]>
Commit | Line | Data |
---|---|---|
84eb345f | 1 | /* Induction variable canonicalization and loop peeling. |
d353bf18 | 2 | Copyright (C) 2004-2015 Free Software Foundation, Inc. |
48e1416a | 3 | |
bb445479 | 4 | This file is part of GCC. |
48e1416a | 5 | |
bb445479 | 6 | GCC is free software; you can redistribute it and/or modify it |
7 | under the terms of the GNU General Public License as published by the | |
8c4c00c1 | 8 | Free Software Foundation; either version 3, or (at your option) any |
bb445479 | 9 | later version. |
48e1416a | 10 | |
bb445479 | 11 | GCC is distributed in the hope that it will be useful, but WITHOUT |
12 | ANY 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. | |
48e1416a | 15 | |
bb445479 | 16 | You should have received a copy of the GNU General Public License |
8c4c00c1 | 17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
bb445479 | 19 | |
20 | /* This pass detects the loops that iterate a constant number of times, | |
48e1416a | 21 | adds a canonical induction variable (step -1, tested against 0) |
bb445479 | 22 | and replaces the exit test. This enables the less powerful rtl |
23 | level analysis to use this information. | |
24 | ||
25 | This might spoil the code in some cases (by increasing register pressure). | |
26 | Note that in the case the new variable is not needed, ivopts will get rid | |
27 | of it, so it might only be a problem when there are no other linear induction | |
28 | variables. In that case the created optimization possibilities are likely | |
29 | to pay up. | |
30 | ||
c836de3f | 31 | We also perform |
4cf494ec | 32 | - complete unrolling (or peeling) when the loops is rolling few enough |
c836de3f | 33 | times |
34 | - simple peeling (i.e. copying few initial iterations prior the loop) | |
35 | when number of iteration estimate is known (typically by the profile | |
36 | info). */ | |
bb445479 | 37 | |
38 | #include "config.h" | |
39 | #include "system.h" | |
40 | #include "coretypes.h" | |
41 | #include "tm.h" | |
b20a8bb4 | 42 | #include "alias.h" |
43 | #include "symtab.h" | |
bb445479 | 44 | #include "tree.h" |
b20a8bb4 | 45 | #include "fold-const.h" |
bb445479 | 46 | #include "tm_p.h" |
886c1262 | 47 | #include "profile.h" |
94ea8568 | 48 | #include "predict.h" |
94ea8568 | 49 | #include "hard-reg-set.h" |
94ea8568 | 50 | #include "function.h" |
51 | #include "dominance.h" | |
52 | #include "cfg.h" | |
bb445479 | 53 | #include "basic-block.h" |
ce084dfc | 54 | #include "gimple-pretty-print.h" |
bc61cadb | 55 | #include "tree-ssa-alias.h" |
56 | #include "internal-fn.h" | |
57 | #include "gimple-fold.h" | |
58 | #include "tree-eh.h" | |
59 | #include "gimple-expr.h" | |
073c1fd5 | 60 | #include "gimple.h" |
dcf1a1ec | 61 | #include "gimple-iterator.h" |
073c1fd5 | 62 | #include "gimple-ssa.h" |
1140c305 | 63 | #include "plugin-api.h" |
64 | #include "ipa-ref.h" | |
073c1fd5 | 65 | #include "cgraph.h" |
66 | #include "tree-cfg.h" | |
67 | #include "tree-phinodes.h" | |
68 | #include "ssa-iterators.h" | |
9ed99284 | 69 | #include "stringpool.h" |
073c1fd5 | 70 | #include "tree-ssanames.h" |
05d9c18a | 71 | #include "tree-ssa-loop-manip.h" |
72 | #include "tree-ssa-loop-niter.h" | |
073c1fd5 | 73 | #include "tree-ssa-loop.h" |
74 | #include "tree-into-ssa.h" | |
bb445479 | 75 | #include "cfgloop.h" |
76 | #include "tree-pass.h" | |
bb445479 | 77 | #include "tree-chrec.h" |
78 | #include "tree-scalar-evolution.h" | |
79 | #include "params.h" | |
80 | #include "flags.h" | |
81 | #include "tree-inline.h" | |
aa2ba534 | 82 | #include "target.h" |
424a4a92 | 83 | #include "tree-cfgcleanup.h" |
f7715905 | 84 | #include "builtins.h" |
bb445479 | 85 | |
604f7b8a | 86 | /* Specifies types of loops that may be unrolled. */ |
87 | ||
88 | enum unroll_level | |
89 | { | |
6414dd4c | 90 | UL_SINGLE_ITER, /* Only loops that exit immediately in the first |
604f7b8a | 91 | iteration. */ |
92 | UL_NO_GROWTH, /* Only loops whose unrolling will not cause increase | |
93 | of code size. */ | |
94 | UL_ALL /* All suitable loops. */ | |
95 | }; | |
96 | ||
bb445479 | 97 | /* Adds a canonical induction variable to LOOP iterating NITER times. EXIT |
98 | is the exit edge whose condition is replaced. */ | |
99 | ||
100 | static void | |
101 | create_canonical_iv (struct loop *loop, edge exit, tree niter) | |
102 | { | |
103 | edge in; | |
75a70cf9 | 104 | tree type, var; |
1a91d914 | 105 | gcond *cond; |
75a70cf9 | 106 | gimple_stmt_iterator incr_at; |
bb445479 | 107 | enum tree_code cmp; |
108 | ||
109 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
110 | { | |
111 | fprintf (dump_file, "Added canonical iv to loop %d, ", loop->num); | |
112 | print_generic_expr (dump_file, niter, TDF_SLIM); | |
113 | fprintf (dump_file, " iterations.\n"); | |
114 | } | |
115 | ||
1a91d914 | 116 | cond = as_a <gcond *> (last_stmt (exit->src)); |
cd665a06 | 117 | in = EDGE_SUCC (exit->src, 0); |
bb445479 | 118 | if (in == exit) |
cd665a06 | 119 | in = EDGE_SUCC (exit->src, 1); |
bb445479 | 120 | |
121 | /* Note that we do not need to worry about overflows, since | |
122 | type of niter is always unsigned and all comparisons are | |
123 | just for equality/nonequality -- i.e. everything works | |
124 | with a modulo arithmetics. */ | |
125 | ||
126 | type = TREE_TYPE (niter); | |
49d00087 | 127 | niter = fold_build2 (PLUS_EXPR, type, |
128 | niter, | |
129 | build_int_cst (type, 1)); | |
75a70cf9 | 130 | incr_at = gsi_last_bb (in->src); |
bb445479 | 131 | create_iv (niter, |
3c6185f1 | 132 | build_int_cst (type, -1), |
bb445479 | 133 | NULL_TREE, loop, |
134 | &incr_at, false, NULL, &var); | |
135 | ||
136 | cmp = (exit->flags & EDGE_TRUE_VALUE) ? EQ_EXPR : NE_EXPR; | |
75a70cf9 | 137 | gimple_cond_set_code (cond, cmp); |
138 | gimple_cond_set_lhs (cond, var); | |
139 | gimple_cond_set_rhs (cond, build_int_cst (type, 0)); | |
22aa74c4 | 140 | update_stmt (cond); |
bb445479 | 141 | } |
142 | ||
aa2ba534 | 143 | /* Describe size of loop as detected by tree_estimate_loop_size. */ |
144 | struct loop_size | |
145 | { | |
146 | /* Number of instructions in the loop. */ | |
147 | int overall; | |
148 | ||
149 | /* Number of instructions that will be likely optimized out in | |
150 | peeled iterations of loop (i.e. computation based on induction | |
151 | variable where induction variable starts at known constant.) */ | |
152 | int eliminated_by_peeling; | |
153 | ||
154 | /* Same statistics for last iteration of loop: it is smaller because | |
155 | instructions after exit are not executed. */ | |
156 | int last_iteration; | |
157 | int last_iteration_eliminated_by_peeling; | |
d583c979 | 158 | |
159 | /* If some IV computation will become constant. */ | |
160 | bool constant_iv; | |
161 | ||
162 | /* Number of call stmts that are not a builtin and are pure or const | |
163 | present on the hot path. */ | |
164 | int num_pure_calls_on_hot_path; | |
165 | /* Number of call stmts that are not a builtin and are not pure nor const | |
166 | present on the hot path. */ | |
167 | int num_non_pure_calls_on_hot_path; | |
168 | /* Number of statements other than calls in the loop. */ | |
169 | int non_call_stmts_on_hot_path; | |
170 | /* Number of branches seen on the hot path. */ | |
171 | int num_branches_on_hot_path; | |
aa2ba534 | 172 | }; |
173 | ||
174 | /* Return true if OP in STMT will be constant after peeling LOOP. */ | |
175 | ||
176 | static bool | |
177 | constant_after_peeling (tree op, gimple stmt, struct loop *loop) | |
178 | { | |
179 | affine_iv iv; | |
180 | ||
181 | if (is_gimple_min_invariant (op)) | |
182 | return true; | |
48e1416a | 183 | |
aa2ba534 | 184 | /* We can still fold accesses to constant arrays when index is known. */ |
185 | if (TREE_CODE (op) != SSA_NAME) | |
186 | { | |
187 | tree base = op; | |
188 | ||
189 | /* First make fast look if we see constant array inside. */ | |
190 | while (handled_component_p (base)) | |
