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. */
38 struct GTY (()) lpt_decision
{
39 enum lpt_dec decision
;
43 /* The type of extend applied to an IV. */
51 /* The structure describing a bound on number of iterations of a loop. */
53 struct GTY ((chain_next ("%h.next"))) nb_iter_bound
{
54 /* The statement STMT is executed at most ... */
57 /* ... BOUND + 1 times (BOUND must be an unsigned constant).
58 The + 1 is added for the following reasons:
60 a) 0 would otherwise be unused, while we would need to care more about
61 overflows (as MAX + 1 is sometimes produced as the estimate on number
62 of executions of STMT).
63 b) it is consistent with the result of number_of_iterations_exit. */
66 /* True if the statement will cause the loop to be leaved the (at most)
67 BOUND + 1-st time it is executed, that is, all the statements after it
68 are executed at most BOUND times. */
71 /* The next bound in the list. */
72 struct nb_iter_bound
*next
;
75 /* Description of the loop exit. */
77 struct GTY ((for_user
)) loop_exit
{
81 /* Previous and next exit in the list of the exits of the loop. */
82 struct loop_exit
*prev
;
83 struct loop_exit
*next
;
85 /* Next element in the list of loops from that E exits. */
86 struct loop_exit
*next_e
;
89 struct loop_exit_hasher
: ggc_hasher
<loop_exit
*>
91 typedef edge compare_type
;
93 static hashval_t
hash (loop_exit
*);
94 static bool equal (loop_exit
*, edge
);
95 static void remove (loop_exit
*);
98 typedef struct loop
*loop_p
;
100 /* An integer estimation of the number of iterations. Estimate_state
101 describes what is the state of the estimation. */
104 /* Estimate was not computed yet. */
106 /* Estimate is ready. */
111 /* Structure to hold information for each natural loop. */
112 struct GTY ((chain_next ("%h.next"))) loop
{
113 /* Index into loops array. */
116 /* Number of loop insns. */
119 /* Basic block of loop header. */
122 /* Basic block of loop latch. */
125 /* For loop unrolling/peeling decision. */
126 struct lpt_decision lpt_decision
;
128 /* Average number of executed insns per iteration. */
131 /* Number of blocks contained within the loop. */
134 /* Superloops of the loop, starting with the outermost loop. */
135 vec
<loop_p
, va_gc
> *superloops
;
137 /* The first inner (child) loop or NULL if innermost loop. */
140 /* Link to the next (sibling) loop. */
143 /* Auxiliary info specific to a pass. */
144 PTR
GTY ((skip (""))) aux
;
146 /* The number of times the latch of the loop is executed. This can be an
147 INTEGER_CST, or a symbolic expression representing the number of
148 iterations like "N - 1", or a COND_EXPR containing the runtime
149 conditions under which the number of iterations is non zero.
151 Don't access this field directly: number_of_latch_executions
152 computes and caches the computed information in this field. */
155 /* An integer guaranteed to be greater or equal to nb_iterations. Only
156 valid if any_upper_bound is true. */
157 widest_int nb_iterations_upper_bound
;
159 /* An integer giving an estimate on nb_iterations. Unlike
160 nb_iterations_upper_bound, there is no guarantee that it is at least
162 widest_int nb_iterations_estimate
;
164 bool any_upper_bound
;
167 /* True if the loop can be parallel. */
168 bool can_be_parallel
;
170 /* True if -Waggressive-loop-optimizations warned about this loop
172 bool warned_aggressive_loop_optimizations
;
174 /* An integer estimation of the number of iterations. Estimate_state
175 describes what is the state of the estimation. */
176 enum loop_estimation estimate_state
;
178 /* If > 0, an integer, where the user asserted that for any
179 I in [ 0, nb_iterations ) and for any J in
180 [ I, min ( I + safelen, nb_iterations ) ), the Ith and Jth iterations
181 of the loop can be safely evaluated concurrently. */
184 /* True if this loop should never be vectorized. */
187 /* True if we should try harder to vectorize this loop. */
188 bool force_vectorize
;
190 /* For SIMD loops, this is a unique identifier of the loop, referenced
