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>.
4
5 This file is part of GCC.
6
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.
11
12GCC is distributed in the hope that it will be useful,
13but WITHOUT ANY WARRANTY; without even the implied warranty of
14MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15GNU General Public License for more details.
16
17You should have received a copy of the GNU General Public License
18along with GCC; see the file COPYING3. If not see
19<http://www.gnu.org/licenses/>. */
20
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"
38
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. */
43
44static basic_block
45find_block_to_duplicate_for_splitting_paths (basic_block latch)
46{
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))
53 {
54 basic_block bb = get_immediate_dominator (CDI_DOMINATORS, latch);
55 gcc_assert (single_pred_edge (latch)->src == bb);
56
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;
61
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;
66
67 /* We're hoping that BB is a join point for an IF-THEN-ELSE diamond
68 region. Verify that it is.
69
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)
73 {
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;
80
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;
88
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;
96
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.
100
101 We're going to want to create two duplicates of BB, one for
102 each successor of BB_IDOM. */
103 return bb;
104 }
105 }
106 return NULL;
107}
108
109/* Return the number of non-debug statements in a block. */
110static unsigned int
111count_stmts_in_block (basic_block bb)
112{
113 gimple_stmt_iterator gsi;
114 unsigned int num_stmts = 0;
115
116 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
117 {
118 gimple *stmt = gsi_stmt (gsi);
119 if (!is_gimple_debug (stmt))
120 num_stmts++;
121 }
122 return num_stmts;
123}
124
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. */
128static bool
129poor_ifcvt_candidate_code (enum tree_code code)
130{
131 return (code == MIN_EXPR
132 || code == MAX_EXPR
133 || code == ABS_EXPR
134 || code == COND_EXPR
135 || code == CALL_EXPR);
136}
137
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.
140
141 Things to consider:
142
143 We do not want to spoil if-conversion if at all possible.
144
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.
148
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. */
152
153static bool
154is_feasible_trace (basic_block bb)
155{
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;
163
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)
169 {
170 gimple *stmt1 = last_and_only_stmt (pred1);
171 gimple *stmt2 = last_and_only_stmt (pred2);
172
173 if (stmt1 && stmt2
174 && gimple_code (stmt1) == GIMPLE_ASSIGN
175 && gimple_code (stmt2) == GIMPLE_ASSIGN)
176 {
177 enum tree_code code1 = gimple_assign_rhs_code (stmt1);
178 enum tree_code code2 = gimple_assign_rhs_code (stmt2);
179
180 if (!poor_ifcvt_candidate_code (code1)
181 && !poor_ifcvt_candidate_code (code2))
182 {
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))
187 {
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))
193 {
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;
200 }
201 }
202 }
203 }
204 }
205
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))
211 {
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))
216 {
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)
223 {
224 found_useful_phi = true;
225 break;
226 }
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)))
234 {
235 bool found_unchanged_path = false;
236 for (unsigned i = 0; i < gimple_phi_num_args (phi); ++i)
237 if (gimple_phi_arg_def (phi, i) == gimple_phi_result (stmt))
238 {
239 found_unchanged_path = true;
240 break;
241 }
242 /* If we found an unchanged path this can only be a threading
243 opportunity if we have uses of the loop header PHI result
244 in a stmt dominating the merge block. Otherwise the
245 splitting may prevent if-conversion. */
246 if (found_unchanged_path)
247 {
248 use_operand_p use2_p;
249 imm_use_iterator iter2;
250 FOR_EACH_IMM_USE_FAST (use2_p, iter2, gimple_phi_result (stmt))
251 {
252 gimple *use_stmt = USE_STMT (use2_p);
253 if (is_gimple_debug (use_stmt))
254 continue;
255 basic_block use_bb = gimple_bb (use_stmt);
256 if (use_bb != bb
257 && dominated_by_p (CDI_DOMINATORS, bb, use_bb))
258 {
259 if (gcond *cond = dyn_cast <gcond *> (use_stmt))
260 if (gimple_cond_code (cond) == EQ_EXPR
261 || gimple_cond_code (cond) == NE_EXPR)
262 found_useful_phi = true;
263 break;
264 }
265 }
266 }
267 if (found_useful_phi)
268 break;
269 }
270 }
271 if (found_useful_phi)
272 break;
273 }
274 /* There is one exception namely a controlling condition we can propagate
275 an equivalence from to the joiner. */
276 bool found_cprop_opportunity = false;
277 basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb);
278 gcond *cond = as_a <gcond *> (last_stmt (dom));
279 if (gimple_cond_code (cond) == EQ_EXPR
280 || gimple_cond_code (cond) == NE_EXPR)
281 for (unsigned i = 0; i < 2; ++i)
282 {
283 tree op = gimple_op (cond, i);
284 if (TREE_CODE (op) == SSA_NAME)
285 {
286 use_operand_p use_p;
287 imm_use_iterator iter;
288 FOR_EACH_IMM_USE_FAST (use_p, iter, op)
289 {
290 if (is_gimple_debug (USE_STMT (use_p)))
291 continue;
292 if (gimple_bb (USE_STMT (use_p)) == bb)
293 {
294 found_cprop_opportunity = true;
295 break;
296 }
297 }
298 }
299 if (found_cprop_opportunity)
300 break;
301 }
302
303 if (! found_useful_phi && ! found_cprop_opportunity)
304 {
305 if (dump_file && (dump_flags & TDF_DETAILS))
306 fprintf (dump_file,
307 "Block %d is a join that does not expose CSE/DCE/jump-thread "
308 "opportunities when duplicated.\n",
309 bb->index);
310 return false;
311 }
312
313 /* We may want something here which looks at dataflow and tries
314 to guess if duplication of BB is likely to result in simplification
315 of instructions in BB in either the original or the duplicate. */
316
317 /* Upper Hard limit on the number statements to copy. */
318 if (num_stmts_in_join
319 >= PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS))
320 return false;
321
322 return true;
323}
324
325/* If the immediate dominator of the latch of the loop is
326 block with conditional branch, then the loop latch is
327 duplicated to its predecessors path preserving the SSA
328 semantics.