191 | base = TREE_OPERAND (base, 0); | |
248022b2 | 192 | if ((DECL_P (base) |
df8d3e89 | 193 | && ctor_for_folding (base) != error_mark_node) |
aa2ba534 | 194 | || CONSTANT_CLASS_P (base)) |
195 | { | |
196 | /* If so, see if we understand all the indices. */ | |
197 | base = op; | |
198 | while (handled_component_p (base)) | |
199 | { | |
200 | if (TREE_CODE (base) == ARRAY_REF | |
201 | && !constant_after_peeling (TREE_OPERAND (base, 1), stmt, loop)) | |
202 | return false; | |
203 | base = TREE_OPERAND (base, 0); | |
204 | } | |
205 | return true; | |
206 | } | |
207 | return false; | |
208 | } | |
209 | ||
210 | /* Induction variables are constants. */ | |
211 | if (!simple_iv (loop, loop_containing_stmt (stmt), op, &iv, false)) | |
212 | return false; | |
213 | if (!is_gimple_min_invariant (iv.base)) | |
214 | return false; | |
215 | if (!is_gimple_min_invariant (iv.step)) | |
216 | return false; | |
217 | return true; | |
218 | } | |
219 | ||
d583c979 | 220 | /* Computes an estimated number of insns in LOOP. |
221 | EXIT (if non-NULL) is an exite edge that will be eliminated in all but last | |
222 | iteration of the loop. | |
223 | EDGE_TO_CANCEL (if non-NULL) is an non-exit edge eliminated in the last iteration | |
224 | of loop. | |
84eb345f | 225 | Return results in SIZE, estimate benefits for complete unrolling exiting by EXIT. |
04437ab6 | 226 | Stop estimating after UPPER_BOUND is met. Return true in this case. */ |
aa2ba534 | 227 | |
84eb345f | 228 | static bool |
229 | tree_estimate_loop_size (struct loop *loop, edge exit, edge edge_to_cancel, struct loop_size *size, | |
230 | int upper_bound) | |
aa2ba534 | 231 | { |
232 | basic_block *body = get_loop_body (loop); | |
233 | gimple_stmt_iterator gsi; | |
234 | unsigned int i; | |
235 | bool after_exit; | |
f1f41a6c | 236 | vec<basic_block> path = get_loop_hot_path (loop); |
aa2ba534 | 237 | |
238 | size->overall = 0; | |
239 | size->eliminated_by_peeling = 0; | |
240 | size->last_iteration = 0; | |
241 | size->last_iteration_eliminated_by_peeling = 0; | |
d583c979 | 242 | size->num_pure_calls_on_hot_path = 0; |
243 | size->num_non_pure_calls_on_hot_path = 0; | |
244 | size->non_call_stmts_on_hot_path = 0; | |
245 | size->num_branches_on_hot_path = 0; | |
246 | size->constant_iv = 0; | |
aa2ba534 | 247 | |
248 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
249 | fprintf (dump_file, "Estimating sizes for loop %i\n", loop->num); | |
250 | for (i = 0; i < loop->num_nodes; i++) | |
251 | { | |
c790d986 | 252 | if (edge_to_cancel && body[i] != edge_to_cancel->src |
253 | && dominated_by_p (CDI_DOMINATORS, body[i], edge_to_cancel->src)) | |
aa2ba534 | 254 | after_exit = true; |
255 | else | |
256 | after_exit = false; | |
257 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
258 | fprintf (dump_file, " BB: %i, after_exit: %i\n", body[i]->index, after_exit); | |
259 | ||
260 | for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi); gsi_next (&gsi)) | |
261 | { | |
262 | gimple stmt = gsi_stmt (gsi); | |
263 | int num = estimate_num_insns (stmt, &eni_size_weights); | |
264 | bool likely_eliminated = false; | |
d583c979 | 265 | bool likely_eliminated_last = false; |
266 | bool likely_eliminated_peeled = false; | |
aa2ba534 | 267 | |
268 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
269 | { | |
270 | fprintf (dump_file, " size: %3i ", num); | |
271 | print_gimple_stmt (dump_file, gsi_stmt (gsi), 0, 0); | |
272 | } | |
273 | ||
274 | /* Look for reasons why we might optimize this stmt away. */ | |
275 | ||
ae8a8b85 | 276 | if (gimple_has_side_effects (stmt)) |
277 | ; | |
aa2ba534 | 278 | /* Exit conditional. */ |
ae8a8b85 | 279 | else if (exit && body[i] == exit->src |
d583c979 | 280 | && stmt == last_stmt (exit->src)) |
aa2ba534 | 281 | { |
282 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
d583c979 | 283 | fprintf (dump_file, " Exit condition will be eliminated " |
284 | "in peeled copies.\n"); | |
285 | likely_eliminated_peeled = true; | |
286 | } | |
287 | else if (edge_to_cancel && body[i] == edge_to_cancel->src | |
288 | && stmt == last_stmt (edge_to_cancel->src)) | |
289 | { | |
290 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
291 | fprintf (dump_file, " Exit condition will be eliminated " | |
292 | "in last copy.\n"); | |
293 | likely_eliminated_last = true; | |
aa2ba534 | 294 | } |
295 | /* Sets of IV variables */ | |
296 | else if (gimple_code (stmt) == GIMPLE_ASSIGN | |
297 | && constant_after_peeling (gimple_assign_lhs (stmt), stmt, loop)) | |
298 | { | |
299 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
300 | fprintf (dump_file, " Induction variable computation will" | |
301 | " be folded away.\n"); | |
302 | likely_eliminated = true; | |
303 | } | |
304 | /* Assignments of IV variables. */ | |
305 | else if (gimple_code (stmt) == GIMPLE_ASSIGN | |
306 | && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME | |
d583c979 | 307 | && constant_after_peeling (gimple_assign_rhs1 (stmt), stmt, loop) |
aa2ba534 | 308 | && (gimple_assign_rhs_class (stmt) != GIMPLE_BINARY_RHS |
309 | || constant_after_peeling (gimple_assign_rhs2 (stmt), | |
310 | stmt, loop))) | |
311 | { | |
d583c979 | 312 | size->constant_iv = true; |
aa2ba534 | 313 | if (dump_file && (dump_flags & TDF_DETAILS)) |
314 | fprintf (dump_file, " Constant expression will be folded away.\n"); | |
315 | likely_eliminated = true; | |
316 | } | |
317 | /* Conditionals. */ | |
d583c979 | 318 | else if ((gimple_code (stmt) == GIMPLE_COND |
319 | && constant_after_peeling (gimple_cond_lhs (stmt), stmt, loop) | |
320 | && constant_after_peeling (gimple_cond_rhs (stmt), stmt, loop)) | |
321 | || (gimple_code (stmt) == GIMPLE_SWITCH | |
1a91d914 | 322 | && constant_after_peeling (gimple_switch_index ( |
323 | as_a <gswitch *> (stmt)), | |
324 | stmt, loop))) | |
aa2ba534 | 325 | { |
326 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
327 | fprintf (dump_file, " Constant conditional.\n"); | |
328 | likely_eliminated = true; | |
329 | } | |
330 | ||
331 | size->overall += num; | |
d583c979 | 332 | if (likely_eliminated || likely_eliminated_peeled) |
aa2ba534 | 333 | size->eliminated_by_peeling += num; |
334 | if (!after_exit) | |
335 | { | |
336 | size->last_iteration += num; | |
d583c979 | 337 | if (likely_eliminated || likely_eliminated_last) |
aa2ba534 | 338 | size->last_iteration_eliminated_by_peeling += num; |
339 | } | |
84eb345f | 340 | if ((size->overall * 3 / 2 - size->eliminated_by_peeling |
341 | - size->last_iteration_eliminated_by_peeling) > upper_bound) | |
342 | { | |
343 | free (body); | |
04437ab6 | 344 | path.release (); |
84eb345f | 345 | return true; |
346 | } | |
aa2ba534 | 347 | } |
348 | } | |
f1f41a6c | 349 | while (path.length ()) |
d583c979 | 350 | { |
f1f41a6c | 351 | basic_block bb = path.pop (); |
d583c979 | 352 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
353 | { | |
354 | gimple stmt = gsi_stmt (gsi); | |
355 | if (gimple_code (stmt) == GIMPLE_CALL) | |
356 | { | |
357 | int flags = gimple_call_flags (stmt); | |
358 | tree decl = gimple_call_fndecl (stmt); | |
359 | ||
360 | if (decl && DECL_IS_BUILTIN (decl) | |
361 | && is_inexpensive_builtin (decl)) | |
362 | ; | |
363 | else if (flags & (ECF_PURE | ECF_CONST)) | |
364 | size->num_pure_calls_on_hot_path++; | |
365 | else | |
366 | size->num_non_pure_calls_on_hot_path++; | |
367 | size->num_branches_on_hot_path ++; | |
368 | } | |
369 | else if (gimple_code (stmt) != GIMPLE_CALL | |
370 | && gimple_code (stmt) != GIMPLE_DEBUG) | |
371 | size->non_call_stmts_on_hot_path++; | |
372 | if (((gimple_code (stmt) == GIMPLE_COND | |
373 | && (!constant_after_peeling (gimple_cond_lhs (stmt), stmt, loop) | |
374 | || constant_after_peeling (gimple_cond_rhs (stmt), stmt, loop))) | |
375 | || (gimple_code (stmt) == GIMPLE_SWITCH | |
1a91d914 | 376 | && !constant_after_peeling (gimple_switch_index ( |
377 | as_a <gswitch *> (stmt)), | |
