]>
Commit | Line | Data |
---|---|---|
3d436d2a | 1 | /* Natural loop functions |
a5544970 | 2 | Copyright (C) 1987-2019 Free Software Foundation, Inc. |
3d436d2a ZD |
3 | |
4 | This file is part of GCC. | |
5 | ||
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 | |
9dcd6f09 | 8 | Software Foundation; either version 3, or (at your option) any later |
3d436d2a ZD |
9 | version. |
10 | ||
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 | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
3d436d2a | 19 | |
59587b18 JQ |
20 | #ifndef GCC_CFGLOOP_H |
21 | #define GCC_CFGLOOP_H | |
22 | ||
893479de | 23 | #include "cfgloopmanip.h" |
7a8cba34 | 24 | |
3d436d2a ZD |
25 | /* Structure to hold decision about unrolling/peeling. */ |
26 | enum lpt_dec | |
27 | { | |
28 | LPT_NONE, | |
3d436d2a ZD |
29 | LPT_UNROLL_CONSTANT, |
30 | LPT_UNROLL_RUNTIME, | |
31 | LPT_UNROLL_STUPID | |
32 | }; | |
33 | ||
d1b38208 | 34 | struct GTY (()) lpt_decision { |
3d436d2a ZD |
35 | enum lpt_dec decision; |
36 | unsigned times; | |
37 | }; | |
38 | ||
1c1ad7bb SB |
39 | /* The type of extend applied to an IV. */ |
40 | enum iv_extend_code | |
41 | { | |
42 | IV_SIGN_EXTEND, | |
43 | IV_ZERO_EXTEND, | |
44 | IV_UNKNOWN_EXTEND | |
45 | }; | |
46 | ||
86df10e3 SP |
47 | /* The structure describing a bound on number of iterations of a loop. */ |
48 | ||
d1b38208 | 49 | struct GTY ((chain_next ("%h.next"))) nb_iter_bound { |
946e1bc7 | 50 | /* The statement STMT is executed at most ... */ |
355fe088 | 51 | gimple *stmt; |
946e1bc7 ZD |
52 | |
53 | /* ... BOUND + 1 times (BOUND must be an unsigned constant). | |
54 | The + 1 is added for the following reasons: | |
55 | ||
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. */ | |
807e902e | 60 | widest_int bound; |
946e1bc7 | 61 | |
b8698a0f | 62 | /* True if the statement will cause the loop to be leaved the (at most) |
946e1bc7 ZD |
63 | BOUND + 1-st time it is executed, that is, all the statements after it |
64 | are executed at most BOUND times. */ | |
65 | bool is_exit; | |
66 | ||
946e1bc7 | 67 | /* The next bound in the list. */ |
86df10e3 | 68 | struct nb_iter_bound *next; |
86df10e3 SP |
69 | }; |
70 | ||
6270df4c ZD |
71 | /* Description of the loop exit. */ |
72 | ||
2a22f99c | 73 | struct GTY ((for_user)) loop_exit { |
6270df4c | 74 | /* The exit edge. */ |
b8244d74 | 75 | edge e; |
6270df4c ZD |
76 | |
77 | /* Previous and next exit in the list of the exits of the loop. */ | |
78 | struct loop_exit *prev; | |
79 | struct loop_exit *next; | |
80 | ||
81 | /* Next element in the list of loops from that E exits. */ | |
82 | struct loop_exit *next_e; | |
83 | }; | |
84 | ||
ca752f39 | 85 | struct loop_exit_hasher : ggc_ptr_hash<loop_exit> |
2a22f99c TS |
86 | { |
87 | typedef edge compare_type; | |
88 | ||
89 | static hashval_t hash (loop_exit *); | |
90 | static bool equal (loop_exit *, edge); | |
91 | static void remove (loop_exit *); | |
92 | }; | |
93 | ||
9ba025a2 | 94 | typedef struct loop *loop_p; |
9ba025a2 | 95 | |
ae50c0cb TN |
96 | /* An integer estimation of the number of iterations. Estimate_state |
97 | describes what is the state of the estimation. */ | |
98 | enum loop_estimation | |
99 | { | |
100 | /* Estimate was not computed yet. */ | |
101 | EST_NOT_COMPUTED, | |
102 | /* Estimate is ready. */ | |
dd366ec3 RB |
103 | EST_AVAILABLE, |
104 | EST_LAST | |
ae50c0cb TN |
105 | }; |
106 | ||
2f07b722 BC |
107 | /* The structure describing non-overflow control induction variable for |
108 | loop's exit edge. */ | |
109 | struct GTY ((chain_next ("%h.next"))) control_iv { | |
110 | tree base; | |
111 | tree step; | |
112 | struct control_iv *next; | |
113 | }; | |
114 | ||
3d436d2a | 115 | /* Structure to hold information for each natural loop. */ |
d1b38208 | 116 | struct GTY ((chain_next ("%h.next"))) loop { |
542e7230 BC |
117 | /* Index into loops array. Note indices will never be reused after loop |
118 | is destroyed. */ | |
3d436d2a ZD |
119 | int num; |
120 | ||
8f5929e1 JJ |
121 | /* Number of loop insns. */ |
122 | unsigned ninsns; | |
123 | ||
3d436d2a | 124 | /* Basic block of loop header. */ |
b8244d74 | 125 | basic_block header; |
3d436d2a ZD |
126 | |
127 | /* Basic block of loop latch. */ | |
b8244d74 | 128 | basic_block latch; |
3d436d2a | 129 | |
3d436d2a ZD |
130 | /* For loop unrolling/peeling decision. */ |
131 | struct lpt_decision lpt_decision; | |
132 | ||
3d436d2a ZD |
133 | /* Average number of executed insns per iteration. */ |
