2015-08-04 Thomas Preud'homme <thomas.preudhomme@arm.com>
[official-gcc.git] / gcc / tree-ssa-loop-ch.c
blob19d1b4cf3efdb4d023d9306016aa639367c27dea
1 /* Loop header copying on trees.
2 Copyright (C) 2004-2015 Free Software Foundation, Inc.
4 This file is part of GCC.
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.
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.
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/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "cfghooks.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "hard-reg-set.h"
28 #include "alias.h"
29 #include "fold-const.h"
30 #include "tm_p.h"
31 #include "internal-fn.h"
32 #include "gimple-iterator.h"
33 #include "gimple-ssa.h"
34 #include "tree-cfg.h"
35 #include "tree-into-ssa.h"
36 #include "tree-pass.h"
37 #include "cfgloop.h"
38 #include "tree-inline.h"
39 #include "flags.h"
40 #include "tree-ssa-scopedtables.h"
41 #include "tree-ssa-threadedge.h"
43 /* Duplicates headers of loops if they are small enough, so that the statements
44 in the loop body are always executed when the loop is entered. This
45 increases effectiveness of code motion optimizations, and reduces the need
46 for loop preconditioning. */
48 /* Check whether we should duplicate HEADER of LOOP. At most *LIMIT
49 instructions should be duplicated, limit is decreased by the actual
50 amount. */
52 static bool
53 should_duplicate_loop_header_p (basic_block header, struct loop *loop,
54 int *limit)
56 gimple_stmt_iterator bsi;
57 gimple last;
59 /* Do not copy one block more than once (we do not really want to do
60 loop peeling here). */
61 if (header->aux)
62 return false;
64 /* Loop header copying usually increases size of the code. This used not to
65 be true, since quite often it is possible to verify that the condition is
66 satisfied in the first iteration and therefore to eliminate it. Jump
67 threading handles these cases now. */
68 if (optimize_loop_for_size_p (loop))
69 return false;
71 gcc_assert (EDGE_COUNT (header->succs) > 0);
72 if (single_succ_p (header))
73 return false;
74 if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest)
75 && flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 1)->dest))
76 return false;
78 /* If this is not the original loop header, we want it to have just
79 one predecessor in order to match the && pattern. */
80 if (header != loop->header && !single_pred_p (header))
81 return false;
83 last = last_stmt (header);
84 if (gimple_code (last) != GIMPLE_COND)
85 return false;
87 /* Approximately copy the conditions that used to be used in jump.c --
88 at most 20 insns and no calls. */
89 for (bsi = gsi_start_bb (header); !gsi_end_p (bsi); gsi_next (&bsi))
91 last = gsi_stmt (bsi);
93 if (gimple_code (last) == GIMPLE_LABEL)
94 continue;
96 if (is_gimple_debug (last))
97 continue;
99 if (is_gimple_call (last))
100 return false;
102 *limit -= estimate_num_insns (last, &eni_size_weights);
103 if (*limit < 0)
104 return false;
107 return true;
110 /* Checks whether LOOP is a do-while style loop. */
112 static bool
113 do_while_loop_p (struct loop *loop)
115 gimple stmt = last_stmt (loop->latch);
117 /* If the latch of the loop is not empty, it is not a do-while loop. */
118 if (stmt
119 && gimple_code (stmt) != GIMPLE_LABEL)
120 return false;
122 /* If the header contains just a condition, it is not a do-while loop. */
123 stmt = last_and_only_stmt (loop->header);
124 if (stmt
125 && gimple_code (stmt) == GIMPLE_COND)
126 return false;
128 return true;
131 namespace {
133 /* Common superclass for both header-copying phases. */
134 class ch_base : public gimple_opt_pass
136 protected:
137 ch_base (pass_data data, gcc::context *ctxt)
138 : gimple_opt_pass (data, ctxt)
141 /* Copies headers of all loops in FUN for which process_loop_p is true. */
142 unsigned int copy_headers (function *fun);