378 | stmt, loop))) | |
d583c979 | 379 | && (!exit || bb != exit->src)) |
380 | size->num_branches_on_hot_path++; | |
381 | } | |
382 | } | |
f1f41a6c | 383 | path.release (); |
aa2ba534 | 384 | if (dump_file && (dump_flags & TDF_DETAILS)) |
385 | fprintf (dump_file, "size: %i-%i, last_iteration: %i-%i\n", size->overall, | |
386 | size->eliminated_by_peeling, size->last_iteration, | |
387 | size->last_iteration_eliminated_by_peeling); | |
48e1416a | 388 | |
aa2ba534 | 389 | free (body); |
84eb345f | 390 | return false; |
aa2ba534 | 391 | } |
604f7b8a | 392 | |
aa2ba534 | 393 | /* Estimate number of insns of completely unrolled loop. |
394 | It is (NUNROLL + 1) * size of loop body with taking into account | |
395 | the fact that in last copy everything after exit conditional | |
396 | is dead and that some instructions will be eliminated after | |
397 | peeling. | |
604f7b8a | 398 | |
c31fb425 | 399 | Loop body is likely going to simplify further, this is difficult |
aa2ba534 | 400 | to guess, we just decrease the result by 1/3. */ |
604f7b8a | 401 | |
402 | static unsigned HOST_WIDE_INT | |
aa2ba534 | 403 | estimated_unrolled_size (struct loop_size *size, |
604f7b8a | 404 | unsigned HOST_WIDE_INT nunroll) |
405 | { | |
aa2ba534 | 406 | HOST_WIDE_INT unr_insns = ((nunroll) |
407 | * (HOST_WIDE_INT) (size->overall | |
408 | - size->eliminated_by_peeling)); | |
409 | if (!nunroll) | |
410 | unr_insns = 0; | |
411 | unr_insns += size->last_iteration - size->last_iteration_eliminated_by_peeling; | |
412 | ||
413 | unr_insns = unr_insns * 2 / 3; | |
604f7b8a | 414 | if (unr_insns <= 0) |
415 | unr_insns = 1; | |
604f7b8a | 416 | |
417 | return unr_insns; | |
418 | } | |
419 | ||
c790d986 | 420 | /* Loop LOOP is known to not loop. See if there is an edge in the loop |
421 | body that can be remove to make the loop to always exit and at | |
422 | the same time it does not make any code potentially executed | |
423 | during the last iteration dead. | |
424 | ||
4cf494ec | 425 | After complete unrolling we still may get rid of the conditional |
c790d986 | 426 | on the exit in the last copy even if we have no idea what it does. |
427 | This is quite common case for loops of form | |
428 | ||
429 | int a[5]; | |
430 | for (i=0;i<b;i++) | |
431 | a[i]=0; | |
432 | ||
433 | Here we prove the loop to iterate 5 times but we do not know | |
434 | it from induction variable. | |
435 | ||
436 | For now we handle only simple case where there is exit condition | |
437 | just before the latch block and the latch block contains no statements | |
438 | with side effect that may otherwise terminate the execution of loop | |
439 | (such as by EH or by terminating the program or longjmp). | |
440 | ||
441 | In the general case we may want to cancel the paths leading to statements | |
442 | loop-niter identified as having undefined effect in the last iteration. | |
443 | The other cases are hopefully rare and will be cleaned up later. */ | |
444 | ||
f86b328b | 445 | static edge |
c790d986 | 446 | loop_edge_to_cancel (struct loop *loop) |
447 | { | |
f1f41a6c | 448 | vec<edge> exits; |
c790d986 | 449 | unsigned i; |
450 | edge edge_to_cancel; | |
451 | gimple_stmt_iterator gsi; | |
452 | ||
453 | /* We want only one predecestor of the loop. */ | |
454 | if (EDGE_COUNT (loop->latch->preds) > 1) | |
455 | return NULL; | |
456 | ||
457 | exits = get_loop_exit_edges (loop); | |
458 | ||
f1f41a6c | 459 | FOR_EACH_VEC_ELT (exits, i, edge_to_cancel) |
c790d986 | 460 | { |
461 | /* Find the other edge than the loop exit | |
462 | leaving the conditoinal. */ | |
463 | if (EDGE_COUNT (edge_to_cancel->src->succs) != 2) | |
464 | continue; | |
465 | if (EDGE_SUCC (edge_to_cancel->src, 0) == edge_to_cancel) | |
466 | edge_to_cancel = EDGE_SUCC (edge_to_cancel->src, 1); | |
467 | else | |
468 | edge_to_cancel = EDGE_SUCC (edge_to_cancel->src, 0); | |
469 | ||
248022b2 | 470 | /* We only can handle conditionals. */ |
471 | if (!(edge_to_cancel->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) | |
472 | continue; | |
473 | ||
c790d986 | 474 | /* We should never have conditionals in the loop latch. */ |
475 | gcc_assert (edge_to_cancel->dest != loop->header); | |
476 | ||
477 | /* Check that it leads to loop latch. */ | |
478 | if (edge_to_cancel->dest != loop->latch) | |
479 | continue; | |
480 | ||
f1f41a6c | 481 | exits.release (); |
c790d986 | 482 | |
483 | /* Verify that the code in loop latch does nothing that may end program | |
484 | execution without really reaching the exit. This may include | |
485 | non-pure/const function calls, EH statements, volatile ASMs etc. */ | |
486 | for (gsi = gsi_start_bb (loop->latch); !gsi_end_p (gsi); gsi_next (&gsi)) | |
487 | if (gimple_has_side_effects (gsi_stmt (gsi))) | |
488 | return NULL; | |
489 | return edge_to_cancel; | |
490 | } | |
f1f41a6c | 491 | exits.release (); |
c790d986 | 492 | return NULL; |
493 | } | |
494 | ||
72276d01 | 495 | /* Remove all tests for exits that are known to be taken after LOOP was |
496 | peeled NPEELED times. Put gcc_unreachable before every statement | |
497 | known to not be executed. */ | |
498 | ||
499 | static bool | |
500 | remove_exits_and_undefined_stmts (struct loop *loop, unsigned int npeeled) | |
501 | { | |
502 | struct nb_iter_bound *elt; | |
503 | bool changed = false; | |
504 | ||
505 | for (elt = loop->bounds; elt; elt = elt->next) | |
506 | { | |
507 | /* If statement is known to be undefined after peeling, turn it | |
508 | into unreachable (or trap when debugging experience is supposed | |
509 | to be good). */ | |
510 | if (!elt->is_exit | |
796b6678 | 511 | && wi::ltu_p (elt->bound, npeeled)) |
72276d01 | 512 | { |
513 | gimple_stmt_iterator gsi = gsi_for_stmt (elt->stmt); | |
1a91d914 | 514 | gcall *stmt = gimple_build_call |
72276d01 | 515 | (builtin_decl_implicit (BUILT_IN_UNREACHABLE), 0); |
72276d01 | 516 | gimple_set_location (stmt, gimple_location (elt->stmt)); |
517 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
3b43af65 | 518 | split_block (gimple_bb (stmt), stmt); |
72276d01 | 519 | changed = true; |
520 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
521 | { | |
522 | fprintf (dump_file, "Forced statement unreachable: "); | |
523 | print_gimple_stmt (dump_file, elt->stmt, 0, 0); | |
524 | } | |
525 | } | |
526 | /* If we know the exit will be taken after peeling, update. */ | |
527 | else if (elt->is_exit | |
796b6678 | 528 | && wi::leu_p (elt->bound, npeeled)) |
72276d01 | 529 | { |
530 | basic_block bb = gimple_bb (elt->stmt); | |
531 | edge exit_edge = EDGE_SUCC (bb, 0); | |
532 | ||
533 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
534 | { | |
535 | fprintf (dump_file, "Forced exit to be taken: "); | |
536 | print_gimple_stmt (dump_file, elt->stmt, 0, 0); | |
537 | } | |
538 | if (!loop_exit_edge_p (loop, exit_edge)) | |
539 | exit_edge = EDGE_SUCC (bb, 1); | |
540 | gcc_checking_assert (loop_exit_edge_p (loop, exit_edge)); | |
1a91d914 | 541 | gcond *cond_stmt = as_a <gcond *> (elt->stmt); |
72276d01 | 542 | if (exit_edge->flags & EDGE_TRUE_VALUE) |
1a91d914 | 543 | gimple_cond_make_true (cond_stmt); |
72276d01 | 544 | else |
1a91d914 | 545 | gimple_cond_make_false (cond_stmt); |
546 | update_stmt (cond_stmt); | |
72276d01 | 547 | changed = true; |
548 | } | |
549 | } | |
550 | return changed; | |
551 | } | |
552 | ||
553 | /* Remove all exits that are known to be never taken because of the loop bound | |
554 | discovered. */ | |
555 | ||
556 | static bool | |
557 | remove_redundant_iv_tests (struct loop *loop) | |
558 | { | |
559 | struct nb_iter_bound *elt; | |
560 | bool changed = false; | |
561 | ||
562 | if (!loop->any_upper_bound) | |
563 | return false; | |
564 | for (elt = loop->bounds; elt; elt = elt->next) | |
565 | { | |
566 | /* Exit is pointless if it won't be taken before loop reaches | |
567 | upper bound. */ | |