134 | unsigned av_ninsns; | |
135 | ||
3d436d2a ZD |
136 | /* Number of blocks contained within the loop. */ |
137 | unsigned num_nodes; | |
138 | ||
9ba025a2 | 139 | /* Superloops of the loop, starting with the outermost loop. */ |
9771b263 | 140 | vec<loop_p, va_gc> *superloops; |
3d436d2a ZD |
141 | |
142 | /* The first inner (child) loop or NULL if innermost loop. */ | |
143 | struct loop *inner; | |
144 | ||
145 | /* Link to the next (sibling) loop. */ | |
146 | struct loop *next; | |
147 | ||
3d436d2a | 148 | /* Auxiliary info specific to a pass. */ |
9e2f83a5 | 149 | PTR GTY ((skip (""))) aux; |
3d436d2a | 150 | |
0a74c758 SP |
151 | /* The number of times the latch of the loop is executed. This can be an |
152 | INTEGER_CST, or a symbolic expression representing the number of | |
153 | iterations like "N - 1", or a COND_EXPR containing the runtime | |
154 | conditions under which the number of iterations is non zero. | |
155 | ||
156 | Don't access this field directly: number_of_latch_executions | |
157 | computes and caches the computed information in this field. */ | |
9baba81b SP |
158 | tree nb_iterations; |
159 | ||
b4a9343c ZD |
160 | /* An integer guaranteed to be greater or equal to nb_iterations. Only |
161 | valid if any_upper_bound is true. */ | |
807e902e | 162 | widest_int nb_iterations_upper_bound; |
9bdb685e | 163 | |
105e29c5 JH |
164 | widest_int nb_iterations_likely_upper_bound; |
165 | ||
b4a9343c ZD |
166 | /* An integer giving an estimate on nb_iterations. Unlike |
167 | nb_iterations_upper_bound, there is no guarantee that it is at least | |
168 | nb_iterations. */ | |
807e902e | 169 | widest_int nb_iterations_estimate; |
86df10e3 | 170 | |
74bf76ed JJ |
171 | /* If > 0, an integer, where the user asserted that for any |
172 | I in [ 0, nb_iterations ) and for any J in | |
173 | [ I, min ( I + safelen, nb_iterations ) ), the Ith and Jth iterations | |
174 | of the loop can be safely evaluated concurrently. */ | |
175 | int safelen; | |
176 | ||
18767ebc BC |
177 | /* Constraints are generally set by consumers and affect certain |
178 | semantics of niter analyzer APIs. Currently the APIs affected are | |
179 | number_of_iterations_exit* functions and their callers. One typical | |
180 | use case of constraints is to vectorize possibly infinite loop: | |
181 | ||
182 | 1) Compute niter->assumptions by calling niter analyzer API and | |
183 | record it as possible condition for loop versioning. | |
184 | 2) Clear buffered result of niter/scev analyzer. | |
185 | 3) Set constraint LOOP_C_FINITE assuming the loop is finite. | |
186 | 4) Analyze data references. Since data reference analysis depends | |
187 | on niter/scev analyzer, the point is that niter/scev analysis | |
188 | is done under circumstance of LOOP_C_FINITE constraint. | |
189 | 5) Version the loop with niter->assumptions computed in step 1). | |
190 | 6) Vectorize the versioned loop in which niter->assumptions is | |
191 | checked to be true. | |
192 | 7) Update constraints in versioned loops so that niter analyzer | |
193 | in following passes can use it. | |
194 | ||
195 | Note consumers are usually the loop optimizers and it is consumers' | |
196 | responsibility to set/clear constraints correctly. Failing to do | |
197 | that might result in hard to track down bugs in niter/scev consumers. */ | |
198 | unsigned constraints; | |
199 | ||
34e82342 RB |
200 | /* An integer estimation of the number of iterations. Estimate_state |
201 | describes what is the state of the estimation. */ | |
202 | ENUM_BITFIELD(loop_estimation) estimate_state : 8; | |
203 | ||
204 | unsigned any_upper_bound : 1; | |
205 | unsigned any_estimate : 1; | |
206 | unsigned any_likely_upper_bound : 1; | |
207 | ||
208 | /* True if the loop can be parallel. */ | |
209 | unsigned can_be_parallel : 1; | |
210 | ||
211 | /* True if -Waggressive-loop-optimizations warned about this loop | |
212 | already. */ | |
213 | unsigned warned_aggressive_loop_optimizations : 1; | |
214 | ||
5ce9450f | 215 | /* True if this loop should never be vectorized. */ |
34e82342 | 216 | unsigned dont_vectorize : 1; |
5ce9450f | 217 | |
718c4601 | 218 | /* True if we should try harder to vectorize this loop. */ |
34e82342 | 219 | unsigned force_vectorize : 1; |
718c4601 | 220 | |
886c388d | 221 | /* True if the loop is part of an oacc kernels region. */ |
34e82342 | 222 | unsigned in_oacc_kernels_region : 1; |
886c388d | 223 | |
ff27462e EB |
