Move pr66680.f90 to libgomp
[gcc.git] / gcc / tree-ssa-loop-ch.c
1 /* Loop header copying on trees.
2 Copyright (C) 2004-2016 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
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
8 Free Software Foundation; either version 3, or (at your option) any
9 later version.
10
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.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
19
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "tree.h"
25 #include "gimple.h"
26 #include "cfghooks.h"
27 #include "tree-pass.h"
28 #include "gimple-ssa.h"
29 #include "gimple-iterator.h"
30 #include "tree-cfg.h"
31 #include "tree-into-ssa.h"
32 #include "cfgloop.h"
33 #include "tree-inline.h"
34 #include "tree-ssa-scopedtables.h"
35 #include "tree-ssa-threadedge.h"
36
37 /* Duplicates headers of loops if they are small enough, so that the statements
38 in the loop body are always executed when the loop is entered. This
39 increases effectiveness of code motion optimizations, and reduces the need
40 for loop preconditioning. */
41
42 /* Check whether we should duplicate HEADER of LOOP. At most *LIMIT
43 instructions should be duplicated, limit is decreased by the actual
44 amount. */
45
46 static bool
47 should_duplicate_loop_header_p (basic_block header, struct loop *loop,
48 int *limit)
49 {
50 gimple_stmt_iterator bsi;
51 gimple *last;
52
53 /* Do not copy one block more than once (we do not really want to do
54 loop peeling here). */
55 if (header->aux)
56 return false;
57
58 /* Loop header copying usually increases size of the code. This used not to
59 be true, since quite often it is possible to verify that the condition is
60 satisfied in the first iteration and therefore to eliminate it. Jump
61 threading handles these cases now. */
62 if (optimize_loop_for_size_p (loop))
63 return false;
64
65 gcc_assert (EDGE_COUNT (header->succs) > 0);
66 if (single_succ_p (header))
67 return false;
68 if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest)
69 && flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 1)->dest))
70 return false;
71
72 /* If this is not the original loop header, we want it to have just
73 one predecessor in order to match the && pattern. */
74 if (header != loop->header && !single_pred_p (header))
75 return false;
76
77 last = last_stmt (header);
78 if (gimple_code (last) != GIMPLE_COND)
79 return false;
80
81 /* Approximately copy the conditions that used to be used in jump.c --
82 at most 20 insns and no calls. */
83 for (bsi = gsi_start_bb (header); !gsi_end_p (bsi); gsi_next (&bsi))
84 {
85 last = gsi_stmt (bsi);
86
87 if (gimple_code (last) == GIMPLE_LABEL)
88 continue;
89
90 if (is_gimple_debug (last))
91 continue;
92
93 if (is_gimple_call (last))
94 return false;
95
96 *limit -= estimate_num_insns (last, &eni_size_weights);
97 if (*limit < 0)
98 return false;
99 }
100
101 return true;
102 }
103
104 /* Checks whether LOOP is a do-while style loop. */
105
106 static bool
107 do_while_loop_p (struct loop *loop)
108 {
109 gimple *stmt = last_stmt (loop->latch);
110
111 /* If the latch of the loop is not empty, it is not a do-while loop. */
112 if (stmt
113 && gimple_code (stmt) != GIMPLE_LABEL)
114 return false;
115
116 /* If the header contains just a condition, it is not a do-while loop. */
117 stmt = last_and_only_stmt (loop->header);
118 if (stmt
119 && gimple_code (stmt) == GIMPLE_COND)
120 return false;
121
122 return true;
123 }
124
125 namespace {
126
127 /* Common superclass for both header-copying phases. */
128 class ch_base : public gimple_opt_pass
129 {
130 protected:
131 ch_base (pass_data data, gcc::context *ctxt)
132 : gimple_opt_pass (data, ctxt)
133 {}
134
135 /* Copies headers of all loops in FUN for which process_loop_p is true. */
136 unsigned int copy_headers (function *fun);
137
138 /* Return true to copy headers of LOOP or false to skip. */
139 virtual bool process_loop_p (struct loop *loop) = 0;
140 };
141
142 const pass_data pass_data_ch =