568 | if (elt->is_exit && loop->any_upper_bound | |
796b6678 | 569 | && wi::ltu_p (loop->nb_iterations_upper_bound, elt->bound)) |
72276d01 | 570 | { |
571 | basic_block bb = gimple_bb (elt->stmt); | |
572 | edge exit_edge = EDGE_SUCC (bb, 0); | |
573 | struct tree_niter_desc niter; | |
574 | ||
575 | if (!loop_exit_edge_p (loop, exit_edge)) | |
576 | exit_edge = EDGE_SUCC (bb, 1); | |
577 | ||
578 | /* Only when we know the actual number of iterations, not | |
579 | just a bound, we can remove the exit. */ | |
580 | if (!number_of_iterations_exit (loop, exit_edge, | |
3a690dce | 581 | &niter, false, false) |
582 | || !integer_onep (niter.assumptions) | |
72276d01 | 583 | || !integer_zerop (niter.may_be_zero) |
584 | || !niter.niter | |
585 | || TREE_CODE (niter.niter) != INTEGER_CST | |
cd9b5516 | 586 | || !wi::ltu_p (loop->nb_iterations_upper_bound, |
587 | wi::to_widest (niter.niter))) | |
72276d01 | 588 | continue; |
589 | ||
590 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
591 | { | |
592 | fprintf (dump_file, "Removed pointless exit: "); | |
593 | print_gimple_stmt (dump_file, elt->stmt, 0, 0); | |
594 | } | |
1a91d914 | 595 | gcond *cond_stmt = as_a <gcond *> (elt->stmt); |
72276d01 | 596 | if (exit_edge->flags & EDGE_TRUE_VALUE) |
1a91d914 | 597 | gimple_cond_make_false (cond_stmt); |
72276d01 | 598 | else |
1a91d914 | 599 | gimple_cond_make_true (cond_stmt); |
600 | update_stmt (cond_stmt); | |
72276d01 | 601 | changed = true; |
602 | } | |
603 | } | |
604 | return changed; | |
605 | } | |
606 | ||
607 | /* Stores loops that will be unlooped after we process whole loop tree. */ | |
f1f41a6c | 608 | static vec<loop_p> loops_to_unloop; |
609 | static vec<int> loops_to_unloop_nunroll; | |
72276d01 | 610 | |
611 | /* Cancel all fully unrolled loops by putting __builtin_unreachable | |
612 | on the latch edge. | |
613 | We do it after all unrolling since unlooping moves basic blocks | |
614 | across loop boundaries trashing loop closed SSA form as well | |
615 | as SCEV info needed to be intact during unrolling. | |
616 | ||
c790d986 | 617 | IRRED_INVALIDATED is used to bookkeep if information about |
618 | irreducible regions may become invalid as a result | |
9f0ac045 | 619 | of the transformation. |
620 | LOOP_CLOSED_SSA_INVALIDATED is used to bookkepp the case | |
621 | when we need to go into loop closed SSA form. */ | |
bb445479 | 622 | |
f86b328b | 623 | static void |
72276d01 | 624 | unloop_loops (bitmap loop_closed_ssa_invalidated, |
625 | bool *irred_invalidated) | |
626 | { | |
f1f41a6c | 627 | while (loops_to_unloop.length ()) |
72276d01 | 628 | { |
f1f41a6c | 629 | struct loop *loop = loops_to_unloop.pop (); |
630 | int n_unroll = loops_to_unloop_nunroll.pop (); | |
72276d01 | 631 | basic_block latch = loop->latch; |
632 | edge latch_edge = loop_latch_edge (loop); | |
633 | int flags = latch_edge->flags; | |
634 | location_t locus = latch_edge->goto_locus; | |
1a91d914 | 635 | gcall *stmt; |
72276d01 | 636 | gimple_stmt_iterator gsi; |
637 | ||
638 | remove_exits_and_undefined_stmts (loop, n_unroll); | |
639 | ||
640 | /* Unloop destroys the latch edge. */ | |
641 | unloop (loop, irred_invalidated, loop_closed_ssa_invalidated); | |
642 | ||
643 | /* Create new basic block for the latch edge destination and wire | |
644 | it in. */ | |
645 | stmt = gimple_build_call (builtin_decl_implicit (BUILT_IN_UNREACHABLE), 0); | |
646 | latch_edge = make_edge (latch, create_basic_block (NULL, NULL, latch), flags); | |
647 | latch_edge->probability = 0; | |
648 | latch_edge->count = 0; | |
649 | latch_edge->flags |= flags; | |
650 | latch_edge->goto_locus = locus; | |
651 | ||
652 | latch_edge->dest->loop_father = current_loops->tree_root; | |
653 | latch_edge->dest->count = 0; | |
654 | latch_edge->dest->frequency = 0; | |
655 | set_immediate_dominator (CDI_DOMINATORS, latch_edge->dest, latch_edge->src); | |
656 | ||
657 | gsi = gsi_start_bb (latch_edge->dest); | |
658 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
659 | } | |
f1f41a6c | 660 | loops_to_unloop.release (); |
661 | loops_to_unloop_nunroll.release (); | |
72276d01 | 662 | } |
663 | ||
664 | /* Tries to unroll LOOP completely, i.e. NITER times. | |
665 | UL determines which loops we are allowed to unroll. | |
f55775aa | 666 | EXIT is the exit of the loop that should be eliminated. |
72276d01 | 667 | MAXITER specfy bound on number of iterations, -1 if it is |
f55775aa | 668 | not known or too large for HOST_WIDE_INT. The location |
669 | LOCUS corresponding to the loop is used when emitting | |
670 | a summary of the unroll to the dump file. */ | |
72276d01 | 671 | |
bb445479 | 672 | static bool |
7194de72 | 673 | try_unroll_loop_completely (struct loop *loop, |
bb445479 | 674 | edge exit, tree niter, |
c790d986 | 675 | enum unroll_level ul, |
f55775aa | 676 | HOST_WIDE_INT maxiter, |
677 | location_t locus) | |
bb445479 | 678 | { |
04e3ee8a | 679 | unsigned HOST_WIDE_INT n_unroll = 0, ninsns, unr_insns; |
aa2ba534 | 680 | struct loop_size size; |
c790d986 | 681 | bool n_unroll_found = false; |
c790d986 | 682 | edge edge_to_cancel = NULL; |
c836de3f | 683 | int report_flags = MSG_OPTIMIZED_LOCATIONS | TDF_RTL | TDF_DETAILS; |
bb445479 | 684 | |
c790d986 | 685 | /* See if we proved number of iterations to be low constant. |
bb445479 | 686 | |
c790d986 | 687 | EXIT is an edge that will be removed in all but last iteration of |
688 | the loop. | |
689 | ||
690 | EDGE_TO_CACNEL is an edge that will be removed from the last iteration | |
691 | of the unrolled sequence and is expected to make the final loop not | |
692 | rolling. | |
693 | ||
694 | If the number of execution of loop is determined by standard induction | |
695 | variable test, then EXIT and EDGE_TO_CANCEL are the two edges leaving | |
696 | from the iv test. */ | |
e913b5cd | 697 | if (tree_fits_uhwi_p (niter)) |
c790d986 | 698 | { |
e913b5cd | 699 | n_unroll = tree_to_uhwi (niter); |
c790d986 | 700 | n_unroll_found = true; |
701 | edge_to_cancel = EDGE_SUCC (exit->src, 0); | |
702 | if (edge_to_cancel == exit) | |
703 | edge_to_cancel = EDGE_SUCC (exit->src, 1); | |
704 | } | |
705 | /* We do not know the number of iterations and thus we can not eliminate | |
706 | the EXIT edge. */ | |
707 | else | |
52c6c602 | 708 | exit = NULL; |
c790d986 | 709 | |
710 | /* See if we can improve our estimate by using recorded loop bounds. */ | |
c790d986 | 711 | if (maxiter >= 0 |
712 | && (!n_unroll_found || (unsigned HOST_WIDE_INT)maxiter < n_unroll)) | |
713 | { | |
714 | n_unroll = maxiter; | |
715 | n_unroll_found = true; | |
716 | /* Loop terminates before the IV variable test, so we can not | |
717 | remove it in the last iteration. */ | |
718 | edge_to_cancel = NULL; | |
719 | } | |
720 | ||
721 | if (!n_unroll_found) | |
bb445479 | 722 | return false; |
bb445479 | 723 | |
04e3ee8a | 724 | if (n_unroll > (unsigned) PARAM_VALUE (PARAM_MAX_COMPLETELY_PEEL_TIMES)) |
725 | { | |
726 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
727 | fprintf (dump_file, "Not unrolling loop %d " | |
728 | "(--param max-completely-peeled-times limit reached).\n", | |
729 | loop->num); | |
730 | return false; | |
731 | } | |
bb445479 | 732 | |
c790d986 | 733 | if (!edge_to_cancel) |
734 | edge_to_cancel = loop_edge_to_cancel (loop); | |
735 | ||
bb445479 | 736 | if (n_unroll) |
737 | { | |
c790d986 | 738 | sbitmap wont_exit; |
739 | edge e; | |
740 | unsigned i; | |
84eb345f | 741 | bool large; |
1e094109 | 742 | vec<edge> to_remove = vNULL; |
604f7b8a | 743 | if (ul == UL_SINGLE_ITER) |
bb445479 | 744 | return false; |
745 | ||
84eb345f | 746 | large = tree_estimate_loop_size |
747 | (loop, exit, edge_to_cancel, &size, | |
748 | PARAM_VALUE (PARAM_MAX_COMPLETELY_PEELED_INSNS)); | |
aa2ba534 | 749 | ninsns = size.overall; |
84eb345f | 750 | if (large) |
751 | { | |
752 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
753 | fprintf (dump_file, "Not unrolling loop %d: it is too large.\n", | |