224 | /* The number of times to unroll the loop. 0 means no information given, |
225 | just do what we always do. A value of 1 means do not unroll the loop. | |
226 | A value of USHRT_MAX means unroll with no specific unrolling factor. | |
227 | Other values means unroll with the given unrolling factor. */ | |
ac9effed EB |
228 | unsigned short unroll; |
229 | ||
74bf76ed JJ |
230 | /* For SIMD loops, this is a unique identifier of the loop, referenced |
231 | by IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LANE and IFN_GOMP_SIMD_LAST_LANE | |
232 | builtins. */ | |
233 | tree simduid; | |
234 | ||
542e7230 BC |
235 | /* In loop optimization, it's common to generate loops from the original |
236 | loop. This field records the index of the original loop which can be | |
237 | used to track the original loop from newly generated loops. This can | |
238 | be done by calling function get_loop (cfun, orig_loop_num). Note the | |
239 | original loop could be destroyed for various reasons thus no longer | |
240 | exists, as a result, function call to get_loop returns NULL pointer. | |
241 | In this case, this field should not be used and needs to be cleared | |
242 | whenever possible. */ | |
243 | int orig_loop_num; | |
244 | ||
e9eb809d ZD |
245 | /* Upper bound on number of iterations of a loop. */ |
246 | struct nb_iter_bound *bounds; | |
82b85a85 | 247 | |
2f07b722 BC |
248 | /* Non-overflow control ivs of a loop. */ |
249 | struct control_iv *control_ivs; | |
250 | ||
6270df4c | 251 | /* Head of the cyclic list of the exits of the loop. */ |
9e2f83a5 | 252 | struct loop_exit *exits; |
ef23e6a2 RB |
253 | |
254 | /* Number of iteration analysis data for RTL. */ | |
255 | struct niter_desc *simple_loop_desc; | |
e4ca2139 | 256 | |
e4ca2139 RB |
257 | /* For sanity checking during loop fixup we record here the former |
258 | loop header for loops marked for removal. Note that this prevents | |
259 | the basic-block from being collected but its index can still be | |
260 | reused. */ | |
261 | basic_block former_header; | |
3d436d2a ZD |
262 | }; |
263 | ||
18767ebc BC |
264 | /* Set if the loop is known to be infinite. */ |
265 | #define LOOP_C_INFINITE (1 << 0) | |
266 | /* Set if the loop is known to be finite without any assumptions. */ | |
267 | #define LOOP_C_FINITE (1 << 1) | |
268 | ||
269 | /* Set C to the LOOP constraint. */ | |
270 | static inline void | |
271 | loop_constraint_set (struct loop *loop, unsigned c) | |
272 | { | |
273 | loop->constraints |= c; | |
274 | } | |
275 | ||
276 | /* Clear C from the LOOP constraint. */ | |
277 | static inline void | |
278 | loop_constraint_clear (struct loop *loop, unsigned c) | |
279 | { | |
280 | loop->constraints &= ~c; | |
281 | } | |
282 | ||
283 | /* Check if C is set in the LOOP constraint. */ | |
284 | static inline bool | |
285 | loop_constraint_set_p (struct loop *loop, unsigned c) | |
286 | { | |
287 | return (loop->constraints & c) == c; | |
288 | } | |
289 | ||
3d436d2a ZD |
290 | /* Flags for state of loop structure. */ |
291 | enum | |
292 | { | |
293 | LOOPS_HAVE_PREHEADERS = 1, | |
294 | LOOPS_HAVE_SIMPLE_LATCHES = 2, | |
82b85a85 | 295 | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS = 4, |
89f8f30f | 296 | LOOPS_HAVE_RECORDED_EXITS = 8, |
c7b852c8 | 297 | LOOPS_MAY_HAVE_MULTIPLE_LATCHES = 16, |
592c303d | 298 | LOOP_CLOSED_SSA = 32, |
e855c69d AB |
299 | LOOPS_NEED_FIXUP = 64, |
300 | LOOPS_HAVE_FALLTHRU_PREHEADERS = 128 | |
3d436d2a ZD |
301 | }; |
302 | ||
d78f3f78 ZD |
303 | #define LOOPS_NORMAL (LOOPS_HAVE_PREHEADERS | LOOPS_HAVE_SIMPLE_LATCHES \ |
304 | | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS) | |
89f8f30f | 305 | #define AVOID_CFG_MODIFICATIONS (LOOPS_MAY_HAVE_MULTIPLE_LATCHES) |
d78f3f78 | 306 | |
3d436d2a | 307 | /* Structure to hold CFG information about natural loops within a function. */ |
d1b38208 | 308 | struct GTY (()) loops { |
0162f155 RG |
309 | /* State of loops. */ |
310 | int state; | |
3d436d2a | 311 | |
42fd6772 | 312 | /* Array of the loops. */ |
9771b263 | 313 | vec<loop_p, va_gc> *larray; |
3d436d2a | 314 | |
6270df4c ZD |
315 | /* Maps edges to the list of their descriptions as loop exits. Edges |
316 | whose sources or destinations have loop_father == NULL (which may | |
317 | happen during the cfg manipulations) should not appear in EXITS. */ | |
2a22f99c | 318 | hash_table<loop_exit_hasher> *GTY(()) exits; |
6270df4c | 319 | |
3d436d2a ZD |
320 | /* Pointer to root of loop hierarchy tree. */ |
