX + 1 is sometimes produced as the estimate on number
58 of executions of STMT).
59 b) it is consistent with the result of number_of_iterations_exit. */
62 /* True if the statement will cause the loop to be leaved the (at most)
63 BOUND + 1-st time it is executed, that is, all the statements after it
64 are executed at most BOUND times. */
67 /* The next bound in the list. */
68 struct nb_iter_bound
*next
;
71 /* Description of the loop exit. */
73 struct GTY ((for_user
)) loop_exit
{
77 /* Previous and next exit in the list of the exits of the loop. */
78 struct loop_exit
*prev
;
79 struct loop_exit
*next
;
81 /* Next element in the list of loops from that E exits. */
82 struct loop_exit
*next_e
;
85 struct loop_exit_hasher
: ggc_ptr_hash
<loop_exit
>
87 typedef edge compare_type
;
89 static hashval_t
hash (loop_exit
*);
90 static bool equal (loop_exit
*, edge
);
91 static void remove (loop_exit
*);
94 typedef struct loop
*loop_p
;
96 /* An integer estimation of the number of iterations. Estimate_state
97 describes what is the state of the estimation. */
100 /* Estimate was not computed yet. */
102 /* Estimate is ready. */
107 /* The structure describing non-overflow control induction variable for
109 struct GTY ((chain_next ("%h.next"))) control_iv
{
112 struct control_iv
*next
;
115 /* Structure to hold information for each natural loop. */
116 struct GTY ((chain_next ("%h.next"))) loop
{
117 /* Index into loops array. */
120 /* Number of loop insns. */
123 /* Basic block of loop header. */
126 /* Basic block of loop latch. */
129 /* For loop unrolling/peeling decision. */
130 struct lpt_decision lpt_decision
;
132 /* Average number of executed insns per iteration. */
135 /* Number of blocks contained within the loop. */
138 /* Superloops of the loop, starting with the outermost loop. */
139 vec
<loop_p
, va_gc
> *superloops
;
141 /* The first inner (child) loop or NULL if innermost loop. */
144 /* Link to the next (sibling) loop. */
147 /* Auxiliary info specific to a pass. */
148 PTR
GTY ((skip (""))) aux
;
150 /* The number of times the latch of the loop is executed. This can be an
151 INTEGER_CST, or a symbolic expression representing the number of
152 iterations like "N - 1", or a COND_EXPR containing the runtime
153 conditions under which the number of iterations is non zero.
155 Don't access this field directly: number_of_latch_executions
156 computes and caches the computed information in this field. */
159 /* An integer guaranteed to be greater or equal to nb_iterations. Only
160 valid if any_upper_bound is true. */
161 widest_int nb_iterations_upper_bound
;
163 widest_int nb_iterations_likely_upper_bound
;
165 /* An integer giving an estimate on nb_iterations. Unlike
166 nb_iterations_upper_bound, there is no guarantee that it is at least
168 widest_int nb_iterations_estimate
;
170 bool any_upper_bound
;
172 bool any_likely_upper_bound
;
174 /* True if the loop can be parallel. */
175 bool can_be_parallel
;
177 /* True if -Waggressive-loop-optimizations warned about this loop
179 bool warned_aggressive_loop_optimizations
;
181 /* An integer estimation of the number of iterations. Estimate_state
182 describes what is the state of the estimation. */
183 enum loop_estimation estimate_state
;
185 /* If > 0, an integer, where the user asserted that for any
186 I in [ 0, nb_iterations ) and for any J in
187 [ I, min ( I + safelen, nb_iterations ) ), the Ith and Jth iterations
188 of the loop can be safely evaluated concurrently. */
191 /* True if this loop should never be vectorized. */
194 /* True if we should try harder to vectorize this loop. */
195 bool force_vectorize
;
197 /* True if the loop is part of an oacc kernels region. */
198 bool in_oacc_kernels_region
;
200 /* For SIMD loops, this is a unique identifier of the loop, referenced
201 by IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LANE and IFN_GOMP_SIMD_LAST_LANE
205 /* Upper bound on number of iterations of a loop. */
206 struct nb_iter_bound
*bounds
;
208 /* Non-overflow control ivs of a loop. */
209 struct control_iv
*control_ivs
;
211 /* Head of the cyclic list of the exits of the loop. */
212 struct loop_exit
*exits
;
214 /* Number of iteration analysis data for RTL. */
215 struct niter_desc
*simple_loop_desc
;
217 /* For sanity checking during loop fixup we record here the former
218 loop header for loops marked for removal. Note that this prevents
219 the basic-block from being collected but its index can still be
221 basic_block former_header
;
224 /* Flags for state of loop structure. */
