Fix memory order description in atomic ops built-ins docs.
[gcc.git] / gcc / tree-ssa-loop-ch.c
1 /* Loop header copying on trees.
2 Copyright (C) 2004-2015 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 "tm.h"
24 #include "hash-set.h"
25 #include "machmode.h"
26 #include "vec.h"
27 #include "double-int.h"
28 #include "input.h"
29 #include "alias.h"
30 #include "symtab.h"
31 #include "wide-int.h"
32 #include "inchash.h"
33 #include "tree.h"
34 #include "fold-const.h"
35 #include "tm_p.h"
36 #include "predict.h"
37 #include "hard-reg-set.h"
38 #include "input.h"
39 #include "function.h"
40 #include "dominance.h"
41 #include "cfg.h"
42 #include "basic-block.h"
43 #include "tree-ssa-alias.h"
44 #include "internal-fn.h"
45 #include "gimple-expr.h"
46 #include "is-a.h"
47 #include "gimple.h"
48 #include "gimple-iterator.h"
49 #include "gimple-ssa.h"
50 #include "tree-cfg.h"
51 #include "tree-into-ssa.h"
52 #include "tree-pass.h"
53 #include "cfgloop.h"
54 #include "tree-inline.h"
55 #include "flags.h"
56 #include "tree-ssa-scopedtables.h"
57 #include "tree-ssa-threadedge.h"
58
59 /* Duplicates headers of loops if they are small enough, so that the statements
60 in the loop body are always executed when the loop is entered. This
61 increases effectiveness of code motion optimizations, and reduces the need
62 for loop preconditioning. */
63
64 /* Check whether we should duplicate HEADER of LOOP. At most *LIMIT
65 instructions should be duplicated, limit is decreased by the actual
66 amount. */
67
68 static bool
69 should_duplicate_loop_header_p (basic_block header, struct loop *loop,
70 int *limit)
71 {
72 gimple_stmt_iterator bsi;
73 gimple last;
74
75 /* Do not copy one block more than once (we do not really want to do
76 loop peeling here). */
77 if (header->aux)
78 return false;
79
80 /* Loop header copying usually increases size of the code. This used not to
81 be true, since quite often it is possible to verify that the condition is
82 satisfied in the first iteration and therefore to eliminate it. Jump
83 threading handles these cases now. */
84 if (optimize_loop_for_size_p (loop))
85 return false;
86
87 gcc_assert (EDGE_COUNT (header->succs) > 0);
88 if (single_succ_p (header))
89 return false;
90 if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest)
91 && flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 1)->dest))
92 return false;
93
94 /* If this is not the original loop header, we want it to have just
95 one predecessor in order to match the && pattern. */
96 if (header != loop->header && !single_pred_p (header))
97 return false;
98
99 last = last_stmt (header);
100 if (gimple_code (last) != GIMPLE_COND)
101 return false;
102
103 /* Approximately copy the conditions that used to be used in jump.c --
104 at most 20 insns and no calls. */
105 for (bsi = gsi_start_bb (header); !gsi_end_p (bsi); gsi_next (&bsi))
106 {
107 last = gsi_stmt (bsi);
108
109 if (gimple_code (last) == GIMPLE_LABEL)
110 continue;
111
112 if (is_gimple_debug (last))
113 continue;
114
115 if (is_gimple_call (last))
116 return false;
117
118 *limit -= estimate_num_insns (last, &eni_size_weights);
119 if (*limit < 0)
120 return false;
121 }
122
123 return true;
124 }
125
126 /* Checks whether LOOP is a do-while style loop. */
127
128 static bool
129 do_while_loop_p (struct loop *loop)
130 {
131 gimple stmt = last_stmt (loop->latch);
132
133 /* If the latch of the loop is not empty, it is not a do-while loop. */
134 if (stmt
135 && gimple_code (stmt) != GIMPLE_LABEL)
136 return false;
137
138 /* If the header contains just a condition, it is not a do-while loop. */
139 stmt = last_and_only_stmt (loop->header);
140 if (stmt
141 && gimple_code (stmt) == GIMPLE_COND)
142 return false;
143
144 return true;
145 }
146
147 /* For all loops, copy the condition at the end of the loop body in front
148 of the loop. This is beneficial since it increases efficiency of
149 code motion optimizations. It also saves one jump on entry to the loop. */
150
151 namespace {
152
153 const pass_data pass_data_ch =
154 {
155 GIMPLE_PASS, /* type */
156 "ch", /* name */
157 OPTGROUP_LOOP, /* optinfo_flags */
158 TV_TREE_CH, /* tv_id */
159 ( PROP_cfg | PROP_ssa ), /* properties_required */
160 0, /* properties_provided */
161 0, /* properties_destroyed */
162 0, /* todo_flags_start */
163 0, /* todo_flags_finish */
164 };
165