321 | struct loop *tree_root; | |
3d436d2a ZD |
322 | }; |
323 | ||
3d436d2a | 324 | /* Loop recognition. */ |
0375167b | 325 | bool bb_loop_header_p (basic_block); |
dd366ec3 | 326 | void init_loops_structure (struct function *, struct loops *, unsigned); |
0375167b | 327 | extern struct loops *flow_loops_find (struct loops *); |
89f8f30f | 328 | extern void disambiguate_loops_with_multiple_latches (void); |
d329e058 | 329 | extern void flow_loops_free (struct loops *); |
d73be268 | 330 | extern void flow_loops_dump (FILE *, |
d329e058 AJ |
331 | void (*)(const struct loop *, FILE *, int), int); |
332 | extern void flow_loop_dump (const struct loop *, FILE *, | |
333 | void (*)(const struct loop *, FILE *, int), int); | |
6270df4c | 334 | struct loop *alloc_loop (void); |
d329e058 | 335 | extern void flow_loop_free (struct loop *); |
2b271002 | 336 | int flow_loop_nodes_find (basic_block, struct loop *); |
8e89b5b5 | 337 | unsigned fix_loop_structure (bitmap changed_bbs); |
2de58650 | 338 | bool mark_irreducible_loops (void); |
61183076 | 339 | void release_recorded_exits (function *); |
6270df4c ZD |
340 | void record_loop_exits (void); |
341 | void rescan_loop_exit (edge, bool, bool); | |
26993e95 | 342 | void sort_sibling_loops (function *); |
3d436d2a | 343 | |
4d6922ee | 344 | /* Loop data structure manipulation/querying. */ |
1cc521f1 MM |
345 | extern void flow_loop_tree_node_add (struct loop *, struct loop *, |
346 | struct loop * = NULL); | |
d329e058 | 347 | extern void flow_loop_tree_node_remove (struct loop *); |
d329e058 | 348 | extern bool flow_loop_nested_p (const struct loop *, const struct loop *); |
ed7a4b4b | 349 | extern bool flow_bb_inside_loop_p (const struct loop *, const_basic_block); |
d329e058 | 350 | extern struct loop * find_common_loop (struct loop *, struct loop *); |
a7e5372d | 351 | struct loop *superloop_at_depth (struct loop *, unsigned); |
a79683d5 | 352 | struct eni_weights; |
ed7a4b4b KG |
353 | extern int num_loop_insns (const struct loop *); |
354 | extern int average_num_loop_insns (const struct loop *); | |
689ba89d | 355 | extern unsigned get_loop_level (const struct loop *); |
ed7a4b4b | 356 | extern bool loop_exit_edge_p (const struct loop *, const_edge); |
f4ce375d VK |
357 | extern bool loop_exits_to_bb_p (struct loop *, basic_block); |
358 | extern bool loop_exits_from_bb_p (struct loop *, basic_block); | |
d73be268 | 359 | extern void mark_loop_exit_edges (void); |
4f5b9c80 | 360 | extern dump_user_location_t get_loop_location (struct loop *loop); |
3d436d2a ZD |
361 | |
362 | /* Loops & cfg manipulation. */ | |
d329e058 | 363 | extern basic_block *get_loop_body (const struct loop *); |
89f8f30f ZD |
364 | extern unsigned get_loop_body_with_size (const struct loop *, basic_block *, |
365 | unsigned); | |
50654f6c | 366 | extern basic_block *get_loop_body_in_dom_order (const struct loop *); |
40923b20 | 367 | extern basic_block *get_loop_body_in_bfs_order (const struct loop *); |
b8698a0f | 368 | extern basic_block *get_loop_body_in_custom_order (const struct loop *, |
e855c69d AB |
369 | int (*) (const void *, const void *)); |
370 | ||
9771b263 | 371 | extern vec<edge> get_loop_exit_edges (const struct loop *); |
6d96bad2 JH |
372 | extern edge single_exit (const struct loop *); |
373 | extern edge single_likely_exit (struct loop *loop); | |
50654f6c | 374 | extern unsigned num_loop_branches (const struct loop *); |
3d436d2a | 375 | |
d329e058 AJ |
376 | extern edge loop_preheader_edge (const struct loop *); |
377 | extern edge loop_latch_edge (const struct loop *); | |
3d436d2a | 378 | |
d329e058 AJ |
379 | extern void add_bb_to_loop (basic_block, struct loop *); |
380 | extern void remove_bb_from_loops (basic_block); | |
3d436d2a | 381 | |
d73be268 | 382 | extern void cancel_loop_tree (struct loop *); |
42fd6772 | 383 | extern void delete_loop (struct loop *); |
3d436d2a | 384 | |
3d436d2a | 385 | |
d73be268 | 386 | extern void verify_loop_structure (void); |
3d436d2a ZD |
387 | |
388 | /* Loop analysis. */ | |
ed7a4b4b | 389 | extern bool just_once_each_iteration_p (const struct loop *, const_basic_block); |
199b1891 | 390 | gcov_type expected_loop_iterations_unbounded (const struct loop *, |
65739a68 | 391 | bool *read_profile_p = NULL, bool by_profile_only = false); |
97c53806 | 392 | extern unsigned expected_loop_iterations (struct loop *); |
30d2ef86 | 393 | extern rtx doloop_condition_get (rtx_insn *); |
617b465c | 394 | |
08c13199 RB |
395 | void mark_loop_for_removal (loop_p); |