227 LOOPS_HAVE_PREHEADERS
= 1,
228 LOOPS_HAVE_SIMPLE_LATCHES
= 2,
229 LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
= 4,
230 LOOPS_HAVE_RECORDED_EXITS
= 8,
231 LOOPS_MAY_HAVE_MULTIPLE_LATCHES
= 16,
232 LOOP_CLOSED_SSA
= 32,
233 LOOPS_NEED_FIXUP
= 64,
234 LOOPS_HAVE_FALLTHRU_PREHEADERS
= 128
237 #define LOOPS_NORMAL (LOOPS_HAVE_PREHEADERS | LOOPS_HAVE_SIMPLE_LATCHES \
238 | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)
239 #define AVOID_CFG_MODIFICATIONS (LOOPS_MAY_HAVE_MULTIPLE_LATCHES)
241 /* Structure to hold CFG information about natural loops within a function. */
242 struct GTY (()) loops
{
243 /* State of loops. */
246 /* Array of the loops. */
247 vec
<loop_p
, va_gc
> *larray
;
249 /* Maps edges to the list of their descriptions as loop exits. Edges
250 whose sources or destinations have loop_father == NULL (which may
251 happen during the cfg manipulations) should not appear in EXITS. */
252 hash_table
<loop_exit_hasher
> *GTY(()) exits
;
254 /* Pointer to root of loop hierarchy tree. */
255 struct loop
*tree_root
;
258 /* Loop recognition. */
259 bool bb_loop_header_p (basic_block
);
260 void init_loops_structure (struct function
*, struct loops
*, unsigned);
261 extern struct loops
*flow_loops_find (struct loops
*);
262 extern void disambiguate_loops_with_multiple_latches (void);
263 extern void flow_loops_free (struct loops
*);
264 extern void flow_loops_dump (FILE *,
265 void (*)(const struct loop
*, FILE *, int), int);
266 extern void flow_loop_dump (const struct loop
*, FILE *,
267 void (*)(const struct loop
*, FILE *, int), int);
268 struct loop
*alloc_loop (void);
269 extern void flow_loop_free (struct loop
*);
270 int flow_loop_nodes_find (basic_block
, struct loop
*);
271 unsigned fix_loop_structure (bitmap changed_bbs
);
272 bool mark_irreducible_loops (void);
273 void release_recorded_exits (function
*);
274 void record_loop_exits (void);
275 void rescan_loop_exit (edge
, bool, bool);
277 /* Loop data structure manipulation/querying. */
278 extern void flow_loop_tree_node_add (struct loop
*, struct loop
*);
279 extern void flow_loop_tree_node_remove (struct loop
*);
280 extern bool flow_loop_nested_p (const struct loop
*, const struct loop
*);
281 extern bool flow_bb_inside_loop_p (const struct loop
*, const_basic_block
);
282 extern struct loop
* find_common_loop (struct loop
*, struct loop
*);
283 struct loop
*superloop_at_depth (struct loop
*, unsigned);
285 extern int num_loop_insns (const struct loop
*);
286 extern int average_num_loop_insns (const struct loop
*);
287 extern unsigned get_loop_level (const struct loop
*);
288 extern bool loop_exit_edge_p (const struct loop
*, const_edge
);
289 extern bool loop_exits_to_bb_p (struct loop
*, basic_block
);
290 extern bool loop_exits_from_bb_p (struct loop
*, basic_block
);
291 extern void mark_loop_exit_edges (void);
292 extern location_t
get_loop_location (struct loop
*loop
);
294 /* Loops & cfg manipulation. */
295 extern basic_block
*get_loop_body (const struct loop
*);
296 extern unsigned get_loop_body_with_size (const struct loop
*, basic_block
*,
298 extern basic_block
*get_loop_body_in_dom_order (const struct loop
*);
299 extern basic_block
*get_loop_body_in_bfs_order (const struct loop
*);
300 extern basic_block
*get_loop_body_in_custom_order (const struct loop
*,
301 int (*) (const void *, const void *));
303 extern vec
<edge
> get_loop_exit_edges (const struct loop
*);
304 extern edge
single_exit (const struct loop
*);
305 extern edge
single_likely_exit (struct loop
*loop
);
306 extern unsigned num_loop_branches (const struct loop
*);
308 extern edge
loop_preheader_edge (const struct loop
*);
309 extern edge
loop_latch_edge (const struct loop
*);
311 extern void add_bb_to_loop (basic_block
, struct loop
*);
312 extern void remove_bb_from_loops (basic_block
);
314 extern void cancel_loop_tree (struct loop
*);
315 extern void delete_loop (struct loop
*);
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 bool *read_profile_p
= NULL
);
324 extern unsigned expected_loop_iterations (struct loop
*);
325 extern rtx
doloop_condition_get (rtx
);
327 void mark_loop_for_removal (loop_p
);
329 /* Induction variable analysis. */
331 /* The description of induction variable. The things are a bit complicated
332 due to need to handle subregs and extends. The value of the object described
333 by it can be obtained as follows (all computations are done in extend_mode):
335 Value in i-th iteration is
336 delta + mult * extend_{extend_mode} (subreg_{mode} (base + i * step)).
