PR c++/67273
[official-gcc.git] / gcc / gimple-ssa-split-paths.c
blobf1bf7ec604c0d0ce41cc329ab36c7267040c2e04
1 /* Support routines for Splitting Paths to loop backedges
2 Copyright (C) 2015-2017 Free Software Foundation, Inc.
3 Contributed by Ajit Kumar Agarwal <ajitkum@xilinx.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
10 any later version.
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "backend.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "tree-pass.h"
28 #include "tree-cfg.h"
29 #include "cfganal.h"
30 #include "cfgloop.h"
31 #include "gimple-iterator.h"
32 #include "tracer.h"
33 #include "predict.h"
34 #include "params.h"
35 #include "gimple-ssa.h"
36 #include "tree-phinodes.h"
37 #include "ssa-iterators.h"
39 /* Given LATCH, the latch block in a loop, see if the shape of the
40 path reaching LATCH is suitable for being split by duplication.
41 If so, return the block that will be duplicated into its predecessor
42 paths. Else return NULL. */
44 static basic_block
45 find_block_to_duplicate_for_splitting_paths (basic_block latch)
47 /* We should have simple latches at this point. So the latch should
48 have a single successor. This implies the predecessor of the latch
49 likely has the loop exit. And it's that predecessor we're most
50 interested in. To keep things simple, we're going to require that
51 the latch have a single predecessor too. */
52 if (single_succ_p (latch) && single_pred_p (latch))
54 basic_block bb = get_immediate_dominator (CDI_DOMINATORS, latch);
55 gcc_assert (single_pred_edge (latch)->src == bb);
57 /* If BB has been marked as not to be duplicated, then honor that
58 request. */
59 if (ignore_bb_p (bb))
60 return NULL;
62 gimple *last = gsi_stmt (gsi_last_nondebug_bb (bb));
63 /* The immediate dominator of the latch must end in a conditional. */
64 if (!last || gimple_code (last) != GIMPLE_COND)
65 return NULL;
67 /* We're hoping that BB is a join point for an IF-THEN-ELSE diamond
68 region. Verify that it is.
70 First, verify that BB has two predecessors (each arm of the
71 IF-THEN-ELSE) and two successors (the latch and exit). */
72 if (EDGE_COUNT (bb->preds) == 2 && EDGE_COUNT (bb->succs) == 2)
74 /* Now verify that BB's immediate dominator ends in a
75 conditional as well. */
76 basic_block bb_idom = get_immediate_dominator (CDI_DOMINATORS, bb);
77 gimple *last = gsi_stmt (gsi_last_nondebug_bb (bb_idom));
78 if (!last || gimple_code (last) != GIMPLE_COND)
79 return NULL;
81 /* And that BB's immediate dominator's successors are the
82 predecessors of BB or BB itself. */
83 if (!(EDGE_PRED (bb, 0)->src == bb_idom
84 || find_edge (bb_idom, EDGE_PRED (bb, 0)->src))
85 || !(EDGE_PRED (bb, 1)->src == bb_idom
86 || find_edge (bb_idom, EDGE_PRED (bb, 1)->src)))
87 return NULL;
89 /* And that the predecessors of BB each have a single successor
90 or are BB's immediate domiator itself. */
91 if (!(EDGE_PRED (bb, 0)->src == bb_idom
92 || single_succ_p (EDGE_PRED (bb, 0)->src))
93 || !(EDGE_PRED (bb, 1)->src == bb_idom
94 || single_succ_p (EDGE_PRED (bb, 1)->src)))
95 return NULL;
97 /* So at this point we have a simple diamond for an IF-THEN-ELSE
98 construct starting at BB_IDOM, with a join point at BB. BB
99 pass control outside the loop or to the loop latch.
101 We're going to want to create two duplicates of BB, one for
102 each successor of BB_IDOM. */
103 return bb;
106 return NULL;
109 /* Return the number of non-debug statements in a block. */
110 static unsigned int
111 count_stmts_in_block (basic_block bb)
113 gimple_stmt_iterator gsi;
114 unsigned int num_stmts = 0;
116 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
118 gimple *stmt = gsi_stmt (gsi);
119 if (!is_gimple_debug (stmt))
120 num_stmts++;
122 return num_stmts;
125 /* Return TRUE if CODE represents a tree code that is not likely to
126 be easily if-convertable because it likely expands into multiple
127 insns, FALSE otherwise. */
128 static bool
129 poor_ifcvt_candidate_code (enum tree_code code)
131 return (code == MIN_EXPR
132 || code == MAX_EXPR
133 || code == ABS_EXPR
134 || code == COND_EXPR
135 || code == CALL_EXPR);
138 /* Return TRUE if BB is a reasonable block to duplicate by examining
139 its size, false otherwise. BB will always be a loop latch block.