396 | ||
50654f6c ZD |
397 | /* Induction variable analysis. */ |
398 | ||
399 | /* The description of induction variable. The things are a bit complicated | |
400 | due to need to handle subregs and extends. The value of the object described | |
401 | by it can be obtained as follows (all computations are done in extend_mode): | |
402 | ||
403 | Value in i-th iteration is | |
404 | delta + mult * extend_{extend_mode} (subreg_{mode} (base + i * step)). | |
405 | ||
406 | If first_special is true, the value in the first iteration is | |
407 | delta + mult * base | |
c22cacf3 | 408 | |
f822d252 | 409 | If extend = UNKNOWN, first_special must be false, delta 0, mult 1 and value is |
50654f6c ZD |
410 | subreg_{mode} (base + i * step) |
411 | ||
412 | The get_iv_value function can be used to obtain these expressions. | |
413 | ||
414 | ??? Add a third mode field that would specify the mode in that inner | |
415 | computation is done, which would enable it to be different from the | |
416 | outer one? */ | |
417 | ||
418 | struct rtx_iv | |
419 | { | |
420 | /* Its base and step (mode of base and step is supposed to be extend_mode, | |
421 | see the description above). */ | |
422 | rtx base, step; | |
423 | ||
1c1ad7bb SB |
424 | /* The type of extend applied to it (IV_SIGN_EXTEND, IV_ZERO_EXTEND, |
425 | or IV_UNKNOWN_EXTEND). */ | |
426 | enum iv_extend_code extend; | |
50654f6c ZD |
427 | |
428 | /* Operations applied in the extended mode. */ | |
429 | rtx delta, mult; | |
430 | ||
431 | /* The mode it is extended to. */ | |
3d88d1cd | 432 | scalar_int_mode extend_mode; |
50654f6c ZD |
433 | |
434 | /* The mode the variable iterates in. */ | |
3d88d1cd | 435 | scalar_int_mode mode; |
50654f6c | 436 | |
50654f6c ZD |
437 | /* Whether the first iteration needs to be handled specially. */ |
438 | unsigned first_special : 1; | |
439 | }; | |
440 | ||
f2dca510 ZD |
441 | /* The description of an exit from the loop and of the number of iterations |
442 | till we take the exit. */ | |
50654f6c | 443 | |
ef23e6a2 | 444 | struct GTY(()) niter_desc |
50654f6c ZD |
445 | { |
446 | /* The edge out of the loop. */ | |
447 | edge out_edge; | |
448 | ||
449 | /* The other edge leading from the condition. */ | |
450 | edge in_edge; | |
451 | ||
452 | /* True if we are able to say anything about number of iterations of the | |
453 | loop. */ | |
454 | bool simple_p; | |
455 | ||
456 | /* True if the loop iterates the constant number of times. */ | |
457 | bool const_iter; | |
458 | ||
459 | /* Number of iterations if constant. */ | |
a9243bfc | 460 | uint64_t niter; |
50654f6c | 461 | |
50654f6c ZD |
462 | /* Assumptions under that the rest of the information is valid. */ |
463 | rtx assumptions; | |
464 | ||
465 | /* Assumptions under that the loop ends before reaching the latch, | |
466 | even if value of niter_expr says otherwise. */ | |
467 | rtx noloop_assumptions; | |
468 | ||
469 | /* Condition under that the loop is infinite. */ | |
470 | rtx infinite; | |
471 | ||
472 | /* Whether the comparison is signed. */ | |
473 | bool signed_p; | |
474 | ||
475 | /* The mode in that niter_expr should be computed. */ | |
3d88d1cd | 476 | scalar_int_mode mode; |
50654f6c ZD |
477 | |
478 | /* The number of iterations of the loop. */ | |
479 | rtx niter_expr; | |
480 | }; | |
481 | ||
482 | extern void iv_analysis_loop_init (struct loop *); | |
3d88d1cd | 483 | extern bool iv_analyze (rtx_insn *, scalar_int_mode, rtx, struct rtx_iv *); |
1b20d55a | 484 | extern bool iv_analyze_result (rtx_insn *, rtx, struct rtx_iv *); |
3d88d1cd | 485 | extern bool iv_analyze_expr (rtx_insn *, scalar_int_mode, rtx, |
1b20d55a | 486 | struct rtx_iv *); |
50654f6c | 487 | extern rtx get_iv_value (struct rtx_iv *, rtx); |
3d88d1cd | 488 | extern bool biv_p (rtx_insn *, scalar_int_mode, rtx); |
50654f6c | 489 | extern void find_simple_exit (struct loop *, struct niter_desc *); |
50654f6c ZD |
490 | extern void iv_analysis_done (void); |
491 | ||
492 | extern struct niter_desc *get_simple_loop_desc (struct loop *loop); | |
493 | extern void free_simple_loop_desc (struct loop *loop); | |
494 | ||
495 | static inline struct niter_desc * | |
496 | simple_loop_desc (struct loop *loop) | |
497 | { | |
ef23e6a2 | 498 | return loop->simple_loop_desc; |
50654f6c ZD |
499 | } |
500 | ||
42fd6772 ZD |
501 | /* Accessors for the loop structures. */ |
502 | ||
0fc822d0 | 503 | /* Returns the loop with index NUM from FNs loop tree. */ |
42fd6772 ZD |