143 {
144 GIMPLE_PASS, /* type */
145 "ch", /* name */
146 OPTGROUP_LOOP, /* optinfo_flags */
147 TV_TREE_CH, /* tv_id */
148 ( PROP_cfg | PROP_ssa ), /* properties_required */
149 0, /* properties_provided */
150 0, /* properties_destroyed */
151 0, /* todo_flags_start */
152 0, /* todo_flags_finish */
153 };
154
155 class pass_ch : public ch_base
156 {
157 public:
158 pass_ch (gcc::context *ctxt)
159 : ch_base (pass_data_ch, ctxt)
160 {}
161
162 /* opt_pass methods: */
163 virtual bool gate (function *) { return flag_tree_ch != 0; }
164
165 /* Initialize and finalize loop structures, copying headers inbetween. */
166 virtual unsigned int execute (function *);
167
168 opt_pass * clone () { return new pass_ch (m_ctxt); }
169
170 protected:
171 /* ch_base method: */
172 virtual bool process_loop_p (struct loop *loop);
173 }; // class pass_ch
174
175 const pass_data pass_data_ch_vect =
176 {
177 GIMPLE_PASS, /* type */
178 "ch_vect", /* name */
179 OPTGROUP_LOOP, /* optinfo_flags */
180 TV_TREE_CH, /* tv_id */
181 ( PROP_cfg | PROP_ssa ), /* properties_required */
182 0, /* properties_provided */
183 0, /* properties_destroyed */
184 0, /* todo_flags_start */
185 0, /* todo_flags_finish */
186 };
187
188 /* This is a more aggressive version of the same pass, designed to run just
189 before if-conversion and vectorization, to put more loops into the form
190 required for those phases. */
191 class pass_ch_vect : public ch_base
192 {
193 public:
194 pass_ch_vect (gcc::context *ctxt)
195 : ch_base (pass_data_ch_vect, ctxt)
196 {}
197
198 /* opt_pass methods: */
199 virtual bool gate (function *fun)
200 {
201 return flag_tree_ch != 0
202 && (flag_tree_loop_vectorize != 0 || fun->has_force_vectorize_loops);
203 }
204
205 /* Just copy headers, no initialization/finalization of loop structures. */
206 virtual unsigned int execute (function *);
207
208 protected:
209 /* ch_base method: */
210 virtual bool process_loop_p (struct loop *loop);
211 }; // class pass_ch_vect
212
213 /* For all loops, copy the condition at the end of the loop body in front
214 of the loop. This is beneficial since it increases efficiency of
215 code motion optimizations. It also saves one jump on entry to the loop. */
216
217 unsigned int
218 ch_base::copy_headers (function *fun)
219 {
220 struct loop *loop;
221 basic_block header;
222 edge exit, entry;
223 basic_block *bbs, *copied_bbs;
224 unsigned n_bbs;
225 unsigned bbs_size;
226 bool changed = false;
227
228 if (number_of_loops (fun) <= 1)
229 return 0;
230
231 bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (fun));
232 copied_bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (fun));
233 bbs_size = n_basic_blocks_for_fn (fun);
234
235 FOR_EACH_LOOP (loop, 0)
236 {
237 /* Copy at most 20 insns. */
238 int limit = 20;
239
240 header = loop->header;
241
242 /* If the loop is already a do-while style one (either because it was
243 written as such, or because jump threading transformed it into one),
244 we might be in fact peeling the first iteration of the loop. This
245 in general is not a good idea. */
246 if (!process_loop_p (loop))
247 continue;
248
249 /* Iterate the header copying up to limit; this takes care of the cases
250 like while (a && b) {...}, where we want to have both of the conditions
251 copied. TODO -- handle while (a || b) - like cases, by not requiring
252 the header to have just a single successor and copying up to
253 postdominator. */
254
255 exit = NULL;
256 n_bbs = 0;
257 while (should_duplicate_loop_header_p (header, loop, &limit))
258 {
259 /* Find a successor of header that is inside a loop; i.e. the new
260 header after the condition is copied. */
261 if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest))
262 exit = EDGE_SUCC (header, 0);
263 else
264 exit = EDGE_SUCC (header, 1);
265 bbs[n_bbs++] = header;
266 gcc_assert (bbs_size > n_bbs);
267 header = exit->dest;
268 }
269
270 if (!exit)
271 continue;
272
273 if (dump_file && (dump_flags & TDF_DETAILS))