754 | loop->num); | |
755 | return false; | |
756 | } | |
bb445479 | 757 | |
aa2ba534 | 758 | unr_insns = estimated_unrolled_size (&size, n_unroll); |
d88fd237 | 759 | if (dump_file && (dump_flags & TDF_DETAILS)) |
760 | { | |
761 | fprintf (dump_file, " Loop size: %d\n", (int) ninsns); | |
762 | fprintf (dump_file, " Estimated size after unrolling: %d\n", | |
763 | (int) unr_insns); | |
764 | } | |
765 | ||
d583c979 | 766 | /* If the code is going to shrink, we don't need to be extra cautious |
767 | on guessing if the unrolling is going to be profitable. */ | |
768 | if (unr_insns | |
769 | /* If there is IV variable that will become constant, we save | |
770 | one instruction in the loop prologue we do not account | |
771 | otherwise. */ | |
772 | <= ninsns + (size.constant_iv != false)) | |
773 | ; | |
c790d986 | 774 | /* We unroll only inner loops, because we do not consider it profitable |
775 | otheriwse. We still can cancel loopback edge of not rolling loop; | |
776 | this is always a good idea. */ | |
d583c979 | 777 | else if (ul == UL_NO_GROWTH) |
778 | { | |
779 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
780 | fprintf (dump_file, "Not unrolling loop %d: size would grow.\n", | |
781 | loop->num); | |
782 | return false; | |
783 | } | |
4cf494ec | 784 | /* Outer loops tend to be less interesting candidates for complete |
d583c979 | 785 | unrolling unless we can do a lot of propagation into the inner loop |
786 | body. For now we disable outer loop unrolling when the code would | |
787 | grow. */ | |
788 | else if (loop->inner) | |
c790d986 | 789 | { |
790 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
d583c979 | 791 | fprintf (dump_file, "Not unrolling loop %d: " |
c790d986 | 792 | "it is not innermost and code would grow.\n", |
793 | loop->num); | |
794 | return false; | |
795 | } | |
d583c979 | 796 | /* If there is call on a hot path through the loop, then |
797 | there is most probably not much to optimize. */ | |
798 | else if (size.num_non_pure_calls_on_hot_path) | |
f00b5e35 | 799 | { |
800 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
d583c979 | 801 | fprintf (dump_file, "Not unrolling loop %d: " |
802 | "contains call and code would grow.\n", | |
f00b5e35 | 803 | loop->num); |
804 | return false; | |
805 | } | |
d583c979 | 806 | /* If there is pure/const call in the function, then we |
807 | can still optimize the unrolled loop body if it contains | |
808 | some other interesting code than the calls and code | |
809 | storing or cumulating the return value. */ | |
810 | else if (size.num_pure_calls_on_hot_path | |
811 | /* One IV increment, one test, one ivtmp store | |
c31fb425 | 812 | and one useful stmt. That is about minimal loop |
d583c979 | 813 | doing pure call. */ |
814 | && (size.non_call_stmts_on_hot_path | |
815 | <= 3 + size.num_pure_calls_on_hot_path)) | |
604f7b8a | 816 | { |
604f7b8a | 817 | if (dump_file && (dump_flags & TDF_DETAILS)) |
d583c979 | 818 | fprintf (dump_file, "Not unrolling loop %d: " |
819 | "contains just pure calls and code would grow.\n", | |
820 | loop->num); | |
821 | return false; | |
822 | } | |
823 | /* Complette unrolling is major win when control flow is removed and | |
824 | one big basic block is created. If the loop contains control flow | |
825 | the optimization may still be a win because of eliminating the loop | |
826 | overhead but it also may blow the branch predictor tables. | |
827 | Limit number of branches on the hot path through the peeled | |
828 | sequence. */ | |
829 | else if (size.num_branches_on_hot_path * (int)n_unroll | |
830 | > PARAM_VALUE (PARAM_MAX_PEEL_BRANCHES)) | |
831 | { | |
832 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
833 | fprintf (dump_file, "Not unrolling loop %d: " | |
834 | " number of branches on hot path in the unrolled sequence" | |
835 | " reach --param max-peel-branches limit.\n", | |
836 | loop->num); | |
837 | return false; | |
838 | } | |
839 | else if (unr_insns | |
840 | > (unsigned) PARAM_VALUE (PARAM_MAX_COMPLETELY_PEELED_INSNS)) | |
841 | { | |
842 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
843 | fprintf (dump_file, "Not unrolling loop %d: " | |
844 | "(--param max-completely-peeled-insns limit reached).\n", | |
c790d986 | 845 | loop->num); |
d88fd237 | 846 | return false; |
604f7b8a | 847 | } |
c836de3f | 848 | dump_printf_loc (report_flags, locus, |
849 | "loop turned into non-loop; it never loops.\n"); | |
fb54ef7c | 850 | |
01020a5f | 851 | initialize_original_copy_tables (); |
fb54ef7c | 852 | wont_exit = sbitmap_alloc (n_unroll + 1); |
53c5d9d4 | 853 | bitmap_ones (wont_exit); |
08b7917c | 854 | bitmap_clear_bit (wont_exit, 0); |
fb54ef7c | 855 | |
75a70cf9 | 856 | if (!gimple_duplicate_loop_to_header_edge (loop, loop_preheader_edge (loop), |
857 | n_unroll, wont_exit, | |
858 | exit, &to_remove, | |
859 | DLTHE_FLAG_UPDATE_FREQ | |
860 | | DLTHE_FLAG_COMPLETTE_PEEL)) | |
bb445479 | 861 | { |
01020a5f | 862 | free_original_copy_tables (); |
fb54ef7c | 863 | free (wont_exit); |
d583c979 | 864 | if (dump_file && (dump_flags & TDF_DETAILS)) |
865 | fprintf (dump_file, "Failed to duplicate the loop\n"); | |
bb445479 | 866 | return false; |
867 | } | |
40ffaada | 868 | |
f1f41a6c | 869 | FOR_EACH_VEC_ELT (to_remove, i, e) |
40ffaada | 870 | { |
871 | bool ok = remove_path (e); | |
872 | gcc_assert (ok); | |
873 | } | |
874 | ||
f1f41a6c | 875 | to_remove.release (); |
fb54ef7c | 876 | free (wont_exit); |
01020a5f | 877 | free_original_copy_tables (); |
bb445479 | 878 | } |
bb445479 | 879 | |
72276d01 | 880 | |
c790d986 | 881 | /* Remove the conditional from the last copy of the loop. */ |
882 | if (edge_to_cancel) | |
883 | { | |
1a91d914 | 884 | gcond *cond = as_a <gcond *> (last_stmt (edge_to_cancel->src)); |
c790d986 | 885 | if (edge_to_cancel->flags & EDGE_TRUE_VALUE) |
886 | gimple_cond_make_false (cond); | |
887 | else | |
888 | gimple_cond_make_true (cond); | |
889 | update_stmt (cond); | |
890 | /* Do not remove the path. Doing so may remove outer loop | |
891 | and confuse bookkeeping code in tree_unroll_loops_completelly. */ | |
892 | } | |
c790d986 | 893 | |
72276d01 | 894 | /* Store the loop for later unlooping and exit removal. */ |
f1f41a6c | 895 | loops_to_unloop.safe_push (loop); |
896 | loops_to_unloop_nunroll.safe_push (n_unroll); | |
095dcfa3 | 897 | |
f55775aa | 898 | if (dump_enabled_p ()) |
c790d986 | 899 | { |
900 | if (!n_unroll) | |
f55775aa | 901 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, locus, |
6ee2edad | 902 | "loop turned into non-loop; it never loops\n"); |
c790d986 | 903 | else |
f55775aa | 904 | { |
905 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, locus, | |
6ee2edad | 906 | "loop with %d iterations completely unrolled", |
907 | (int) (n_unroll + 1)); | |
f55775aa | 908 | if (profile_info) |
909 | dump_printf (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, | |
910 | " (header execution count %d)", | |
911 | (int)loop->header->count); | |
912 | dump_printf (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, "\n"); | |
913 | } | |
914 | } | |
915 | ||
916 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
917 | { | |
c790d986 | 918 | if (exit) |
919 | fprintf (dump_file, "Exit condition of peeled iterations was " | |
920 | "eliminated.\n"); | |
921 | if (edge_to_cancel) | |
922 | fprintf (dump_file, "Last iteration exit edge was proved true.\n"); | |
923 | else | |
924 | fprintf (dump_file, "Latch of last iteration was marked by " | |
925 | "__builtin_unreachable ().\n"); | |
926 | } | |
bb445479 | 927 | |
928 | return true; | |
929 | } | |
930 | ||
c836de3f | 931 | /* Return number of instructions after peeling. */ |
932 | static unsigned HOST_WIDE_INT | |
933 | estimated_peeled_sequence_size (struct loop_size *size, | |
934 | unsigned HOST_WIDE_INT npeel) | |
935 | { | |
936 | return MAX (npeel * (HOST_WIDE_INT) (size->overall | |