191 by IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LANE and IFN_GOMP_SIMD_LAST_LANE
195 /* Upper bound on number of iterations of a loop. */
196 struct nb_iter_bound
*bounds
;
198 /* Head of the cyclic list of the exits of the loop. */
199 struct loop_exit
*exits
;
201 /* Number of iteration analysis data for RTL. */
202 struct niter_desc
*simple_loop_desc
;
204 /* For sanity checking during loop fixup we record here the former
205 loop header for loops marked for removal. Note that this prevents
206 the basic-block from being collected but its index can still be
208 basic_block former_header
;
211 /* Flags for state of loop structure. */
214 LOOPS_HAVE_PREHEADERS
= 1,
215 LOOPS_HAVE_SIMPLE_LATCHES
= 2,
216 LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
= 4,
217 LOOPS_HAVE_RECORDED_EXITS
= 8,
218 LOOPS_MAY_HAVE_MULTIPLE_LATCHES
= 16,
219 LOOP_CLOSED_SSA
= 32,
220 LOOPS_NEED_FIXUP
= 64,
221 LOOPS_HAVE_FALLTHRU_PREHEADERS
= 128
224 #define LOOPS_NORMAL (LOOPS_HAVE_PREHEADERS | LOOPS_HAVE_SIMPLE_LATCHES \
225 | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)
226 #define AVOID_CFG_MODIFICATIONS (LOOPS_MAY_HAVE_MULTIPLE_LATCHES)
228 /* Structure to hold CFG information about natural loops within a function. */
229 struct GTY (()) loops
{
230 /* State of loops. */
233 /* Array of the loops. */
234 vec
<loop_p
, va_gc
> *larray
;
236 /* Maps edges to the list of their descriptions as loop exits. Edges
237 whose sources or destinations have loop_father == NULL (which may
238 happen during the cfg manipulations) should not appear in EXITS. */
239 hash_table
<loop_exit_hasher
> *GTY(()) exits
;
241 /* Pointer to root of loop hierarchy tree. */
242 struct loop
*tree_root
;
245 /* Loop recognition. */
246 bool bb_loop_header_p (basic_block
);
247 void init_loops_structure (struct function
*, struct loops
*, unsigned);
248 extern struct loops
*flow_loops_find (struct loops
*);
249 extern void disambiguate_loops_with_multiple_latches (void);
250 extern void flow_loops_free (struct loops
*);
251 extern void flow_loops_dump (FILE *,
252 void (*)(const struct loop
*, FILE *, int), int);
253 extern void flow_loop_dump (const struct loop
*, FILE *,
254 void (*)(const struct loop
*, FILE *, int), int);
255 struct loop
*alloc_loop (void);
256 extern void flow_loop_free (struct loop
*);
257 int flow_loop_nodes_find (basic_block
, struct loop
*);
258 unsigned fix_loop_structure (bitmap changed_bbs
);
259 bool mark_irreducible_loops (void);
260 void release_recorded_exits (void);
261 void record_loop_exits (void);
262 void rescan_loop_exit (edge
, bool, bool);
264 /* Loop data structure manipulation/querying. */
265 extern void flow_loop_tree_node_add (struct loop
*, struct loop
*);
266 extern void flow_loop_tree_node_remove (struct loop
*);
267 extern void place_new_loop (struct function
*, struct loop
*);
268 extern void add_loop (struct loop
*, struct loop
*);
269 extern bool flow_loop_nested_p (const struct loop
*, const struct loop
*);
270 extern bool flow_bb_inside_loop_p (const struct loop
*, const_basic_block
);
271 extern struct loop
* find_common_loop (struct loop
*, struct loop
*);
272 struct loop
*superloop_at_depth (struct loop
*, unsigned);
273 struct eni_weights_d
;
274 extern int num_loop_insns (const struct loop
*);
275 extern int average_num_loop_insns (const struct loop
*);
276 extern unsigned get_loop_level (const struct loop
*);
277 extern bool loop_exit_edge_p (const struct loop
*, const_edge
);
278 extern bool loop_exits_to_bb_p (struct loop
*, basic_block
);
279 extern bool loop_exits_from_bb_p (struct loop
*, basic_block
);
280 extern void mark_loop_exit_edges (void);
281 extern location_t
get_loop_location (struct loop
*loop
);
283 /* Loops & cfg manipulation. */
284 extern basic_block
*get_loop_body (const struct loop
*);
285 extern unsigned get_loop_body_with_size (const struct loop
*, basic_block
*,
287 extern basic_block
*get_loop_body_in_dom_order (const struct loop
*);
288 extern basic_block
*get_loop_body_in_bfs_order (const struct loop
*);
289 extern basic_block