144 /* Return true to copy headers of LOOP or false to skip. */
145 virtual bool process_loop_p (struct loop *loop) = 0;
148 const pass_data pass_data_ch =
150 GIMPLE_PASS, /* type */
151 "ch", /* name */
152 OPTGROUP_LOOP, /* optinfo_flags */
153 TV_TREE_CH, /* tv_id */
154 ( PROP_cfg | PROP_ssa ), /* properties_required */
155 0, /* properties_provided */
156 0, /* properties_destroyed */
157 0, /* todo_flags_start */
158 0, /* todo_flags_finish */
161 class pass_ch : public ch_base
163 public:
164 pass_ch (gcc::context *ctxt)
165 : ch_base (pass_data_ch, ctxt)
168 /* opt_pass methods: */
169 virtual bool gate (function *) { return flag_tree_ch != 0; }
171 /* Initialize and finalize loop structures, copying headers inbetween. */
172 virtual unsigned int execute (function *);
174 protected:
175 /* ch_base method: */
176 virtual bool process_loop_p (struct loop *loop);
177 }; // class pass_ch
179 const pass_data pass_data_ch_vect =
181 GIMPLE_PASS, /* type */
182 "ch_vect", /* name */
183 OPTGROUP_LOOP, /* optinfo_flags */
184 TV_TREE_CH, /* tv_id */
185 ( PROP_cfg | PROP_ssa ), /* properties_required */
186 0, /* properties_provided */
187 0, /* properties_destroyed */
188 0, /* todo_flags_start */
189 0, /* todo_flags_finish */
192 /* This is a more aggressive version of the same pass, designed to run just
193 before if-conversion and vectorization, to put more loops into the form
194 required for those phases. */
195 class pass_ch_vect : public ch_base
197 public:
198 pass_ch_vect (gcc::context *ctxt)
199 : ch_base (pass_data_ch_vect, ctxt)
202 /* opt_pass methods: */
203 virtual bool gate (function *fun)
205 return flag_tree_ch != 0
206 && (flag_tree_loop_vectorize != 0 || fun->has_force_vectorize_loops);
209 /* Just copy headers, no initialization/finalization of loop structures. */
210 virtual unsigned int execute (function *);
212 protected:
213 /* ch_base method: */
214 virtual bool process_loop_p (struct loop *loop);
215 }; // class pass_ch_vect
217 /* For all loops, copy the condition at the end of the loop body in front
218 of the loop. This is beneficial since it increases efficiency of
219 code motion optimizations. It also saves one jump on entry to the loop. */
221 unsigned int
222 ch_base::copy_headers (function *fun)
224 struct loop *loop;
225 basic_block header;
226 edge exit, entry;
227 basic_block *bbs, *copied_bbs;
228 unsigned n_bbs;
229 unsigned bbs_size;
230 bool changed = false;
232 if (number_of_loops (fun) <= 1)
233 return 0;
235 bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (fun));
236 copied_bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (fun));
237 bbs_size = n_basic_blocks_for_fn (fun);
239 FOR_EACH_LOOP (loop, 0)
241 /* Copy at most 20 insns. */
242 int limit = 20;
244 header = loop->header;
246 /* If the loop is already a do-while style one (either because it was
247 written as such, or because jump threading transformed it into one),
248 we might be in fact peeling the first iteration of the loop. This
249 in general is not a good idea. */
250 if (!process_loop_p (loop))
251 continue;
253 /* Iterate the header copying up to limit; this takes care of the cases
254 like while (a && b) {...}, where we want to have both of the conditions
255 copied. TODO -- handle while (a || b) - like cases, by not requiring
256 the header to have just a single successor and copying up to
257 postdominator. */
259 exit = NULL;
260 n_bbs = 0;
261 while (should_duplicate_loop_header_p (header, loop, &limit))
263 /* Find a successor of header that is inside a loop; i.e. the new
264 header after the condition is copied. */
265 if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest))
266 exit = EDGE_SUCC (header, 0);
267 else
268 exit = EDGE_SUCC (header, 1);
269 bbs[n_bbs++] = header;
270 gcc_assert (bbs_size > n_bbs);