274 fprintf (dump_file,
275 "Duplicating header of the loop %d up to edge %d->%d.\n",
276 loop->num, exit->src->index, exit->dest->index);
277
278 /* Ensure that the header will have just the latch as a predecessor
279 inside the loop. */
280 if (!single_pred_p (exit->dest))
281 exit = single_pred_edge (split_edge (exit));
282
283 entry = loop_preheader_edge (loop);
284
285 propagate_threaded_block_debug_into (exit->dest, entry->dest);
286 if (!gimple_duplicate_sese_region (entry, exit, bbs, n_bbs, copied_bbs,
287 true))
288 {
289 fprintf (dump_file, "Duplication failed.\n");
290 continue;
291 }
292
293 /* If the loop has the form "for (i = j; i < j + 10; i++)" then
294 this copying can introduce a case where we rely on undefined
295 signed overflow to eliminate the preheader condition, because
296 we assume that "j < j + 10" is true. We don't want to warn
297 about that case for -Wstrict-overflow, because in general we
298 don't warn about overflow involving loops. Prevent the
299 warning by setting the no_warning flag in the condition. */
300 if (warn_strict_overflow > 0)
301 {
302 unsigned int i;
303
304 for (i = 0; i < n_bbs; ++i)
305 {
306 gimple_stmt_iterator bsi;
307
308 for (bsi = gsi_start_bb (copied_bbs[i]);
309 !gsi_end_p (bsi);
310 gsi_next (&bsi))
311 {
312 gimple *stmt = gsi_stmt (bsi);
313 if (gimple_code (stmt) == GIMPLE_COND)
314 gimple_set_no_warning (stmt, true);
315 else if (is_gimple_assign (stmt))
316 {
317 enum tree_code rhs_code = gimple_assign_rhs_code (stmt);
318 if (TREE_CODE_CLASS (rhs_code) == tcc_comparison)
319 gimple_set_no_warning (stmt, true);
320 }
321 }
322 }
323 }
324
325 /* Ensure that the latch and the preheader is simple (we know that they
326 are not now, since there was the loop exit condition. */
327 split_edge (loop_preheader_edge (loop));
328 split_edge (loop_latch_edge (loop));
329
330 changed = true;
331 }
332
333 if (changed)
334 update_ssa (TODO_update_ssa);
335 free (bbs);
336 free (copied_bbs);
337
338 return changed ? TODO_cleanup_cfg : 0;
339 }
340
341 /* Initialize the loop structures we need, and finalize after. */
342
343 unsigned int
344 pass_ch::execute (function *fun)
345 {
346 loop_optimizer_init (LOOPS_HAVE_PREHEADERS
347 | LOOPS_HAVE_SIMPLE_LATCHES);
348
349 unsigned int res = copy_headers (fun);
350
351 loop_optimizer_finalize ();
352 return res;
353 }
354
355 /* Assume an earlier phase has already initialized all the loop structures that
356 we need here (and perhaps others too), and that these will be finalized by
357 a later phase. */
358
359 unsigned int
360 pass_ch_vect::execute (function *fun)
361 {
362 return copy_headers (fun);
363 }
364
365 /* Apply header copying according to a very simple test of do-while shape. */
366
367 bool
368 pass_ch::process_loop_p (struct loop *loop)
369 {
370 return !do_while_loop_p (loop);
371 }
372
373 /* Apply header-copying to loops where we might enable vectorization. */
374
375 bool
376 pass_ch_vect::process_loop_p (struct loop *loop)
377 {
378 if (!flag_tree_vectorize && !loop->force_vectorize)
379 return false;
380
381 if (loop->dont_vectorize)
382 return false;
383
384 if (!do_while_loop_p (loop))
385 return true;
386
387 /* The vectorizer won't handle anything with multiple exits, so skip. */
388 edge exit = single_exit (loop);
389 if (!exit)
390 return false;
391
392 /* Copy headers iff there looks to be code in the loop after the exit block,
393 i.e. the exit block has an edge to another block (besides the latch,
394 which should be empty). */
395 edge_iterator ei;
396 edge e;
397 FOR_EACH_EDGE (e, ei, exit->src->succs)
398 if (!loop_exit_edge_p (loop, e)
399 && e->dest != loop->header
400 && e->dest != loop->latch)
401 return true;
402
403 return false;
404 }
405
406 } // anon namespace
407
408 gimple_opt_pass *
409 make_pass_ch_vect (gcc::context *ctxt)
410 {
411 return new pass_ch_vect (ctxt);
412 }
413
414 gimple_opt_pass *
415 make_pass_ch (gcc::context *ctxt)
416 {
417 return new pass_ch (ctxt);
418 }