111
|
1 /* Back-propagation of usage information to definitions.
|
131
|
2 Copyright (C) 2015-2018 Free Software Foundation, Inc.
|
111
|
3
|
|
4 This file is part of GCC.
|
|
5
|
|
6 GCC is free software; you can redistribute it and/or modify
|
|
7 it under the terms of the GNU General Public License as published by
|
|
8 the Free Software Foundation; either version 3, or (at your option)
|
|
9 any later version.
|
|
10
|
|
11 GCC is distributed in the hope that it will be useful,
|
|
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
14 GNU General Public License 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 /* This pass propagates information that is common to all uses of an SSA
|
|
21 name back up through the sequence of statements that generate it,
|
|
22 simplifying the statements where possible. Sometimes this can expose
|
|
23 fully or partially dead code, but the main focus is simplifying
|
|
24 computations.
|
|
25
|
|
26 At the moment the pass only handles one piece of information: whether the
|
|
27 sign of a value matters, and therefore whether sign-changing operations
|
|
28 can be skipped. The pass could be extended to more interesting
|
|
29 information in future, such as which bits of an integer are significant.
|
|
30
|
|
31 For example, take the function:
|
|
32
|
|
33 double
|
|
34 f (double *a, int n, double start)
|
|
35 {
|
|
36 double x = fabs (start);
|
|
37 for (int i = 0; i < n; ++i)
|
|
38 x *= a[i];
|
|
39 return __builtin_cos (x);
|
|
40 }
|
|
41
|
|
42 cos(x) == cos(-x), so the sign of the final x doesn't matter.
|
|
43 That x is the result of a series of multiplications, and if
|
|
44 the sign of the result of a multiplication doesn't matter,
|
|
45 the signs of the inputs don't matter either.
|
|
46
|
|
47 The pass would replace the incoming value of x (i.e. fabs(start))
|
|
48 with start. Since there are no other uses of the fabs result,
|
|
49 the call would get deleted as dead.
|
|
50
|
|
51 The algorithm is:
|
|
52
|
|
53 (1) Do a post-order traversal of the blocks in the function, walking
|
|
54 each block backwards. For each potentially-simplifiable statement
|
|
55 that defines an SSA name X, examine all uses of X to see what
|
|
56 information is actually significant. Record this as INFO_MAP[X].
|
|
57 Optimistically ignore for now any back-edge references to
|
|
58 unprocessed phis.
|
|
59
|
|
60 (An alternative would be to record each use when we visit its
|
|
61 statement and take the intersection as we go along. However,
|
|
62 this would lead to more SSA names being entered into INFO_MAP
|
|
63 unnecessarily, only to be taken out again later. At the moment
|
|
64 very few SSA names end up with useful information.)
|
|
65
|
|
66 (2) Iteratively reduce the optimistic result of (1) until we reach
|
|
67 a maximal fixed point (which at the moment would mean revisiting
|
|
68 statements at most once). First push all SSA names that used an
|
|
69 optimistic assumption about a backedge phi onto a worklist.
|
|
70 While the worklist is nonempty, pick off an SSA name X and recompute
|
|
71 INFO_MAP[X]. If the value changes, push all SSA names used in the
|
|
72 definition of X onto the worklist.
|
|
73
|
|
74 (3) Iterate over each SSA name X with info in INFO_MAP, in the
|
|
75 opposite order to (1), i.e. a forward reverse-post-order walk.
|
|
76 Try to optimize the definition of X using INFO_MAP[X] and fold
|
|
77 the result. (This ensures that we fold definitions before uses.)
|
|
78
|
|
79 (4) Iterate over each SSA name X with info in INFO_MAP, in the same
|
|
80 order as (1), and delete any statements that are now dead.
|
|
81 (This ensures that if a sequence of statements is dead,
|
|
82 we delete the last statement first.)