271 header = exit->dest;
274 if (!exit)
275 continue;
277 if (dump_file && (dump_flags & TDF_DETAILS))
278 fprintf (dump_file,
279 "Duplicating header of the loop %d up to edge %d->%d.\n",
280 loop->num, exit->src->index, exit->dest->index);
282 /* Ensure that the header will have just the latch as a predecessor
283 inside the loop. */
284 if (!single_pred_p (exit->dest))
285 exit = single_pred_edge (split_edge (exit));
287 entry = loop_preheader_edge (loop);
289 propagate_threaded_block_debug_into (exit->dest, entry->dest);
290 if (!gimple_duplicate_sese_region (entry, exit, bbs, n_bbs, copied_bbs,
291 true))
293 fprintf (dump_file, "Duplication failed.\n");
294 continue;
297 /* If the loop has the form "for (i = j; i < j + 10; i++)" then
298 this copying can introduce a case where we rely on undefined
299 signed overflow to eliminate the preheader condition, because
300 we assume that "j < j + 10" is true. We don't want to warn
301 about that case for -Wstrict-overflow, because in general we
302 don't warn about overflow involving loops. Prevent the
303 warning by setting the no_warning flag in the condition. */
304 if (warn_strict_overflow > 0)
306 unsigned int i;
308 for (i = 0; i < n_bbs; ++i)
310 gimple_stmt_iterator bsi;
312 for (bsi = gsi_start_bb (copied_bbs[i]);
313 !gsi_end_p (bsi);
314 gsi_next (&bsi))
316 gimple stmt = gsi_stmt (bsi);
317 if (gimple_code (stmt) == GIMPLE_COND)
318 gimple_set_no_warning (stmt, true);
319 else if (is_gimple_assign (stmt))
321 enum tree_code rhs_code = gimple_assign_rhs_code (stmt);
322 if (TREE_CODE_CLASS (rhs_code) == tcc_comparison)
323 gimple_set_no_warning (stmt, true);
329 /* Ensure that the latch and the preheader is simple (we know that they
330 are not now, since there was the loop exit condition. */
331 split_edge (loop_preheader_edge (loop));
332 split_edge (loop_latch_edge (loop));
334 changed = true;
337 if (changed)
338 update_ssa (TODO_update_ssa);
339 free (bbs);
340 free (copied_bbs);
342 return changed ? TODO_cleanup_cfg : 0;
345 /* Initialize the loop structures we need, and finalize after. */
347 unsigned int
348 pass_ch::execute (function *fun)
350 loop_optimizer_init (LOOPS_HAVE_PREHEADERS
351 | LOOPS_HAVE_SIMPLE_LATCHES);
353 unsigned int res = copy_headers (fun);
355 loop_optimizer_finalize ();
356 return res;
359 /* Assume an earlier phase has already initialized all the loop structures that
360 we need here (and perhaps others too), and that these will be finalized by
361 a later phase. */
363 unsigned int
364 pass_ch_vect::execute (function *fun)
366 return copy_headers (fun);
369 /* Apply header copying according to a very simple test of do-while shape. */
371 bool
372 pass_ch::process_loop_p (struct loop *loop)
374 return !do_while_loop_p (loop);
377 /* Apply header-copying to loops where we might enable vectorization. */
379 bool
380 pass_ch_vect::process_loop_p (struct loop *loop)
382 if (!flag_tree_vectorize && !loop->force_vectorize)
383 return false;
385 if (loop->dont_vectorize)
386 return false;
388 if (!do_while_loop_p (loop))
389 return true;
391 /* The vectorizer won't handle anything with multiple exits, so skip. */
392 edge exit = single_exit (loop);
393 if (!exit)
394 return false;
396 /* Copy headers iff there looks to be code in the loop after the exit block,
397 i.e. the exit block has an edge to another block (besides the latch,
398 which should be empty). */
399 edge_iterator ei;
400 edge e;
401 FOR_EACH_EDGE (e, ei, exit->src->succs)
402 if (!loop_exit_edge_p (loop, e)
403 && e->dest != loop->header
404 && e->dest != loop->latch)
405 return true;
407 return false;
410 } // anon namespace
412 gimple_opt_pass *
413 make_pass_ch_vect (gcc::context *ctxt)
415 return new pass_ch_vect (ctxt);
418 gimple_opt_pass *
419 make_pass_ch (gcc::context *ctxt)
421 return new pass_ch (ctxt);