166 class pass_ch : public gimple_opt_pass
167 {
168 public:
169 pass_ch (gcc::context *ctxt)
170 : gimple_opt_pass (pass_data_ch, ctxt)
171 {}
172
173 /* opt_pass methods: */
174 virtual bool gate (function *) { return flag_tree_ch != 0; }
175 virtual unsigned int execute (function *);
176
177 }; // class pass_ch
178
179 unsigned int
180 pass_ch::execute (function *fun)
181 {
182 struct loop *loop;
183 basic_block header;
184 edge exit, entry;
185 basic_block *bbs, *copied_bbs;
186 unsigned n_bbs;
187 unsigned bbs_size;
188 bool changed = false;
189
190 loop_optimizer_init (LOOPS_HAVE_PREHEADERS
191 | LOOPS_HAVE_SIMPLE_LATCHES);
192 if (number_of_loops (fun) <= 1)
193 {
194 loop_optimizer_finalize ();
195 return 0;
196 }
197
198 bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (fun));
199 copied_bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (fun));
200 bbs_size = n_basic_blocks_for_fn (fun);
201
202 FOR_EACH_LOOP (loop, 0)
203 {
204 /* Copy at most 20 insns. */
205 int limit = 20;
206
207 header = loop->header;
208
209 /* If the loop is already a do-while style one (either because it was
210 written as such, or because jump threading transformed it into one),
211 we might be in fact peeling the first iteration of the loop. This
212 in general is not a good idea. */
213 if (do_while_loop_p (loop))
214 continue;
215
216 /* Iterate the header copying up to limit; this takes care of the cases
217 like while (a && b) {...}, where we want to have both of the conditions
218 copied. TODO -- handle while (a || b) - like cases, by not requiring
219 the header to have just a single successor and copying up to
220 postdominator. */
221
222 exit = NULL;
223 n_bbs = 0;
224 while (should_duplicate_loop_header_p (header, loop, &limit))
225 {
226 /* Find a successor of header that is inside a loop; i.e. the new
227 header after the condition is copied. */
228 if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest))
229 exit = EDGE_SUCC (header, 0);
230 else
231 exit = EDGE_SUCC (header, 1);
232 bbs[n_bbs++] = header;
233 gcc_assert (bbs_size > n_bbs);
234 header = exit->dest;
235 }
236
237 if (!exit)
238 continue;
239
240 if (dump_file && (dump_flags & TDF_DETAILS))
241 fprintf (dump_file,
242 "Duplicating header of the loop %d up to edge %d->%d.\n",
243 loop->num, exit->src->index, exit->dest->index);
244
245 /* Ensure that the header will have just the latch as a predecessor
246 inside the loop. */
247 if (!single_pred_p (exit->dest))
248 exit = single_pred_edge (split_edge (exit));
249
250 entry = loop_preheader_edge (loop);
251
252 propagate_threaded_block_debug_into (exit->dest, entry->dest);
253 if (!gimple_duplicate_sese_region (entry, exit, bbs, n_bbs, copied_bbs,
254 true))
255 {
256 fprintf (dump_file, "Duplication failed.\n");
257 continue;
258 }
259
260 /* If the loop has the form "for (i = j; i < j + 10; i++)" then
261 this copying can introduce a case where we rely on undefined
262 signed overflow to eliminate the preheader condition, because
263 we assume that "j < j + 10" is true. We don't want to warn
264 about that case for -Wstrict-overflow, because in general we
265 don't warn about overflow involving loops. Prevent the
266 warning by setting the no_warning flag in the condition. */
267 if (warn_strict_overflow > 0)
268 {
269 unsigned int i;
270
271 for (i = 0; i < n_bbs; ++i)
272 {
273 gimple_stmt_iterator bsi;
274
275 for (bsi = gsi_start_bb (copied_bbs[i]);
276 !gsi_end_p (bsi);
277 gsi_next (&bsi))
278 {
279 gimple stmt = gsi_stmt (bsi);
280 if (gimple_code (stmt) == GIMPLE_COND)
281 gimple_set_no_warning (stmt, true);
282 else if (is_gimple_assign (stmt))
283 {
284 enum tree_code rhs_code = gimple_assign_rhs_code (stmt);
285 if (TREE_CODE_CLASS (rhs_code) == tcc_comparison)
286 gimple_set_no_warning (stmt, true);
287 }
288 }
289 }
290 }
291
292 /* Ensure that the latch and the preheader is simple (we know that they
293 are not now, since there was the loop exit condition. */
294 split_edge (loop_preheader_edge (loop));
295 split_edge (loop_latch_edge (loop));
296
297 changed = true;
298 }
299
300 update_ssa (TODO_update_ssa);
301 free (bbs);
302 free (copied_bbs);
303
304 loop_optimizer_finalize ();
305 return changed ? TODO_cleanup_cfg : 0;
306 }
307
308 } // anon namespace
309
310 gimple_opt_pass *
311 make_pass_ch (gcc::context *ctxt)
312 {
313 return new pass_ch (ctxt);
314 }