937 | - size->eliminated_by_peeling), 1); | |
938 | } | |
939 | ||
940 | /* If the loop is expected to iterate N times and is | |
941 | small enough, duplicate the loop body N+1 times before | |
942 | the loop itself. This way the hot path will never | |
943 | enter the loop. | |
944 | Parameters are the same as for try_unroll_loops_completely */ | |
945 | ||
946 | static bool | |
947 | try_peel_loop (struct loop *loop, | |
948 | edge exit, tree niter, | |
949 | HOST_WIDE_INT maxiter) | |
950 | { | |
951 | int npeel; | |
952 | struct loop_size size; | |
953 | int peeled_size; | |
954 | sbitmap wont_exit; | |
955 | unsigned i; | |
956 | vec<edge> to_remove = vNULL; | |
957 | edge e; | |
958 | ||
959 | /* If the iteration bound is known and large, then we can safely eliminate | |
960 | the check in peeled copies. */ | |
961 | if (TREE_CODE (niter) != INTEGER_CST) | |
962 | exit = NULL; | |
963 | ||
964 | if (!flag_peel_loops || PARAM_VALUE (PARAM_MAX_PEEL_TIMES) <= 0) | |
965 | return false; | |
966 | ||
967 | /* Peel only innermost loops. */ | |
968 | if (loop->inner) | |
969 | { | |
970 | if (dump_file) | |
971 | fprintf (dump_file, "Not peeling: outer loop\n"); | |
972 | return false; | |
973 | } | |
974 | ||
975 | if (!optimize_loop_for_speed_p (loop)) | |
976 | { | |
977 | if (dump_file) | |
978 | fprintf (dump_file, "Not peeling: cold loop\n"); | |
979 | return false; | |
980 | } | |
981 | ||
982 | /* Check if there is an estimate on the number of iterations. */ | |
983 | npeel = estimated_loop_iterations_int (loop); | |
984 | if (npeel < 0) | |
985 | { | |
986 | if (dump_file) | |
987 | fprintf (dump_file, "Not peeling: number of iterations is not " | |
988 | "estimated\n"); | |
989 | return false; | |
990 | } | |
991 | if (maxiter >= 0 && maxiter <= npeel) | |
992 | { | |
993 | if (dump_file) | |
994 | fprintf (dump_file, "Not peeling: upper bound is known so can " | |
4cf494ec | 995 | "unroll completely\n"); |
c836de3f | 996 | return false; |
997 | } | |
998 | ||
999 | /* We want to peel estimated number of iterations + 1 (so we never | |
1000 | enter the loop on quick path). Check against PARAM_MAX_PEEL_TIMES | |
1001 | and be sure to avoid overflows. */ | |
1002 | if (npeel > PARAM_VALUE (PARAM_MAX_PEEL_TIMES) - 1) | |
1003 | { | |
1004 | if (dump_file) | |
1005 | fprintf (dump_file, "Not peeling: rolls too much " | |
1006 | "(%i + 1 > --param max-peel-times)\n", npeel); | |
1007 | return false; | |
1008 | } | |
1009 | npeel++; | |
1010 | ||
1011 | /* Check peeled loops size. */ | |
1012 | tree_estimate_loop_size (loop, exit, NULL, &size, | |
1013 | PARAM_VALUE (PARAM_MAX_PEELED_INSNS)); | |
1014 | if ((peeled_size = estimated_peeled_sequence_size (&size, npeel)) | |
1015 | > PARAM_VALUE (PARAM_MAX_PEELED_INSNS)) | |
1016 | { | |
1017 | if (dump_file) | |
1018 | fprintf (dump_file, "Not peeling: peeled sequence size is too large " | |
1019 | "(%i insns > --param max-peel-insns)", peeled_size); | |
1020 | return false; | |
1021 | } | |
1022 | ||
1023 | /* Duplicate possibly eliminating the exits. */ | |
1024 | initialize_original_copy_tables (); | |
1025 | wont_exit = sbitmap_alloc (npeel + 1); | |
1026 | bitmap_ones (wont_exit); | |
1027 | bitmap_clear_bit (wont_exit, 0); | |
1028 | if (!gimple_duplicate_loop_to_header_edge (loop, loop_preheader_edge (loop), | |
1029 | npeel, wont_exit, | |
1030 | exit, &to_remove, | |
1031 | DLTHE_FLAG_UPDATE_FREQ | |
1032 | | DLTHE_FLAG_COMPLETTE_PEEL)) | |
1033 | { | |
1034 | free_original_copy_tables (); | |
1035 | free (wont_exit); | |
1036 | return false; | |
1037 | } | |
1038 | FOR_EACH_VEC_ELT (to_remove, i, e) | |
1039 | { | |
1040 | bool ok = remove_path (e); | |
1041 | gcc_assert (ok); | |
1042 | } | |
1043 | free (wont_exit); | |
1044 | free_original_copy_tables (); | |
1045 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1046 | { | |
1047 | fprintf (dump_file, "Peeled loop %d, %i times.\n", | |
1048 | loop->num, npeel); | |
1049 | } | |
1050 | if (loop->any_upper_bound) | |
1051 | loop->nb_iterations_upper_bound -= npeel; | |
1052 | loop->nb_iterations_estimate = 0; | |
1053 | /* Make sure to mark loop cold so we do not try to peel it more. */ | |
1054 | scale_loop_profile (loop, 1, 0); | |
1055 | loop->header->count = 0; | |
1056 | return true; | |
1057 | } | |
7194de72 | 1058 | /* Adds a canonical induction variable to LOOP if suitable. |
48e1416a | 1059 | CREATE_IV is true if we may create a new iv. UL determines |
604f7b8a | 1060 | which loops we are allowed to completely unroll. If TRY_EVAL is true, we try |
48e1416a | 1061 | to determine the number of iterations of a loop by direct evaluation. |
72276d01 | 1062 | Returns true if cfg is changed. */ |
bb445479 | 1063 | |
1064 | static bool | |
7194de72 | 1065 | canonicalize_loop_induction_variables (struct loop *loop, |
604f7b8a | 1066 | bool create_iv, enum unroll_level ul, |
72276d01 | 1067 | bool try_eval) |
bb445479 | 1068 | { |
1069 | edge exit = NULL; | |
1070 | tree niter; | |
72276d01 | 1071 | HOST_WIDE_INT maxiter; |
1072 | bool modified = false; | |
f55775aa | 1073 | location_t locus = UNKNOWN_LOCATION; |
bb445479 | 1074 | |
0c3c2e56 | 1075 | niter = number_of_latch_executions (loop); |
f55775aa | 1076 | exit = single_exit (loop); |
bb445479 | 1077 | if (TREE_CODE (niter) == INTEGER_CST) |
f55775aa | 1078 | locus = gimple_location (last_stmt (exit->src)); |
b091dc59 | 1079 | else |
1080 | { | |
1081 | /* If the loop has more than one exit, try checking all of them | |
1082 | for # of iterations determinable through scev. */ | |
f55775aa | 1083 | if (!exit) |
b091dc59 | 1084 | niter = find_loop_niter (loop, &exit); |
1085 | ||
1086 | /* Finally if everything else fails, try brute force evaluation. */ | |
1087 | if (try_eval | |
1088 | && (chrec_contains_undetermined (niter) | |
1089 | || TREE_CODE (niter) != INTEGER_CST)) | |
1090 | niter = find_loop_niter_by_eval (loop, &exit); | |
1091 | ||
f55775aa | 1092 | if (exit) |
1093 | locus = gimple_location (last_stmt (exit->src)); | |
1094 | ||
c790d986 | 1095 | if (TREE_CODE (niter) != INTEGER_CST) |
1096 | exit = NULL; | |
b091dc59 | 1097 | } |
bb445479 | 1098 | |
c790d986 | 1099 | /* We work exceptionally hard here to estimate the bound |
1100 | by find_loop_niter_by_eval. Be sure to keep it for future. */ | |
1101 | if (niter && TREE_CODE (niter) == INTEGER_CST) | |
57337fec | 1102 | { |
5de9d3ed | 1103 | record_niter_bound (loop, wi::to_widest (niter), |
57337fec | 1104 | exit == single_likely_exit (loop), true); |
1105 | } | |
c790d986 | 1106 | |
72276d01 | 1107 | /* Force re-computation of loop bounds so we can remove redundant exits. */ |
1108 | maxiter = max_loop_iterations_int (loop); | |
1109 | ||
c790d986 | 1110 | if (dump_file && (dump_flags & TDF_DETAILS) |
1111 | && TREE_CODE (niter) == INTEGER_CST) | |
bb445479 | 1112 | { |
1113 | fprintf (dump_file, "Loop %d iterates ", loop->num); | |
1114 | print_generic_expr (dump_file, niter, TDF_SLIM); | |
1115 | fprintf (dump_file, " times.\n"); | |
1116 | } | |
c790d986 | 1117 | if (dump_file && (dump_flags & TDF_DETAILS) |
72276d01 | 1118 | && maxiter >= 0) |
c790d986 | 1119 | { |
1120 | fprintf (dump_file, "Loop %d iterates at most %i times.\n", loop->num, | |
72276d01 | 1121 | (int)maxiter); |
c790d986 | 1122 | } |
bb445479 | 1123 | |
72276d01 | 1124 | /* Remove exits that are known to be never taken based on loop bound. |
1125 | Needs to be called after compilation of max_loop_iterations_int that | |
1126 | populates the loop bounds. */ | |
1127 | modified |= remove_redundant_iv_tests (loop); | |
1128 | ||
f55775aa | 1129 | if (try_unroll_loop_completely (loop, exit, niter, ul, maxiter, locus)) |
bb445479 | 1130 | return true; |
1131 | ||
c790d986 | 1132 | if (create_iv |
57337fec | 1133 | && niter && !chrec_contains_undetermined (niter) |
1134 | && exit && just_once_each_iteration_p (loop, exit->src)) | |