504 | |
505 | static inline struct loop * | |
0fc822d0 | 506 | get_loop (struct function *fn, unsigned num) |
42fd6772 | 507 | { |
0fc822d0 | 508 | return (*loops_for_fn (fn)->larray)[num]; |
42fd6772 ZD |
509 | } |
510 | ||
9ba025a2 ZD |
511 | /* Returns the number of superloops of LOOP. */ |
512 | ||
513 | static inline unsigned | |
514 | loop_depth (const struct loop *loop) | |
515 | { | |
9771b263 | 516 | return vec_safe_length (loop->superloops); |
9ba025a2 ZD |
517 | } |
518 | ||
519 | /* Returns the immediate superloop of LOOP, or NULL if LOOP is the outermost | |
520 | loop. */ | |
521 | ||
522 | static inline struct loop * | |
523 | loop_outer (const struct loop *loop) | |
524 | { | |
9771b263 | 525 | unsigned n = vec_safe_length (loop->superloops); |
9ba025a2 ZD |
526 | |
527 | if (n == 0) | |
528 | return NULL; | |
529 | ||
9771b263 | 530 | return (*loop->superloops)[n - 1]; |
9ba025a2 ZD |
531 | } |
532 | ||
07643d76 AM |
533 | /* Returns true if LOOP has at least one exit edge. */ |
534 | ||
535 | static inline bool | |
536 | loop_has_exit_edges (const struct loop *loop) | |
537 | { | |
538 | return loop->exits->next->e != NULL; | |
539 | } | |
540 | ||
0fc822d0 | 541 | /* Returns the list of loops in FN. */ |
42fd6772 | 542 | |
0fc822d0 RB |
543 | inline vec<loop_p, va_gc> * |
544 | get_loops (struct function *fn) | |
42fd6772 | 545 | { |
0fc822d0 RB |
546 | struct loops *loops = loops_for_fn (fn); |
547 | if (!loops) | |
42fd6772 ZD |
548 | return NULL; |
549 | ||
0fc822d0 | 550 | return loops->larray; |
42fd6772 ZD |
551 | } |
552 | ||
0fc822d0 | 553 | /* Returns the number of loops in FN (including the removed |
42fd6772 ZD |
554 | ones and the fake loop that forms the root of the loop tree). */ |
555 | ||
556 | static inline unsigned | |
0fc822d0 | 557 | number_of_loops (struct function *fn) |
42fd6772 | 558 | { |
0fc822d0 | 559 | struct loops *loops = loops_for_fn (fn); |
0d0e2af6 | 560 | if (!loops) |
42fd6772 ZD |
561 | return 0; |
562 | ||
0fc822d0 | 563 | return vec_safe_length (loops->larray); |
42fd6772 ZD |
564 | } |
565 | ||
f87000d0 ZD |
566 | /* Returns true if state of the loops satisfies all properties |
567 | described by FLAGS. */ | |
568 | ||
61183076 RB |
569 | static inline bool |
570 | loops_state_satisfies_p (function *fn, unsigned flags) | |
571 | { | |
572 | return (loops_for_fn (fn)->state & flags) == flags; | |
573 | } | |
574 | ||
f87000d0 ZD |
575 | static inline bool |
576 | loops_state_satisfies_p (unsigned flags) | |
577 | { | |
61183076 | 578 | return loops_state_satisfies_p (cfun, flags); |
f87000d0 ZD |
579 | } |
580 | ||
581 | /* Sets FLAGS to the loops state. */ | |
582 | ||
61183076 RB |
583 | static inline void |
584 | loops_state_set (function *fn, unsigned flags) | |
585 | { | |
586 | loops_for_fn (fn)->state |= flags; | |
587 | } | |
588 | ||
f87000d0 ZD |
589 | static inline void |
590 | loops_state_set (unsigned flags) | |
591 | { | |
61183076 | 592 | loops_state_set (cfun, flags); |
f87000d0 ZD |
593 | } |
594 | ||
595 | /* Clears FLAGS from the loops state. */ | |
596 | ||
61183076 RB |
597 | static inline void |
598 | loops_state_clear (function *fn, unsigned flags) | |
599 | { | |
600 | loops_for_fn (fn)->state &= ~flags; | |
601 | } | |
602 | ||
f87000d0 ZD |
603 | static inline void |
604 | loops_state_clear (unsigned flags) | |
605 | { | |
606 | if (!current_loops) | |
607 | return; | |
61183076 | 608 | loops_state_clear (cfun, flags); |
f87000d0 ZD |
609 | } |
610 | ||
a84a49b7 RB |
611 | /* Check loop structure invariants, if internal consistency checks are |
612 | enabled. */ | |
613 | ||
614 | static inline void | |
615 | checking_verify_loop_structure (void) | |
616 | { | |
617 | /* VERIFY_LOOP_STRUCTURE essentially asserts that no loops need fixups. | |
618 | ||
619 | The loop optimizers should never make changes to the CFG which | |
620 | require loop fixups. But the low level CFG manipulation code may | |
621 | set the flag conservatively. | |
622 | ||
623 | Go ahead and clear the flag here. That avoids the assert inside | |
624 | VERIFY_LOOP_STRUCTURE, and if there is an inconsistency in the loop | |
625 | structures VERIFY_LOOP_STRUCTURE will detect it. | |
626 | ||
627 | This also avoid the compile time cost of excessive fixups. */ | |
628 | loops_state_clear (LOOPS_NEED_FIXUP); | |
629 | if (flag_checking) | |
630 | verify_loop_structure (); | |
631 | } | |
632 | ||
42fd6772 ZD |
633 | /* Loop iterators. */ |
634 | ||