|
|
83
|
|
84 Note that this pass does not deal with direct redundancies,
|
|
85 such as cos(-x)->cos(x). match.pd handles those cases instead. */
|
|
86
|
|
87 #include "config.h"
|
|
88 #include "system.h"
|
|
89 #include "coretypes.h"
|
|
90 #include "backend.h"
|
|
91 #include "tree.h"
|
|
92 #include "gimple.h"
|
|
93 #include "gimple-iterator.h"
|
|
94 #include "ssa.h"
|
|
95 #include "fold-const.h"
|
|
96 #include "tree-pass.h"
|
|
97 #include "cfganal.h"
|
|
98 #include "gimple-pretty-print.h"
|
|
99 #include "tree-cfg.h"
|
|
100 #include "tree-ssa.h"
|
|
101 #include "tree-ssa-propagate.h"
|
|
102 #include "gimple-fold.h"
|
|
103 #include "alloc-pool.h"
|
|
104 #include "tree-hash-traits.h"
|
|
105 #include "case-cfn-macros.h"
|
|
106
|
|
107 namespace {
|
|
108
|
|
109 /* Information about a group of uses of an SSA name. */
|
|
110 struct usage_info
|
|
111 {
|
|
112 usage_info () : flag_word (0) {}
|
|
113 usage_info &operator &= (const usage_info &);
|
|
114 usage_info operator & (const usage_info &) const;
|
|
115 bool operator == (const usage_info &) const;
|
|
116 bool operator != (const usage_info &) const;
|
|
117 bool is_useful () const;
|
|
118
|
|
119 static usage_info intersection_identity ();
|
|
120
|
|
121 union
|
|
122 {
|
|
123 struct
|
|
124 {
|
|
125 /* True if the uses treat x and -x in the same way. */
|
|
126 unsigned int ignore_sign : 1;
|
|
127 } flags;
|
|
128 /* All the flag bits as a single int. */
|
|
129 unsigned int flag_word;
|
|
130 };
|
|
131 };
|
|
132
|
|
133 /* Return an X such that X & Y == Y for all Y. This is the most
|
|
134 optimistic assumption possible. */
|
|
135
|
|
136 usage_info
|
|
137 usage_info::intersection_identity ()
|
|
138 {
|
|
139 usage_info ret;
|
|
140 ret.flag_word = -1;
|
|
141 return ret;
|
|
142 }
|
|
143
|
|
144 /* Intersect *THIS with OTHER, so that *THIS describes all uses covered
|
|
145 by the original *THIS and OTHER. */
|
|
146
|
|
147 usage_info &
|
|
148 usage_info::operator &= (const usage_info &other)
|
|
149 {
|
|
150 flag_word &= other.flag_word;
|
|
151 return *this;
|
|
152 }
|
|
153
|
|
154 /* Return the intersection of *THIS and OTHER, i.e. a structure that
|
|
155 describes all uses covered by *THIS and OTHER. */
|
|
156
|
|
157 usage_info
|
|
158 usage_info::operator & (const usage_info &other) const
|
|
159 {
|
|
160 usage_info info (*this);
|
|
161 info &= other;
|
|
162 return info;
|
|
163 }
|
|
164
|
|
165 bool
|
|
166 usage_info::operator == (const usage_info &other) const
|
|
167 {
|
|
168 return flag_word == other.flag_word;
|
|
169 }
|
|
170
|
|
171 bool
|
|
172 usage_info::operator != (const usage_info &other) const
|
|
173 {
|
|
174 return !operator == (other);
|
|
175 }
|
|
176
|
|
177 /* Return true if *THIS is not simply the default, safe assumption. */
|
|
178
|
|
179 bool
|
|
180 usage_info::is_useful () const
|
|
181 {
|
|
182 return flag_word != 0;
|
|
183 }
|
|
184
|
|
185 /* Start a dump line about SSA name VAR. */
|
|
186
|
|
187 static void
|
|
188 dump_usage_prefix (FILE *file, tree var)
|
|
189 {
|
|
190 fprintf (file, " ");
|
|
191 print_generic_expr (file, var);
|
|
192 fprintf (file, ": ");
|
|
193 }
|
|
194
|
|
195 /* Print INFO to FILE. */
|
|
196
|
|
197 static void