bb445479 | 1135 | create_canonical_iv (loop, exit, niter); |
1136 | ||
c836de3f | 1137 | if (ul == UL_ALL) |
1138 | modified |= try_peel_loop (loop, exit, niter, maxiter); | |
1139 | ||
72276d01 | 1140 | return modified; |
bb445479 | 1141 | } |
1142 | ||
1143 | /* The main entry point of the pass. Adds canonical induction variables | |
7194de72 | 1144 | to the suitable loops. */ |
bb445479 | 1145 | |
4c641bf8 | 1146 | unsigned int |
7194de72 | 1147 | canonicalize_induction_variables (void) |
bb445479 | 1148 | { |
bb445479 | 1149 | struct loop *loop; |
053fdd99 | 1150 | bool changed = false; |
c790d986 | 1151 | bool irred_invalidated = false; |
9f0ac045 | 1152 | bitmap loop_closed_ssa_invalidated = BITMAP_ALLOC (NULL); |
48e1416a | 1153 | |
72276d01 | 1154 | free_numbers_of_iterations_estimates (); |
1155 | estimate_numbers_of_iterations (); | |
1156 | ||
f21d4d00 | 1157 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
bb445479 | 1158 | { |
17519ba0 | 1159 | changed |= canonicalize_loop_induction_variables (loop, |
1160 | true, UL_SINGLE_ITER, | |
72276d01 | 1161 | true); |
bb445479 | 1162 | } |
ea1c5c31 | 1163 | gcc_assert (!need_ssa_update_p (cfun)); |
bb445479 | 1164 | |
72276d01 | 1165 | unloop_loops (loop_closed_ssa_invalidated, &irred_invalidated); |
c790d986 | 1166 | if (irred_invalidated |
1167 | && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) | |
1168 | mark_irreducible_loops (); | |
1169 | ||
08162157 | 1170 | /* Clean up the information about numbers of iterations, since brute force |
1171 | evaluation could reveal new information. */ | |
1172 | scev_reset (); | |
1173 | ||
9f0ac045 | 1174 | if (!bitmap_empty_p (loop_closed_ssa_invalidated)) |
1175 | { | |
1176 | gcc_checking_assert (loops_state_satisfies_p (LOOP_CLOSED_SSA)); | |
1177 | rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); | |
1178 | } | |
1179 | BITMAP_FREE (loop_closed_ssa_invalidated); | |
1180 | ||
bb445479 | 1181 | if (changed) |
4c641bf8 | 1182 | return TODO_cleanup_cfg; |
1183 | return 0; | |
bb445479 | 1184 | } |
1185 | ||
2ebfc881 | 1186 | /* Propagate VAL into all uses of SSA_NAME. */ |
1187 | ||
1188 | static void | |
1189 | propagate_into_all_uses (tree ssa_name, tree val) | |
1190 | { | |
1191 | imm_use_iterator iter; | |
1192 | gimple use_stmt; | |
1193 | ||
1194 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, ssa_name) | |
1195 | { | |
1196 | gimple_stmt_iterator use_stmt_gsi = gsi_for_stmt (use_stmt); | |
1197 | use_operand_p use; | |
1198 | ||
1199 | FOR_EACH_IMM_USE_ON_STMT (use, iter) | |
1200 | SET_USE (use, val); | |
1201 | ||
1202 | if (is_gimple_assign (use_stmt) | |
1203 | && get_gimple_rhs_class (gimple_assign_rhs_code (use_stmt)) | |
1204 | == GIMPLE_SINGLE_RHS) | |
1205 | { | |
1206 | tree rhs = gimple_assign_rhs1 (use_stmt); | |
1207 | ||
1208 | if (TREE_CODE (rhs) == ADDR_EXPR) | |
1209 | recompute_tree_invariant_for_addr_expr (rhs); | |
1210 | } | |
1211 | ||
1212 | fold_stmt_inplace (&use_stmt_gsi); | |
1213 | update_stmt (use_stmt); | |
f4ea772b | 1214 | maybe_clean_or_replace_eh_stmt (use_stmt, use_stmt); |
2ebfc881 | 1215 | } |
1216 | } | |
1217 | ||
1218 | /* Propagate constant SSA_NAMEs defined in basic block BB. */ | |
1219 | ||
1220 | static void | |
1221 | propagate_constants_for_unrolling (basic_block bb) | |
1222 | { | |
2ebfc881 | 1223 | /* Look for degenerate PHI nodes with constant argument. */ |
1a91d914 | 1224 | for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi); ) |
2ebfc881 | 1225 | { |
1a91d914 | 1226 | gphi *phi = gsi.phi (); |
2ebfc881 | 1227 | tree result = gimple_phi_result (phi); |
1228 | tree arg = gimple_phi_arg_def (phi, 0); | |
1229 | ||
1230 | if (gimple_phi_num_args (phi) == 1 && TREE_CODE (arg) == INTEGER_CST) | |
1231 | { | |
1232 | propagate_into_all_uses (result, arg); | |
1233 | gsi_remove (&gsi, true); | |
1234 | release_ssa_name (result); | |
1235 | } | |
1236 | else | |
1237 | gsi_next (&gsi); | |
1238 | } | |
1239 | ||
1240 | /* Look for assignments to SSA names with constant RHS. */ | |
1a91d914 | 1241 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); ) |
2ebfc881 | 1242 | { |
1243 | gimple stmt = gsi_stmt (gsi); | |
1244 | tree lhs; | |
1245 | ||
1246 | if (is_gimple_assign (stmt) | |
fca2aa67 | 1247 | && gimple_assign_rhs_code (stmt) == INTEGER_CST |
2ebfc881 | 1248 | && (lhs = gimple_assign_lhs (stmt), TREE_CODE (lhs) == SSA_NAME) |
fca2aa67 | 1249 | && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)) |
2ebfc881 | 1250 | { |
1251 | propagate_into_all_uses (lhs, gimple_assign_rhs1 (stmt)); | |
1252 | gsi_remove (&gsi, true); | |
1253 | release_ssa_name (lhs); | |
1254 | } | |
1255 | else | |
1256 | gsi_next (&gsi); | |
1257 | } | |
1258 | } | |
1259 | ||
042301ef | 1260 | /* Process loops from innermost to outer, stopping at the innermost |
1261 | loop we unrolled. */ | |
1262 | ||
1263 | static bool | |
1264 | tree_unroll_loops_completely_1 (bool may_increase_size, bool unroll_outer, | |
d70aebca | 1265 | vec<loop_p, va_heap>& father_stack, |
042301ef | 1266 | struct loop *loop) |
1267 | { | |
1268 | struct loop *loop_father; | |
1269 | bool changed = false; | |
1270 | struct loop *inner; | |
1271 | enum unroll_level ul; | |
1272 | ||
1273 | /* Process inner loops first. */ | |
1274 | for (inner = loop->inner; inner != NULL; inner = inner->next) | |
1275 | changed |= tree_unroll_loops_completely_1 (may_increase_size, | |
1276 | unroll_outer, father_stack, | |
1277 | inner); | |
1278 | ||
1279 | /* If we changed an inner loop we cannot process outer loops in this | |
1280 | iteration because SSA form is not up-to-date. Continue with | |
1281 | siblings of outer loops instead. */ | |
1282 | if (changed) | |
1283 | return true; | |
1284 | ||
3d483a94 | 1285 | /* Don't unroll #pragma omp simd loops until the vectorizer |
1286 | attempts to vectorize those. */ | |
4c73695b | 1287 | if (loop->force_vectorize) |
3d483a94 | 1288 | return false; |
1289 | ||
042301ef | 1290 | /* Try to unroll this loop. */ |
1291 | loop_father = loop_outer (loop); | |
1292 | if (!loop_father) | |
1293 | return false; | |
1294 | ||
1295 | if (may_increase_size && optimize_loop_nest_for_speed_p (loop) | |
1296 | /* Unroll outermost loops only if asked to do so or they do | |
1297 | not cause code growth. */ | |
1298 | && (unroll_outer || loop_outer (loop_father))) | |
1299 | ul = UL_ALL; | |
1300 | else | |
1301 | ul = UL_NO_GROWTH; | |
1302 | ||
1303 | if (canonicalize_loop_induction_variables | |
1304 | (loop, false, ul, !flag_tree_loop_ivcanon)) | |
1305 | { | |
1306 | /* If we'll continue unrolling, we need to propagate constants | |
1307 | within the new basic blocks to fold away induction variable | |
1308 | computations; otherwise, the size might blow up before the | |
1309 | iteration is complete and the IR eventually cleaned up. */ | |
1310 | if (loop_outer (loop_father) && !loop_father->aux) | |
1311 | { | |
1312 | father_stack.safe_push (loop_father); | |
1313 | loop_father->aux = loop_father; | |
1314 | } | |
1315 | ||
1316 | return true; | |
1317 | } | |
1318 | ||
1319 | return false; | |
1320 | } | |
1321 | ||
604f7b8a | 1322 | /* Unroll LOOPS completely if they iterate just few times. Unless |
1323 | MAY_INCREASE_SIZE is true, perform the unrolling only if the | |
1324 | size of the code does not increase. */ | |
bb445479 | 1325 | |
4c641bf8 | 1326 | unsigned int |
d88fd237 | 1327 | tree_unroll_loops_completely (bool may_increase_size, bool unroll_outer) |
bb445479 | 1328 | { |
4997014d | 1329 | auto_vec<loop_p, 16> father_stack; |
d88fd237 | 1330 | bool changed; |
793a0ab5 | 1331 | int iteration = 0; |
9f0ac045 | 1332 | bool irred_invalidated = false; |
bb445479 | 1333 | |
d88fd237 | 1334 | do |
bb445479 | 1335 | { |
d88fd237 | 1336 | changed = false; |
9f0ac045 | 1337 | bitmap loop_closed_ssa_invalidated = NULL; |
1338 | ||
1339 | if (loops_state_satisfies_p (LOOP_CLOSED_SSA)) | |