635 | /* Flags for loop iteration. */ | |
636 | ||
637 | enum li_flags | |
638 | { | |
677cc14d ZD |
639 | LI_INCLUDE_ROOT = 1, /* Include the fake root of the loop tree. */ |
640 | LI_FROM_INNERMOST = 2, /* Iterate over the loops in the reverse order, | |
641 | starting from innermost ones. */ | |
642 | LI_ONLY_INNERMOST = 4 /* Iterate only over innermost loops. */ | |
42fd6772 ZD |
643 | }; |
644 | ||
645 | /* The iterator for loops. */ | |
646 | ||
f0bd40b1 | 647 | struct loop_iterator |
42fd6772 | 648 | { |
61183076 | 649 | loop_iterator (function *fn, loop_p *loop, unsigned flags); |
f0bd40b1 RB |
650 | ~loop_iterator (); |
651 | ||
652 | inline loop_p next (); | |
653 | ||
61183076 RB |
654 | /* The function we are visiting. */ |
655 | function *fn; | |
656 | ||
677cc14d | 657 | /* The list of loops to visit. */ |
9771b263 | 658 | vec<int> to_visit; |
677cc14d ZD |
659 | |
660 | /* The index of the actual loop. */ | |
661 | unsigned idx; | |
f0bd40b1 | 662 | }; |
42fd6772 | 663 | |
f0bd40b1 RB |
664 | inline loop_p |
665 | loop_iterator::next () | |
42fd6772 | 666 | { |
677cc14d ZD |
667 | int anum; |
668 | ||
f0bd40b1 | 669 | while (this->to_visit.iterate (this->idx, &anum)) |
42fd6772 | 670 | { |
f0bd40b1 | 671 | this->idx++; |
61183076 | 672 | loop_p loop = get_loop (fn, anum); |
f0bd40b1 RB |
673 | if (loop) |
674 | return loop; | |
42fd6772 ZD |
675 | } |
676 | ||
f0bd40b1 | 677 | return NULL; |
42fd6772 ZD |
678 | } |
679 | ||
f0bd40b1 | 680 | inline |
61183076 | 681 | loop_iterator::loop_iterator (function *fn, loop_p *loop, unsigned flags) |
42fd6772 | 682 | { |
677cc14d ZD |
683 | struct loop *aloop; |
684 | unsigned i; | |
685 | int mn; | |
686 | ||
f0bd40b1 | 687 | this->idx = 0; |
61183076 RB |
688 | this->fn = fn; |
689 | if (!loops_for_fn (fn)) | |
42fd6772 | 690 | { |
f0bd40b1 | 691 | this->to_visit.create (0); |
42fd6772 ZD |
692 | *loop = NULL; |
693 | return; | |
694 | } | |
695 | ||
61183076 | 696 | this->to_visit.create (number_of_loops (fn)); |
677cc14d ZD |
697 | mn = (flags & LI_INCLUDE_ROOT) ? 0 : 1; |
698 | ||
699 | if (flags & LI_ONLY_INNERMOST) | |
700 | { | |
61183076 | 701 | for (i = 0; vec_safe_iterate (loops_for_fn (fn)->larray, i, &aloop); i++) |
677cc14d ZD |
702 | if (aloop != NULL |
703 | && aloop->inner == NULL | |
704 | && aloop->num >= mn) | |
f0bd40b1 | 705 | this->to_visit.quick_push (aloop->num); |
677cc14d ZD |
706 | } |
707 | else if (flags & LI_FROM_INNERMOST) | |
42fd6772 | 708 | { |
677cc14d | 709 | /* Push the loops to LI->TO_VISIT in postorder. */ |
61183076 | 710 | for (aloop = loops_for_fn (fn)->tree_root; |
677cc14d ZD |
711 | aloop->inner != NULL; |
712 | aloop = aloop->inner) | |
713 | continue; | |
714 | ||
715 | while (1) | |
716 | { | |
717 | if (aloop->num >= mn) | |
f0bd40b1 | 718 | this->to_visit.quick_push (aloop->num); |
677cc14d ZD |
719 | |
720 | if (aloop->next) | |
721 | { | |
722 | for (aloop = aloop->next; | |
723 | aloop->inner != NULL; | |
724 | aloop = aloop->inner) | |
725 | continue; | |
726 | } | |
9ba025a2 | 727 | else if (!loop_outer (aloop)) |
677cc14d ZD |
728 | break; |
729 | else | |
9ba025a2 | 730 | aloop = loop_outer (aloop); |
677cc14d | 731 | } |
42fd6772 ZD |
732 | } |
733 | else | |
734 | { | |
677cc14d | 735 | /* Push the loops to LI->TO_VISIT in preorder. */ |
61183076 | 736 | aloop = loops_for_fn (fn)->tree_root; |
677cc14d ZD |
737 | while (1) |
738 | { | |
739 | if (aloop->num >= mn) | |
f0bd40b1 | 740 | this->to_visit.quick_push (aloop->num); |
677cc14d ZD |
741 | |
742 | if (aloop->inner != NULL) | |
743 | aloop = aloop->inner; | |
744 | else | |
745 | { | |
746 | while (aloop != NULL && aloop->next == NULL) | |
9ba025a2 | 747 | aloop = loop_outer (aloop); |
677cc14d ZD |
748 | if (aloop == NULL) |
749 | break; | |
750 | aloop = aloop->next; | |
751 | } | |
752 | } | |
42fd6772 | 753 | } |
677cc14d | 754 | |
f0bd40b1 | 755 | *loop = this->next (); |
42fd6772 ZD |
756 | } |
757 | ||
f0bd40b1 RB |
758 | inline |
759 | loop_iterator::~loop_iterator () | |
760 | { | |
761 | this->to_visit.release (); | |
762 | } | |
677cc14d | 763 | |
f0bd40b1 | 764 | #define FOR_EACH_LOOP(LOOP, FLAGS) \ |
61183076 RB |
765 | for (loop_iterator li(cfun, &(LOOP), FLAGS); \ |
766 | (LOOP); \ | |
767 | (LOOP) = li.next ()) | |
768 | ||
769 | #define FOR_EACH_LOOP_FN(FN, LOOP, FLAGS) \ | |
650a8fcf | 770 | for (loop_iterator li(FN, &(LOOP), FLAGS); \ |
f0bd40b1 RB |
771 | (LOOP); \ |
772 | (LOOP) = li.next ()) | |
42fd6772 | 773 | |