338 If first_special is true, the value in the first iteration is
341 If extend = UNKNOWN, first_special must be false, delta 0, mult 1 and value is
342 subreg_{mode} (base + i * step)
344 The get_iv_value function can be used to obtain these expressions.
346 ??? Add a third mode field that would specify the mode in that inner
347 computation is done, which would enable it to be different from the
352 /* Its base and step (mode of base and step is supposed to be extend_mode,
353 see the description above). */
356 /* The type of extend applied to it (IV_SIGN_EXTEND, IV_ZERO_EXTEND,
357 or IV_UNKNOWN_EXTEND). */
358 enum iv_extend_code extend
;
360 /* Operations applied in the extended mode. */
363 /* The mode it is extended to. */
364 machine_mode extend_mode
;
366 /* The mode the variable iterates in. */
369 /* Whether the first iteration needs to be handled specially. */
370 unsigned first_special
: 1;
373 /* The description of an exit from the loop and of the number of iterations
374 till we take the exit. */
376 struct GTY(()) niter_desc
378 /* The edge out of the loop. */
381 /* The other edge leading from the condition. */
384 /* True if we are able to say anything about number of iterations of the
388 /* True if the loop iterates the constant number of times. */
391 /* Number of iterations if constant. */
394 /* Assumptions under that the rest of the information is valid. */
397 /* Assumptions under that the loop ends before reaching the latch,
398 even if value of niter_expr says otherwise. */
399 rtx noloop_assumptions
;
401 /* Condition under that the loop is infinite. */
404 /* Whether the comparison is signed. */
407 /* The mode in that niter_expr should be computed. */
410 /* The number of iterations of the loop. */
414 extern void iv_analysis_loop_init (struct loop
*);
415 extern bool iv_analyze (rtx_insn
*, rtx
, struct rtx_iv
*);
416 extern bool iv_analyze_result (rtx_insn
*, rtx
, struct rtx_iv
*);
417 extern bool iv_analyze_expr (rtx_insn
*, rtx
, machine_mode
,
419 extern rtx
get_iv_value (struct rtx_iv
*, rtx
);
420 extern bool biv_p (rtx_insn
*, rtx
);
421 extern void find_simple_exit (struct loop
*, struct niter_desc
*);
422 extern void iv_analysis_done (void);
424 extern struct niter_desc
*get_simple_loop_desc (struct loop
*loop
);
425 extern void free_simple_loop_desc (struct loop
*loop
);
427 static inline struct niter_desc
*
428 simple_loop_desc (struct loop
*loop
)
430 return loop
->simple_loop_desc
;
433 /* Accessors for the loop structures. */
435 /* Returns the loop with index NUM from FNs loop tree. */
437 static inline struct loop
*
438 get_loop (struct function
*fn
, unsigned num
)
440 return (*loops_for_fn (fn
)->larray
)[num
];
443 /* Returns the number of superloops of LOOP. */
445 static inline unsigned
446 loop_depth (const struct loop
*loop
)
448 return vec_safe_length (loop
->superloops
);
451 /* Returns the immediate superloop of LOOP, or NULL if LOOP is the outermost
454 static inline struct loop
*
455 loop_outer (const struct loop
*loop
)
457 unsigned n
= vec_safe_length (loop
->superloops
);
462 return (*loop
->superloops
)[n
- 1];
465 /* Returns true if LOOP has at least one exit edge. */
468 loop_has_exit_edges (const struct loop
*loop
)
470 return loop
->exits
->next
->e
!= NULL
;
473 /* Returns the list of loops in FN. */
475 inline vec
<loop_p
, va_gc
> *
476 get_loops (struct function
*fn
)
478 struct loops
*loops
= loops_for_fn (fn
);
482 return loops
->larray
;
485 /* Returns the number of loops in FN (including the removed