141 Things to consider:
143 We do not want to spoil if-conversion if at all possible.
145 Most of the benefit seems to be from eliminating the unconditional
146 jump rather than CSE/DCE opportunities. So favor duplicating
147 small latches. A latch with just a conditional branch is ideal.
149 CSE/DCE opportunties crop up when statements from the predecessors
150 feed statements in the latch and allow statements in the latch to
151 simplify. */
153 static bool
154 is_feasible_trace (basic_block bb)
156 basic_block pred1 = EDGE_PRED (bb, 0)->src;
157 basic_block pred2 = EDGE_PRED (bb, 1)->src;
158 int num_stmts_in_join = count_stmts_in_block (bb);
159 int num_stmts_in_pred1
160 = EDGE_COUNT (pred1->succs) == 1 ? count_stmts_in_block (pred1) : 0;
161 int num_stmts_in_pred2
162 = EDGE_COUNT (pred2->succs) == 1 ? count_stmts_in_block (pred2) : 0;
164 /* This is meant to catch cases that are likely opportunities for
165 if-conversion. Essentially we look for the case where
166 BB's predecessors are both single statement blocks where
167 the output of that statement feed the same PHI in BB. */
168 if (num_stmts_in_pred1 == 1 && num_stmts_in_pred2 == 1)
170 gimple *stmt1 = last_and_only_stmt (pred1);
171 gimple *stmt2 = last_and_only_stmt (pred2);
173 if (stmt1 && stmt2
174 && gimple_code (stmt1) == GIMPLE_ASSIGN
175 && gimple_code (stmt2) == GIMPLE_ASSIGN)
177 enum tree_code code1 = gimple_assign_rhs_code (stmt1);
178 enum tree_code code2 = gimple_assign_rhs_code (stmt2);
180 if (!poor_ifcvt_candidate_code (code1)
181 && !poor_ifcvt_candidate_code (code2))
183 tree lhs1 = gimple_assign_lhs (stmt1);
184 tree lhs2 = gimple_assign_lhs (stmt2);
185 gimple_stmt_iterator gsi;
186 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
188 gimple *phi = gsi_stmt (gsi);
189 if ((gimple_phi_arg_def (phi, 0) == lhs1
190 && gimple_phi_arg_def (phi, 1) == lhs2)
191 || (gimple_phi_arg_def (phi, 1) == lhs1
192 && gimple_phi_arg_def (phi, 0) == lhs2))
194 if (dump_file && (dump_flags & TDF_DETAILS))
195 fprintf (dump_file,
196 "Block %d appears to be a join point for "
197 "if-convertable diamond.\n",
198 bb->index);
199 return false;
206 /* If the joiner has no PHIs with useful uses there is zero chance
207 of CSE/DCE/jump-threading possibilities exposed by duplicating it. */
208 bool found_useful_phi = false;
209 for (gphi_iterator si = gsi_start_phis (bb); ! gsi_end_p (si);
210 gsi_next (&si))
212 gphi *phi = si.phi ();
213 use_operand_p use_p;
214 imm_use_iterator iter;
215 FOR_EACH_IMM_USE_FAST (use_p, iter, gimple_phi_result (phi))
217 gimple *stmt = USE_STMT (use_p);
218 if (is_gimple_debug (stmt))
219 continue;
220 /* If there's a use in the joiner this might be a CSE/DCE
221 opportunity. */
222 if (gimple_bb (stmt) == bb)
224 found_useful_phi = true;
225 break;
227 /* If the use is on a loop header PHI and on one path the
228 value is unchanged this might expose a jump threading
229 opportunity. */
230 if (gimple_code (stmt) == GIMPLE_PHI
231 && gimple_bb (stmt) == bb->loop_father->header
232 /* But for memory the PHI alone isn't good enough. */
233 && ! virtual_operand_p (gimple_phi_result (stmt)))
235 for (unsigned i = 0; i < gimple_phi_num_args (phi); ++i)
236 if (gimple_phi_arg_def (phi, i) == gimple_phi_result (stmt))
238 found_useful_phi = true;
239 break;
241 if (found_useful_phi)
242 break;
245 if (found_useful_phi)
246 break;
248 if (! found_useful_phi)
250 if (dump_file && (dump_flags & TDF_DETAILS))
251 fprintf (dump_file,
252 "Block %d is a join that does not expose CSE/DCE/jump-thread "
253 "opportunities when duplicated.\n",
254 bb->index);
255 return false;
258 /* We may want something here which looks at dataflow and tries
259 to guess if duplication of BB is likely to result in simplification
260 of instructions in BB in either the original or the duplicate. */
262 /* Upper Hard limit on the number statements to copy. */
263 if (num_stmts_in_join
264 >= PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS))
265 return false;
267 return true;
270 /* If the immediate dominator of the latch of the loop is
271 block with conditional branch, then the loop latch is
272 duplicated to its predecessors path preserving the SSA
273 semantics.