*get_loop_body_in_custom_order (const struct loop
*,
290 int (*) (const void *, const void *));
292 extern vec
<edge
> get_loop_exit_edges (const struct loop
*);
293 extern edge
single_exit (const struct loop
*);
294 extern edge
single_likely_exit (struct loop
*loop
);
295 extern unsigned num_loop_branches (const struct loop
*);
297 extern edge
loop_preheader_edge (const struct loop
*);
298 extern edge
loop_latch_edge (const struct loop
*);
300 extern void add_bb_to_loop (basic_block
, struct loop
*);
301 extern void remove_bb_from_loops (basic_block
);
303 extern void cancel_loop_tree (struct loop
*);
304 extern void delete_loop (struct loop
*);
308 CP_SIMPLE_PREHEADERS
= 1,
309 CP_FALLTHRU_PREHEADERS
= 2
312 basic_block
create_preheader (struct loop
*, int);
313 extern void create_preheaders (int);
314 extern void force_single_succ_latches (void);
316 extern void verify_loop_structure (void);
319 extern bool just_once_each_iteration_p (const struct loop
*, const_basic_block
);
320 gcov_type
expected_loop_iterations_unbounded (const struct loop
*);
321 extern unsigned expected_loop_iterations (const struct loop
*);
322 extern rtx
doloop_condition_get (rtx
);
325 /* Loop manipulation. */
326 extern bool can_duplicate_loop_p (const struct loop
*loop
);
328 #define DLTHE_FLAG_UPDATE_FREQ 1 /* Update frequencies in
329 duplicate_loop_to_header_edge. */
330 #define DLTHE_RECORD_COPY_NUMBER 2 /* Record copy number in the aux
331 field of newly create BB. */
332 #define DLTHE_FLAG_COMPLETTE_PEEL 4 /* Update frequencies expecting
333 a complete peeling. */
335 extern edge
create_empty_if_region_on_edge (edge
, tree
);
336 extern struct loop
*create_empty_loop_on_edge (edge
, tree
, tree
, tree
, tree
,
337 tree
*, tree
*, struct loop
*);
338 extern struct loop
* duplicate_loop (struct loop
*, struct loop
*);
339 extern void copy_loop_info (struct loop
*loop
, struct loop
*target
);
340 extern void duplicate_subloops (struct loop
*, struct loop
*);
341 extern bool duplicate_loop_to_header_edge (struct loop
*, edge
,
342 unsigned, sbitmap
, edge
,
344 extern struct loop
*loopify (edge
, edge
,
345 basic_block
, edge
, edge
, bool,
347 struct loop
* loop_version (struct loop
*, void *,
348 basic_block
*, unsigned, unsigned, unsigned, bool);
349 extern bool remove_path (edge
);
350 extern void unloop (struct loop
*, bool *, bitmap
);
351 extern void scale_loop_frequencies (struct loop
*, int, int);
352 void mark_loop_for_removal (loop_p
);
355 /* Induction variable analysis. */
357 /* The description of induction variable. The things are a bit complicated
358 due to need to handle subregs and extends. The value of the object described
359 by it can be obtained as follows (all computations are done in extend_mode):
361 Value in i-th iteration is
362 delta + mult * extend_{extend_mode} (subreg_{mode} (base + i * step)).
364 If first_special is true, the value in the first iteration is
367 If extend = UNKNOWN, first_special must be false, delta 0, mult 1 and value is
368 subreg_{mode} (base + i * step)
370 The get_iv_value function can be used to obtain these expressions.
372 ??? Add a third mode field that would specify the mode in that inner
373 computation is done, which would enable it to be different from the
378 /* Its base and step (mode of base and step is supposed to be extend_mode,
379 see the description above). */
382 /* The type of extend applied to it (IV_SIGN_EXTEND, IV_ZERO_EXTEND,
383 or IV_UNKNOWN_EXTEND). */
384 enum iv_extend_code extend
;
386 /* Operations applied in the extended mode. */
389 /* The mode it is extended to. */
390 enum machine_mode extend_mode
;
392 /* The mode the variable iterates in. */
393 enum machine_mode mode
;
395 /* Whether the first iteration needs to be handled specially. */
396 unsigned first_special
: 1;
399 /* The description of an exit from the loop and of the number of iterations
400 till we take the exit. */
402 struct GTY(()) niter_desc
404 /* The edge out of the loop. */
407 /* The other edge leading from the condition. */