486 ones and the fake loop that forms the root of the loop tree). */
488 static inline unsigned
489 number_of_loops (struct function
*fn
)
491 struct loops
*loops
= loops_for_fn (fn
);
495 return vec_safe_length (loops
->larray
);
498 /* Returns true if state of the loops satisfies all properties
499 described by FLAGS. */
502 loops_state_satisfies_p (function
*fn
, unsigned flags
)
504 return (loops_for_fn (fn
)->state
& flags
) == flags
;
508 loops_state_satisfies_p (unsigned flags
)
510 return loops_state_satisfies_p (cfun
, flags
);
513 /* Sets FLAGS to the loops state. */
516 loops_state_set (function
*fn
, unsigned flags
)
518 loops_for_fn (fn
)->state
|= flags
;
522 loops_state_set (unsigned flags
)
524 loops_state_set (cfun
, flags
);
527 /* Clears FLAGS from the loops state. */
530 loops_state_clear (function
*fn
, unsigned flags
)
532 loops_for_fn (fn
)->state
&= ~flags
;
536 loops_state_clear (unsigned flags
)
540 loops_state_clear (cfun
, flags
);
543 /* Check loop structure invariants, if internal consistency checks are
547 checking_verify_loop_structure (void)
549 /* VERIFY_LOOP_STRUCTURE essentially asserts that no loops need fixups.
551 The loop optimizers should never make changes to the CFG which
552 require loop fixups. But the low level CFG manipulation code may
553 set the flag conservatively.
555 Go ahead and clear the flag here. That avoids the assert inside
556 VERIFY_LOOP_STRUCTURE, and if there is an inconsistency in the loop
557 structures VERIFY_LOOP_STRUCTURE will detect it.
559 This also avoid the compile time cost of excessive fixups. */
560 loops_state_clear (LOOPS_NEED_FIXUP
);
562 verify_loop_structure ();
565 /* Loop iterators. */
567 /* Flags for loop iteration. */
571 LI_INCLUDE_ROOT
= 1, /* Include the fake root of the loop tree. */
572 LI_FROM_INNERMOST
= 2, /* Iterate over the loops in the reverse order,
573 starting from innermost ones. */
574 LI_ONLY_INNERMOST
= 4 /* Iterate only over innermost loops. */
577 /* The iterator for loops. */
581 loop_iterator (function
*fn
, loop_p
*loop
, unsigned flags
);
584 inline loop_p
next ();
586 /* The function we are visiting. */
589 /* The list of loops to visit. */
592 /* The index of the actual loop. */
597 loop_iterator::next ()
601 while (this->to_visit
.iterate (this->idx
, &anum
))
604 loop_p loop
= get_loop (fn
, anum
);
613 loop_iterator::loop_iterator (function
*fn
, loop_p
*loop
, unsigned flags
)
621 if (!loops_for_fn (fn
))
623 this->to_visit
.create (0);
628 this->to_visit
.create (number_of_loops (fn
));
629 mn
= (flags
& LI_INCLUDE_ROOT
) ? 0 : 1;
631 if (flags
& LI_ONLY_INNERMOST
)
633 for (i
= 0; vec_safe_iterate (loops_for_fn (fn
)->larray
, i
, &aloop
); i
++)
635 && aloop
->inner
== NULL
637 this->to_visit
.quick_push (aloop
->num
);
639 else if (flags
& LI_FROM_INNERMOST
)
641 /* Push the loops to LI->TO_VISIT in postorder. */
642 for (aloop
= loops_for_fn (fn
)->tree_root
;
643 aloop
->inner
!= NULL
;
644 aloop
= aloop
->inner
)
649 if (aloop
->num
>= mn
)
650 this->to_visit
.quick_push (aloop
->num
);
654 for (aloop
= aloop
->next
;
655 aloop
->inner
!= NULL
;
656 aloop
= aloop
->inner
)
659 else if (!loop_outer (aloop
))
662 aloop
= loop_outer (aloop
);
667 /* Push the loops to LI->TO_VISIT in preorder. */
668 aloop
= loops_for_fn (fn
)->tree_root
;
671 if (aloop
->num
>= mn
)
672 this->to_visit
.quick_push (aloop
->num
);
674 if (aloop
->inner
!= NULL
)
675 aloop
= aloop
->inner
;
678 while (aloop
!= NULL
&& aloop
->next
== NULL