275 CFG before transformation.
280 +---->3
281 | / \
282 | / \
283 | 4 5
284 | \ /
285 | \ /
287 | / \
288 | / \
289 | 8 7
290 | | |
291 ---+ E
295 Block 8 is the latch. We're going to make copies of block 6 (9 & 10)
296 and wire things up so they look like this:
301 +---->3
302 | / \
303 | / \
304 | 4 5
305 | | |
306 | | |
307 | 9 10
308 | |\ /|
309 | | \ / |
310 | | 7 |
311 | | | |
312 | | E |
313 | | |
314 | \ /
315 | \ /
316 +-----8
319 Blocks 9 and 10 will get merged into blocks 4 & 5 respectively which
320 enables CSE, DCE and other optimizations to occur on a larger block
321 of code. */
323 static bool
324 split_paths ()
326 bool changed = false;
327 loop_p loop;
329 loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
330 initialize_original_copy_tables ();
331 calculate_dominance_info (CDI_DOMINATORS);
333 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
335 /* Only split paths if we are optimizing this loop for speed. */
336 if (!optimize_loop_for_speed_p (loop))
337 continue;
339 /* See if there is a block that we can duplicate to split the
340 path to the loop latch. */
341 basic_block bb
342 = find_block_to_duplicate_for_splitting_paths (loop->latch);
344 /* BB is the merge point for an IF-THEN-ELSE we want to transform.
346 Essentially we want to create a duplicate of bb and redirect the
347 first predecessor of BB to the duplicate (leaving the second
348 predecessor as is. This will split the path leading to the latch
349 re-using BB to avoid useless copying. */
350 if (bb && is_feasible_trace (bb))
352 if (dump_file && (dump_flags & TDF_DETAILS))
353 fprintf (dump_file,
354 "Duplicating join block %d into predecessor paths\n",
355 bb->index);
356 basic_block pred0 = EDGE_PRED (bb, 0)->src;
357 if (EDGE_COUNT (pred0->succs) != 1)
358 pred0 = EDGE_PRED (bb, 1)->src;
359 transform_duplicate (pred0, bb);
360 changed = true;
362 /* If BB has an outgoing edge marked as IRREDUCIBLE, then
363 duplicating BB may result in an irreducible region turning
364 into a natural loop.
366 Long term we might want to hook this into the block
367 duplication code, but as we've seen with similar changes
368 for edge removal, that can be somewhat risky. */
369 if (EDGE_SUCC (bb, 0)->flags & EDGE_IRREDUCIBLE_LOOP
370 || EDGE_SUCC (bb, 1)->flags & EDGE_IRREDUCIBLE_LOOP)
372 if (dump_file && (dump_flags & TDF_DETAILS))
373 fprintf (dump_file,
374 "Join block %d has EDGE_IRREDUCIBLE_LOOP set. "
375 "Scheduling loop fixups.\n",
376 bb->index);
377 loops_state_set (LOOPS_NEED_FIXUP);
382 loop_optimizer_finalize ();
383 free_original_copy_tables ();
384 return changed;
387 /* Main entry point for splitting paths. Returns TODO_cleanup_cfg if any
388 paths where split, otherwise return zero. */
390 static unsigned int
391 execute_split_paths ()
393 /* If we don't have at least 2 real blocks and backedges in the
394 CFG, then there's no point in trying to perform path splitting. */
395 if (n_basic_blocks_for_fn (cfun) <= NUM_FIXED_BLOCKS + 1
396 || !mark_dfs_back_edges ())
397 return 0;
399 bool changed = split_paths();
400 if (changed)
401 free_dominance_info (CDI_DOMINATORS);
403 return changed ? TODO_cleanup_cfg : 0;
406 static bool
407 gate_split_paths ()
409 return flag_split_paths;
412 namespace {
414 const pass_data pass_data_split_paths =
416 GIMPLE_PASS, /* type */
417 "split-paths", /* name */
418 OPTGROUP_NONE, /* optinfo_flags */
419 TV_SPLIT_PATHS, /* tv_id */
420 PROP_ssa, /* properties_required */
421 0, /* properties_provided */
422 0, /* properties_destroyed */
423 0, /* todo_flags_start */
424 TODO_update_ssa, /* todo_flags_finish */
427 class pass_split_paths : public gimple_opt_pass
429 public:
430 pass_split_paths (gcc::context *ctxt)
431 : gimple_opt_pass (pass_data_split_paths, ctxt)
433 /* opt_pass methods: */
434 opt_pass * clone () { return new pass_split_paths (m_ctxt); }
435 virtual bool gate (function *) { return gate_split_paths (); }
436 virtual unsigned int execute (function *) { return execute_split_paths (); }
438 }; // class pass_split_paths
440 } // anon namespace
442 gimple_opt_pass *
443 make_pass_split_paths (gcc::context *ctxt)
445 return new pass_split_paths (ctxt);