1340 | loop_closed_ssa_invalidated = BITMAP_ALLOC (NULL); | |
bb445479 | 1341 | |
72276d01 | 1342 | free_numbers_of_iterations_estimates (); |
1343 | estimate_numbers_of_iterations (); | |
1344 | ||
042301ef | 1345 | changed = tree_unroll_loops_completely_1 (may_increase_size, |
1346 | unroll_outer, father_stack, | |
1347 | current_loops->tree_root); | |
d88fd237 | 1348 | if (changed) |
1349 | { | |
2ebfc881 | 1350 | struct loop **iter; |
1351 | unsigned i; | |
1352 | ||
72276d01 | 1353 | /* Be sure to skip unlooped loops while procesing father_stack |
1354 | array. */ | |
f1f41a6c | 1355 | FOR_EACH_VEC_ELT (loops_to_unloop, i, iter) |
72276d01 | 1356 | (*iter)->aux = NULL; |
f1f41a6c | 1357 | FOR_EACH_VEC_ELT (father_stack, i, iter) |
72276d01 | 1358 | if (!(*iter)->aux) |
1359 | *iter = NULL; | |
1360 | unloop_loops (loop_closed_ssa_invalidated, &irred_invalidated); | |
c790d986 | 1361 | |
72276d01 | 1362 | /* We can not use TODO_update_ssa_no_phi because VOPS gets confused. */ |
9f0ac045 | 1363 | if (loop_closed_ssa_invalidated |
1364 | && !bitmap_empty_p (loop_closed_ssa_invalidated)) | |
1365 | rewrite_into_loop_closed_ssa (loop_closed_ssa_invalidated, | |
1366 | TODO_update_ssa); | |
1367 | else | |
1368 | update_ssa (TODO_update_ssa); | |
ea1c5c31 | 1369 | |
2ebfc881 | 1370 | /* Propagate the constants within the new basic blocks. */ |
f1f41a6c | 1371 | FOR_EACH_VEC_ELT (father_stack, i, iter) |
72276d01 | 1372 | if (*iter) |
1373 | { | |
1374 | unsigned j; | |
1375 | basic_block *body = get_loop_body_in_dom_order (*iter); | |
1376 | for (j = 0; j < (*iter)->num_nodes; j++) | |
1377 | propagate_constants_for_unrolling (body[j]); | |
1378 | free (body); | |
1379 | (*iter)->aux = NULL; | |
1380 | } | |
f1f41a6c | 1381 | father_stack.truncate (0); |
2ebfc881 | 1382 | |
d88fd237 | 1383 | /* This will take care of removing completely unrolled loops |
1384 | from the loop structures so we can continue unrolling now | |
1385 | innermost loops. */ | |
b2a225ba | 1386 | if (cleanup_tree_cfg ()) |
1387 | update_ssa (TODO_update_ssa_only_virtuals); | |
d88fd237 | 1388 | |
1389 | /* Clean up the information about numbers of iterations, since | |
1390 | complete unrolling might have invalidated it. */ | |
1391 | scev_reset (); | |
9f0ac045 | 1392 | #ifdef ENABLE_CHECKING |
1393 | if (loops_state_satisfies_p (LOOP_CLOSED_SSA)) | |
1394 | verify_loop_closed_ssa (true); | |
1395 | #endif | |
d88fd237 | 1396 | } |
9f0ac045 | 1397 | if (loop_closed_ssa_invalidated) |
1398 | BITMAP_FREE (loop_closed_ssa_invalidated); | |
d88fd237 | 1399 | } |
793a0ab5 | 1400 | while (changed |
1401 | && ++iteration <= PARAM_VALUE (PARAM_MAX_UNROLL_ITERATIONS)); | |
08162157 | 1402 | |
f1f41a6c | 1403 | father_stack.release (); |
2ebfc881 | 1404 | |
9f0ac045 | 1405 | if (irred_invalidated |
1406 | && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)) | |
1407 | mark_irreducible_loops (); | |
1408 | ||
4c641bf8 | 1409 | return 0; |
bb445479 | 1410 | } |
f86b328b | 1411 | |
1412 | /* Canonical induction variable creation pass. */ | |
1413 | ||
f86b328b | 1414 | namespace { |
1415 | ||
1416 | const pass_data pass_data_iv_canon = | |
1417 | { | |
1418 | GIMPLE_PASS, /* type */ | |
1419 | "ivcanon", /* name */ | |
1420 | OPTGROUP_LOOP, /* optinfo_flags */ | |
f86b328b | 1421 | TV_TREE_LOOP_IVCANON, /* tv_id */ |
1422 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
1423 | 0, /* properties_provided */ | |
1424 | 0, /* properties_destroyed */ | |
1425 | 0, /* todo_flags_start */ | |
1426 | 0, /* todo_flags_finish */ | |
1427 | }; | |
1428 | ||
1429 | class pass_iv_canon : public gimple_opt_pass | |
1430 | { | |
1431 | public: | |
1432 | pass_iv_canon (gcc::context *ctxt) | |
1433 | : gimple_opt_pass (pass_data_iv_canon, ctxt) | |
1434 | {} | |
1435 | ||
1436 | /* opt_pass methods: */ | |
31315c24 | 1437 | virtual bool gate (function *) { return flag_tree_loop_ivcanon != 0; } |
65b0537f | 1438 | virtual unsigned int execute (function *fun); |
f86b328b | 1439 | |
1440 | }; // class pass_iv_canon | |
1441 | ||
65b0537f | 1442 | unsigned int |
1443 | pass_iv_canon::execute (function *fun) | |
1444 | { | |
1445 | if (number_of_loops (fun) <= 1) | |
1446 | return 0; | |
1447 | ||
1448 | return canonicalize_induction_variables (); | |
1449 | } | |
1450 | ||
f86b328b | 1451 | } // anon namespace |
1452 | ||
1453 | gimple_opt_pass * | |
1454 | make_pass_iv_canon (gcc::context *ctxt) | |
1455 | { | |
1456 | return new pass_iv_canon (ctxt); | |
1457 | } | |
1458 | ||
1459 | /* Complete unrolling of loops. */ | |
1460 | ||
f86b328b | 1461 | namespace { |
1462 | ||
1463 | const pass_data pass_data_complete_unroll = | |
1464 | { | |
1465 | GIMPLE_PASS, /* type */ | |
1466 | "cunroll", /* name */ | |
1467 | OPTGROUP_LOOP, /* optinfo_flags */ | |
f86b328b | 1468 | TV_COMPLETE_UNROLL, /* tv_id */ |
1469 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
1470 | 0, /* properties_provided */ | |
1471 | 0, /* properties_destroyed */ | |
1472 | 0, /* todo_flags_start */ | |
1473 | 0, /* todo_flags_finish */ | |
1474 | }; | |
1475 | ||
1476 | class pass_complete_unroll : public gimple_opt_pass | |
1477 | { | |
1478 | public: | |
1479 | pass_complete_unroll (gcc::context *ctxt) | |
1480 | : gimple_opt_pass (pass_data_complete_unroll, ctxt) | |
1481 | {} | |
1482 | ||
1483 | /* opt_pass methods: */ | |
65b0537f | 1484 | virtual unsigned int execute (function *); |
f86b328b | 1485 | |
1486 | }; // class pass_complete_unroll | |
1487 | ||
65b0537f | 1488 | unsigned int |
1489 | pass_complete_unroll::execute (function *fun) | |
1490 | { | |
1491 | if (number_of_loops (fun) <= 1) | |
1492 | return 0; | |
1493 | ||
1494 | return tree_unroll_loops_completely (flag_unroll_loops | |
1495 | || flag_peel_loops | |
1496 | || optimize >= 3, true); | |
1497 | } | |
1498 | ||
f86b328b | 1499 | } // anon namespace |
1500 | ||
1501 | gimple_opt_pass * | |
1502 | make_pass_complete_unroll (gcc::context *ctxt) | |
1503 | { | |
1504 | return new pass_complete_unroll (ctxt); | |
1505 | } | |
1506 | ||
1507 | /* Complete unrolling of inner loops. */ | |
1508 | ||
f86b328b | 1509 | namespace { |
1510 | ||
1511 | const pass_data pass_data_complete_unrolli = | |
1512 | { | |
1513 | GIMPLE_PASS, /* type */ | |
1514 | "cunrolli", /* name */ | |
1515 | OPTGROUP_LOOP, /* optinfo_flags */ | |
f86b328b | 1516 | TV_COMPLETE_UNROLL, /* tv_id */ |
1517 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
1518 | 0, /* properties_provided */ | |
1519 | 0, /* properties_destroyed */ | |
1520 | 0, /* todo_flags_start */ | |
8b88439e | 1521 | 0, /* todo_flags_finish */ |
f86b328b | 1522 | }; |
1523 | ||
1524 | class pass_complete_unrolli : public gimple_opt_pass | |
1525 | { | |
1526 | public: | |
1527 | pass_complete_unrolli (gcc::context *ctxt) | |
1528 | : gimple_opt_pass (pass_data_complete_unrolli, ctxt) | |
1529 | {} | |
1530 | ||
1531 | /* opt_pass methods: */ | |
31315c24 | 1532 | virtual bool gate (function *) { return optimize >= 2; } |
65b0537f | 1533 | virtual unsigned int execute (function *); |
f86b328b | 1534 | |
1535 | }; // class pass_complete_unrolli | |
1536 | ||
65b0537f | 1537 | unsigned int |
1538 | pass_complete_unrolli::execute (function *fun) | |
1539 | { | |
1540 | unsigned ret = 0; | |
1541 | ||
1542 | loop_optimizer_init (LOOPS_NORMAL | |
1543 | | LOOPS_HAVE_RECORDED_EXITS); | |
1544 | if (number_of_loops (fun) > 1) | |
1545 | { | |
1546 | scev_initialize (); | |
1547 | ret = tree_unroll_loops_completely (optimize >= 3, false); | |
1548 | free_numbers_of_iterations_estimates (); | |
1549 | scev_finalize (); | |
1550 | } | |
1551 | loop_optimizer_finalize (); | |
1552 | ||
1553 | return ret; | |
1554 | } | |
1555 | ||
f86b328b | 1556 | } // anon namespace |
1557 | ||
1558 | gimple_opt_pass * | |
1559 | make_pass_complete_unrolli (gcc::context *ctxt) | |
1560 | { | |
1561 | return new pass_complete_unrolli (ctxt); | |
1562 | } | |
1563 | ||
1564 |