410 /* True if we are able to say anything about number of iterations of the
414 /* True if the loop iterates the constant number of times. */
417 /* Number of iterations if constant. */
420 /* Assumptions under that the rest of the information is valid. */
423 /* Assumptions under that the loop ends before reaching the latch,
424 even if value of niter_expr says otherwise. */
425 rtx noloop_assumptions
;
427 /* Condition under that the loop is infinite. */
430 /* Whether the comparison is signed. */
433 /* The mode in that niter_expr should be computed. */
434 enum machine_mode mode
;
436 /* The number of iterations of the loop. */
440 extern void iv_analysis_loop_init (struct loop
*);
441 extern bool iv_analyze (rtx_insn
*, rtx
, struct rtx_iv
*);
442 extern bool iv_analyze_result (rtx_insn
*, rtx
, struct rtx_iv
*);
443 extern bool iv_analyze_expr (rtx_insn
*, rtx
, enum machine_mode
,
445 extern rtx
get_iv_value (struct rtx_iv
*, rtx
);
446 extern bool biv_p (rtx_insn
*, rtx
);
447 extern void find_simple_exit (struct loop
*, struct niter_desc
*);
448 extern void iv_analysis_done (void);
450 extern struct niter_desc
*get_simple_loop_desc (struct loop
*loop
);
451 extern void free_simple_loop_desc (struct loop
*loop
);
453 static inline struct niter_desc
*
454 simple_loop_desc (struct loop
*loop
)
456 return loop
->simple_loop_desc
;
459 /* Accessors for the loop structures. */
461 /* Returns the loop with index NUM from FNs loop tree. */
463 static inline struct loop
*
464 get_loop (struct function
*fn
, unsigned num
)
466 return (*loops_for_fn (fn
)->larray
)[num
];
469 /* Returns the number of superloops of LOOP. */
471 static inline unsigned
472 loop_depth (const struct loop
*loop
)
474 return vec_safe_length (loop
->superloops
);
477 /* Returns the immediate superloop of LOOP, or NULL if LOOP is the outermost
480 static inline struct loop
*
481 loop_outer (const struct loop
*loop
)
483 unsigned n
= vec_safe_length (loop
->superloops
);
488 return (*loop
->superloops
)[n
- 1];
491 /* Returns true if LOOP has at least one exit edge. */
494 loop_has_exit_edges (const struct loop
*loop
)
496 return loop
->exits
->next
->e
!= NULL
;
499 /* Returns the list of loops in FN. */
501 inline vec
<loop_p
, va_gc
> *
502 get_loops (struct function
*fn
)
504 struct loops
*loops
= loops_for_fn (fn
);
508 return loops
->larray
;
511 /* Returns the number of loops in FN (including the removed
512 ones and the fake loop that forms the root of the loop tree). */
514 static inline unsigned
515 number_of_loops (struct function
*fn
)
517 struct loops
*loops
= loops_for_fn (fn
);
521 return vec_safe_length (loops
->larray
);
524 /* Returns true if state of the loops satisfies all properties
525 described by FLAGS. */
528 loops_state_satisfies_p (unsigned flags
)
530 return (current_loops
->state
& flags
) == flags
;
533 /* Sets FLAGS to the loops state. */
536 loops_state_set (unsigned flags
)
538 current_loops
->state
|= flags
;
541 /* Clears FLAGS from the loops state. */
544 loops_state_clear (unsigned flags
)
548 current_loops
->state
&= ~flags
;
551 /* Loop iterators. */
553 /* Flags for loop iteration. */
557 LI_INCLUDE_ROOT
= 1, /* Include the fake root of the loop tree. */
558 LI_FROM_INNERMOST
= 2, /* Iterate over the loops in the reverse order,
559 starting from innermost ones. */
560 LI_ONLY_INNERMOST
= 4 /* Iterate only over innermost loops. */
563 /* The iterator for loops. */
567 loop_iterator (loop_p
*loop
, unsigned flags
);
570 inline loop_p
next ();
572 /* The list of loops to visit. */
575 /* The index of the actual loop. */
580 loop_iterator::next ()
584 while (this->to_visit
.iterate (this->idx
, &anum
))
587 loop_p loop
= get_loop (cfun
, anum
);
596 loop_iterator::loop_iterator (loop_p
*loop
, unsigned flags
)
605 this->to_visit
.create (0);
610 this->to_visit
.create (number_of_loops (cfun
));
611 mn
= (flags
& LI_INCLUDE_ROOT
) ? 0 : 1;
613 if (flags
& LI_ONLY_INNERMOST
)
615 for (i
= 0; vec_safe_iterate (current_loops
->larray
, i
, &aloop
); i
++)
617 && aloop
->inner
== NULL