)
679 aloop
= loop_outer (aloop
);
687 *loop
= this->next ();
691 loop_iterator::~loop_iterator ()
693 this->to_visit
.release ();
696 #define FOR_EACH_LOOP(LOOP, FLAGS) \
697 for (loop_iterator li(cfun, &(LOOP), FLAGS); \
701 #define FOR_EACH_LOOP_FN(FN, LOOP, FLAGS) \
702 for (loop_iterator li(fn, &(LOOP), FLAGS); \
706 /* The properties of the target. */
707 struct target_cfgloop
{
708 /* Number of available registers. */
709 unsigned x_target_avail_regs
;
711 /* Number of available registers that are call-clobbered. */
712 unsigned x_target_clobbered_regs
;
714 /* Number of registers reserved for temporary expressions. */
715 unsigned x_target_res_regs
;
717 /* The cost for register when there still is some reserve, but we are
718 approaching the number of available registers. */
719 unsigned x_target_reg_cost
[2];
721 /* The cost for register when we need to spill. */
722 unsigned x_target_spill_cost
[2];
725 extern struct target_cfgloop default_target_cfgloop
;
726 #if SWITCHABLE_TARGET
727 extern struct target_cfgloop
*this_target_cfgloop
;
729 #define this_target_cfgloop (&default_target_cfgloop)
732 #define target_avail_regs \
733 (this_target_cfgloop->x_target_avail_regs)
734 #define target_clobbered_regs \
735 (this_target_cfgloop->x_target_clobbered_regs)
736 #define target_res_regs \
737 (this_target_cfgloop->x_target_res_regs)
738 #define target_reg_cost \
739 (this_target_cfgloop->x_target_reg_cost)
740 #define target_spill_cost \
741 (this_target_cfgloop->x_target_spill_cost)
743 /* Register pressure estimation for induction variable optimizations & loop
745 extern unsigned estimate_reg_pressure_cost (unsigned, unsigned, bool, bool);
746 extern void init_set_costs (void);
748 /* Loop optimizer initialization. */
749 extern void loop_optimizer_init (unsigned);
750 extern void loop_optimizer_finalize (function
*);
752 loop_optimizer_finalize ()
754 loop_optimizer_finalize (cfun
);
757 /* Optimization passes. */
760 UAP_UNROLL
= 1, /* Enables unrolling of loops if it seems profitable. */
761 UAP_UNROLL_ALL
= 2 /* Enables unrolling of all loops. */
764 extern void doloop_optimize_loops (void);
765 extern void move_loop_invariants (void);
766 extern vec
<basic_block
> get_loop_hot_path (const struct loop
*loop
);
768 /* Returns the outermost loop of the loop nest that contains LOOP.*/
769 static inline struct loop
*
770 loop_outermost (struct loop
*loop
)
772 unsigned n
= vec_safe_length (loop
->superloops
);
777 return (*loop
->superloops
)[1];
780 extern void record_niter_bound (struct loop
*, const widest_int
&, bool, bool);
781 extern HOST_WIDE_INT
get_estimated_loop_iterations_int (struct loop
*);
782 extern HOST_WIDE_INT
get_max_loop_iterations_int (const struct loop
*);
783 extern HOST_WIDE_INT
get_likely_max_loop_iterations_int (struct loop
*);
784 extern bool get_estimated_loop_iterations (struct loop
*loop
, widest_int
*nit
);
785 extern bool get_max_loop_iterations (const struct loop
*loop
, widest_int
*nit
);
786 extern bool get_likely_max_loop_iterations (struct loop
*loop
, widest_int
*nit
);
787 extern int bb_loop_depth (const_basic_block
);
789 /* Converts VAL to widest_int. */
791 static inline widest_int
792 gcov_type_to_wide_int (gcov_type val
)
796 a
[0] = (unsigned HOST_WIDE_INT
) val
;
797 /* If HOST_BITS_PER_WIDE_INT == HOST_BITS_PER_WIDEST_INT, avoid shifting by
799 val
>>= HOST_BITS_PER_WIDE_INT
- 1;
801 a
[1] = (unsigned HOST_WIDE_INT
) val
;
803 return widest_int::from_array (a
, 2);
805 #endif /* GCC_CFGLOOP_H */