8b11a64c | 774 | /* The properties of the target. */ |
4391924a RS |
775 | struct target_cfgloop { |
776 | /* Number of available registers. */ | |
777 | unsigned x_target_avail_regs; | |
8b11a64c | 778 | |
4391924a RS |
779 | /* Number of available registers that are call-clobbered. */ |
780 | unsigned x_target_clobbered_regs; | |
781 | ||
782 | /* Number of registers reserved for temporary expressions. */ | |
783 | unsigned x_target_res_regs; | |
784 | ||
785 | /* The cost for register when there still is some reserve, but we are | |
786 | approaching the number of available registers. */ | |
787 | unsigned x_target_reg_cost[2]; | |
788 | ||
789 | /* The cost for register when we need to spill. */ | |
790 | unsigned x_target_spill_cost[2]; | |
791 | }; | |
792 | ||
793 | extern struct target_cfgloop default_target_cfgloop; | |
794 | #if SWITCHABLE_TARGET | |
795 | extern struct target_cfgloop *this_target_cfgloop; | |
796 | #else | |
797 | #define this_target_cfgloop (&default_target_cfgloop) | |
798 | #endif | |
799 | ||
800 | #define target_avail_regs \ | |
801 | (this_target_cfgloop->x_target_avail_regs) | |
802 | #define target_clobbered_regs \ | |
803 | (this_target_cfgloop->x_target_clobbered_regs) | |
804 | #define target_res_regs \ | |
805 | (this_target_cfgloop->x_target_res_regs) | |
806 | #define target_reg_cost \ | |
807 | (this_target_cfgloop->x_target_reg_cost) | |
808 | #define target_spill_cost \ | |
809 | (this_target_cfgloop->x_target_spill_cost) | |
8b11a64c | 810 | |
5e962776 ZD |
811 | /* Register pressure estimation for induction variable optimizations & loop |
812 | invariant motion. */ | |
bec922f0 | 813 | extern unsigned estimate_reg_pressure_cost (unsigned, unsigned, bool, bool); |
5e962776 ZD |
814 | extern void init_set_costs (void); |
815 | ||
617b465c | 816 | /* Loop optimizer initialization. */ |
598ec7bd | 817 | extern void loop_optimizer_init (unsigned); |
61183076 RB |
818 | extern void loop_optimizer_finalize (function *); |
819 | inline void | |
820 | loop_optimizer_finalize () | |
821 | { | |
822 | loop_optimizer_finalize (cfun); | |
823 | } | |
617b465c ZD |
824 | |
825 | /* Optimization passes. */ | |
b17d5d7c ZD |
826 | enum |
827 | { | |
f8934be7 JH |
828 | UAP_UNROLL = 1, /* Enables unrolling of loops if it seems profitable. */ |
829 | UAP_UNROLL_ALL = 2 /* Enables unrolling of all loops. */ | |
b17d5d7c ZD |
830 | }; |
831 | ||
d73be268 ZD |
832 | extern void doloop_optimize_loops (void); |
833 | extern void move_loop_invariants (void); | |
9771b263 | 834 | extern vec<basic_block> get_loop_hot_path (const struct loop *loop); |
59587b18 | 835 | |
a5497b12 VK |
836 | /* Returns the outermost loop of the loop nest that contains LOOP.*/ |
837 | static inline struct loop * | |
838 | loop_outermost (struct loop *loop) | |
839 | { | |
9771b263 | 840 | unsigned n = vec_safe_length (loop->superloops); |
a5497b12 VK |
841 | |
842 | if (n <= 1) | |
843 | return loop; | |
844 | ||
9771b263 | 845 | return (*loop->superloops)[1]; |
a5497b12 VK |
846 | } |
847 | ||
807e902e | 848 | extern void record_niter_bound (struct loop *, const widest_int &, bool, bool); |
1ef88893 | 849 | extern HOST_WIDE_INT get_estimated_loop_iterations_int (struct loop *); |
199b1891 | 850 | extern HOST_WIDE_INT get_max_loop_iterations_int (const struct loop *); |
105e29c5 | 851 | extern HOST_WIDE_INT get_likely_max_loop_iterations_int (struct loop *); |
807e902e | 852 | extern bool get_estimated_loop_iterations (struct loop *loop, widest_int *nit); |
199b1891 | 853 | extern bool get_max_loop_iterations (const struct loop *loop, widest_int *nit); |
105e29c5 | 854 | extern bool get_likely_max_loop_iterations (struct loop *loop, widest_int *nit); |
4484a35a | 855 | extern int bb_loop_depth (const_basic_block); |
a5497b12 | 856 | |
807e902e | 857 | /* Converts VAL to widest_int. */ |
71343877 | 858 | |
807e902e KZ |
859 | static inline widest_int |
860 | gcov_type_to_wide_int (gcov_type val) | |
71343877 | 861 | { |
807e902e | 862 | HOST_WIDE_INT a[2]; |
71343877 | 863 | |
807e902e | 864 | a[0] = (unsigned HOST_WIDE_INT) val; |
71343877 AM |
865 | /* If HOST_BITS_PER_WIDE_INT == HOST_BITS_PER_WIDEST_INT, avoid shifting by |
866 | the size of type. */ | |
867 | val >>= HOST_BITS_PER_WIDE_INT - 1; | |
868 | val >>= 1; | |
807e902e | 869 | a[1] = (unsigned HOST_WIDE_INT) val; |
71343877 | 870 | |
807e902e | 871 | return widest_int::from_array (a, 2); |
71343877 | 872 | } |
59587b18 | 873 | #endif /* GCC_CFGLOOP_H */ |