619 this->to_visit
.quick_push (aloop
->num
);
621 else if (flags
& LI_FROM_INNERMOST
)
623 /* Push the loops to LI->TO_VISIT in postorder. */
624 for (aloop
= current_loops
->tree_root
;
625 aloop
->inner
!= NULL
;
626 aloop
= aloop
->inner
)
631 if (aloop
->num
>= mn
)
632 this->to_visit
.quick_push (aloop
->num
);
636 for (aloop
= aloop
->next
;
637 aloop
->inner
!= NULL
;
638 aloop
= aloop
->inner
)
641 else if (!loop_outer (aloop
))
644 aloop
= loop_outer (aloop
);
649 /* Push the loops to LI->TO_VISIT in preorder. */
650 aloop
= current_loops
->tree_root
;
653 if (aloop
->num
>= mn
)
654 this->to_visit
.quick_push (aloop
->num
);
656 if (aloop
->inner
!= NULL
)
657 aloop
= aloop
->inner
;
660 while (aloop
!= NULL
&& aloop
->next
== NULL
)
661 aloop
= loop_outer (aloop
);
669 *loop
= this->next ();
673 loop_iterator::~loop_iterator ()
675 this->to_visit
.release ();
678 #define FOR_EACH_LOOP(LOOP, FLAGS) \
679 for (loop_iterator li(&(LOOP), FLAGS); \
683 /* The properties of the target. */
684 struct target_cfgloop
{
685 /* Number of available registers. */
686 unsigned x_target_avail_regs
;
688 /* Number of available registers that are call-clobbered. */
689 unsigned x_target_clobbered_regs
;
691 /* Number of registers reserved for temporary expressions. */
692 unsigned x_target_res_regs
;
694 /* The cost for register when there still is some reserve, but we are
695 approaching the number of available registers. */
696 unsigned x_target_reg_cost
[2];
698 /* The cost for register when we need to spill. */
699 unsigned x_target_spill_cost
[2];
702 extern struct target_cfgloop default_target_cfgloop
;
703 #if SWITCHABLE_TARGET
704 extern struct target_cfgloop
*this_target_cfgloop
;
706 #define this_target_cfgloop (&default_target_cfgloop)
709 #define target_avail_regs \
710 (this_target_cfgloop->x_target_avail_regs)
711 #define target_clobbered_regs \
712 (this_target_cfgloop->x_target_clobbered_regs)
713 #define target_res_regs \
714 (this_target_cfgloop->x_target_res_regs)
715 #define target_reg_cost \
716 (this_target_cfgloop->x_target_reg_cost)
717 #define target_spill_cost \
718 (this_target_cfgloop->x_target_spill_cost)
720 /* Register pressure estimation for induction variable optimizations & loop
722 extern unsigned estimate_reg_pressure_cost (unsigned, unsigned, bool, bool);
723 extern void init_set_costs (void);
725 /* Loop optimizer initialization. */
726 extern void loop_optimizer_init (unsigned);
727 extern void loop_optimizer_finalize (void);
729 /* Optimization passes. */
732 UAP_UNROLL
= 1, /* Enables unrolling of loops if it seems profitable. */
733 UAP_UNROLL_ALL
= 2 /* Enables unrolling of all loops. */
736 extern void unroll_loops (int);
737 extern void doloop_optimize_loops (void);
738 extern void move_loop_invariants (void);
739 extern void scale_loop_profile (struct loop
*loop
, int scale
, gcov_type iteration_bound
);
740 extern vec
<basic_block
> get_loop_hot_path (const struct loop
*loop
);
742 /* Returns the outermost loop of the loop nest that contains LOOP.*/
743 static inline struct loop
*
744 loop_outermost (struct loop
*loop
)
746 unsigned n
= vec_safe_length (loop
->superloops
);
751 return (*loop
->superloops
)[1];
754 extern void record_niter_bound (struct loop
*, const widest_int
&, bool, bool);
755 extern HOST_WIDE_INT
get_estimated_loop_iterations_int (struct loop
*);
756 extern HOST_WIDE_INT
get_max_loop_iterations_int (struct loop
*);
757 extern bool get_estimated_loop_iterations (struct loop
*loop
, widest_int
*nit
);
758 extern bool get_max_loop_iterations (struct loop
*loop
, widest_int
*nit
);
759 extern int bb_loop_depth (const_basic_block
);
761 /* Converts VAL to widest_int. */
763 static inline widest_int
764 gcov_type_to_wide_int (gcov_type val
)
768 a
[0] = (unsigned HOST_WIDE_INT
) val
;
769 /* If HOST_BITS_PER_WIDE_INT == HOST_BITS_PER_WIDEST_INT, avoid shifting by
771 val
>>= HOST_BITS_PER_WIDE_INT
- 1;
773 a
[1] = (unsigned HOST_WIDE_INT
) val
;
775 return widest_int::from_array (a
, 2);
777 #endif /* GCC_CFGLOOP_H */