329
330 CFG before transformation.
331
332 2
333 |
334 |
335 +---->3
336 | / \
337 | / \
338 | 4 5
339 | \ /
340 | \ /
341 | 6
342 | / \
343 | / \
344 | 8 7
345 | | |
346 ---+ E
347
348
349
350 Block 8 is the latch. We're going to make copies of block 6 (9 & 10)
351 and wire things up so they look like this:
352
353 2
354 |
355 |
356 +---->3
357 | / \
358 | / \
359 | 4 5
360 | | |
361 | | |
362 | 9 10
363 | |\ /|
364 | | \ / |
365 | | 7 |
366 | | | |
367 | | E |
368 | | |
369 | \ /
370 | \ /
371 +-----8
372
373
374 Blocks 9 and 10 will get merged into blocks 4 & 5 respectively which
375 enables CSE, DCE and other optimizations to occur on a larger block
376 of code. */
377
378static bool
379split_paths ()
380{
381 bool changed = false;
382 loop_p loop;
383
384 loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
385 initialize_original_copy_tables ();
386 calculate_dominance_info (CDI_DOMINATORS);
387
388 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
389 {
390 /* Only split paths if we are optimizing this loop for speed. */
391 if (!optimize_loop_for_speed_p (loop))
392 continue;
393
394 /* See if there is a block that we can duplicate to split the
395 path to the loop latch. */
396 basic_block bb
397 = find_block_to_duplicate_for_splitting_paths (loop->latch);
398
399 /* BB is the merge point for an IF-THEN-ELSE we want to transform.
400
401 Essentially we want to create a duplicate of bb and redirect the
402 first predecessor of BB to the duplicate (leaving the second
403 predecessor as is. This will split the path leading to the latch
404 re-using BB to avoid useless copying. */
405 if (bb && is_feasible_trace (bb))
406 {
407 if (dump_file && (dump_flags & TDF_DETAILS))
408 fprintf (dump_file,
409 "Duplicating join block %d into predecessor paths\n",
410 bb->index);
411 basic_block pred0 = EDGE_PRED (bb, 0)->src;
412 if (EDGE_COUNT (pred0->succs) != 1)
413 pred0 = EDGE_PRED (bb, 1)->src;
414 transform_duplicate (pred0, bb);
415 changed = true;
416
417 /* If BB has an outgoing edge marked as IRREDUCIBLE, then
418 duplicating BB may result in an irreducible region turning
419 into a natural loop.
420
421 Long term we might want to hook this into the block
422 duplication code, but as we've seen with similar changes
423 for edge removal, that can be somewhat risky. */
424 if (EDGE_SUCC (bb, 0)->flags & EDGE_IRREDUCIBLE_LOOP
425 || EDGE_SUCC (bb, 1)->flags & EDGE_IRREDUCIBLE_LOOP)
426 {
427 if (dump_file && (dump_flags & TDF_DETAILS))
428 fprintf (dump_file,
429 "Join block %d has EDGE_IRREDUCIBLE_LOOP set. "
430 "Scheduling loop fixups.\n",
431 bb->index);
432 loops_state_set (LOOPS_NEED_FIXUP);
433 }
434 }
435 }
436
437 loop_optimizer_finalize ();
438 free_original_copy_tables ();
439 return changed;
440}
441
442/* Main entry point for splitting paths. Returns TODO_cleanup_cfg if any
443 paths where split, otherwise return zero. */
444
445static unsigned int
446execute_split_paths ()
447{
448 /* If we don't have at least 2 real blocks and backedges in the
449 CFG, then there's no point in trying to perform path splitting. */
450 if (n_basic_blocks_for_fn (cfun) <= NUM_FIXED_BLOCKS + 1
451 || !mark_dfs_back_edges ())
452 return 0;
453
454 bool changed = split_paths();
455 if (changed)
456 free_dominance_info (CDI_DOMINATORS);
457
458 return changed ? TODO_cleanup_cfg : 0;
459}
460
461static bool
462gate_split_paths ()
463{
464 return flag_split_paths;
465}
466
467namespace {
468
469const pass_data pass_data_split_paths =
470{
471 GIMPLE_PASS, /* type */
472 "split-paths", /* name */
473 OPTGROUP_NONE, /* optinfo_flags */
474 TV_SPLIT_PATHS, /* tv_id */
475 PROP_ssa, /* properties_required */
476 0, /* properties_provided */
477 0, /* properties_destroyed */
478 0, /* todo_flags_start */
479 TODO_update_ssa, /* todo_flags_finish */
480};
481
482class pass_split_paths : public gimple_opt_pass
483{
484 public:
485 pass_split_paths (gcc::context *ctxt)
486 : gimple_opt_pass (pass_data_split_paths, ctxt)
487 {}
488 /* opt_pass methods: */
489 opt_pass * clone () { return new pass_split_paths (m_ctxt); }
490 virtual bool gate (function *) { return gate_split_paths (); }
491 virtual unsigned int execute (function *) { return execute_split_paths (); }
492
493}; // class pass_split_paths
494
495} // anon namespace
496
497gimple_opt_pass *
498make_pass_split_paths (gcc::context *ctxt)
499{
500 return new pass_split_paths (ctxt);
501}
502