|
|
198 dump_usage_info (FILE *file, tree var, usage_info *info)
|
|
199 {
|
|
200 if (info->flags.ignore_sign)
|
|
201 {
|
|
202 dump_usage_prefix (file, var);
|
|
203 fprintf (file, "sign bit not important\n");
|
|
204 }
|
|
205 }
|
|
206
|
|
207 /* Represents one execution of the pass. */
|
|
208 class backprop
|
|
209 {
|
|
210 public:
|
|
211 backprop (function *);
|
|
212 ~backprop ();
|
|
213
|
|
214 void execute ();
|
|
215
|
|
216 private:
|
|
217 const usage_info *lookup_operand (tree);
|
|
218
|
|
219 void push_to_worklist (tree);
|
|
220 tree pop_from_worklist ();
|
|
221
|
|
222 void process_builtin_call_use (gcall *, tree, usage_info *);
|
|
223 void process_assign_use (gassign *, tree, usage_info *);
|
|
224 void process_phi_use (gphi *, usage_info *);
|
|
225 void process_use (gimple *, tree, usage_info *);
|
|
226 bool intersect_uses (tree, usage_info *);
|
|
227 void reprocess_inputs (gimple *);
|
|
228 void process_var (tree);
|
|
229 void process_block (basic_block);
|
|
230
|
|
231 void prepare_change (tree);
|
|
232 void complete_change (gimple *);
|
|
233 void optimize_builtin_call (gcall *, tree, const usage_info *);
|
|
234 void replace_assign_rhs (gassign *, tree, tree, tree, tree);
|
|
235 void optimize_assign (gassign *, tree, const usage_info *);
|
|
236 void optimize_phi (gphi *, tree, const usage_info *);
|
|
237
|
|
238 typedef hash_map <tree_ssa_name_hash, usage_info *> info_map_type;
|
|
239 typedef std::pair <tree, usage_info *> var_info_pair;
|
|
240
|
|
241 /* The function we're optimizing. */
|
|
242 function *m_fn;
|
|
243
|
|
244 /* Pool for allocating usage_info structures. */
|
|
245 object_allocator <usage_info> m_info_pool;
|
|
246
|
|
247 /* Maps an SSA name to a description of all uses of that SSA name.
|
|
248 All the usage_infos satisfy is_useful.
|
|
249
|
|
250 We use a hash_map because the map is expected to be sparse
|
|
251 (i.e. most SSA names won't have useful information attached to them).
|
|
252 We could move to a directly-indexed array if that situation changes. */
|
|
253 info_map_type m_info_map;
|
|
254
|
|
255 /* Post-ordered list of all potentially-interesting SSA names,
|
|
256 along with information that describes all uses. */
|
|
257 auto_vec <var_info_pair, 128> m_vars;
|
|
258
|
|
259 /* A bitmap of blocks that we have finished processing in the initial
|
|
260 post-order walk. */
|
|
261 auto_sbitmap m_visited_blocks;
|
|
262
|
131
|
263 /* A bitmap of phis that we have finished processing in the initial
|
|
264 post-order walk, excluding those from blocks mentioned in
|
|
265 M_VISITED_BLOCKS. */
|
|
266 auto_bitmap m_visited_phis;
|
|
267
|
111
|
268 /* A worklist of SSA names whose definitions need to be reconsidered. */
|
|
269 auto_vec <tree, 64> m_worklist;
|
|
270
|
|
271 /* The SSA names in M_WORKLIST, identified by their SSA_NAME_VERSION.
|
|
272 We use a bitmap rather than an sbitmap because most SSA names are
|
|
273 never added to the worklist. */
|
|
274 bitmap m_worklist_names;
|
|
275 };
|
|
276
|
|
277 backprop::backprop (function *fn)
|
|
278 : m_fn (fn),
|
|
279 m_info_pool ("usage_info"),
|
|
280 m_visited_blocks (last_basic_block_for_fn (m_fn)),
|
|
281 m_worklist_names (BITMAP_ALLOC (NULL))
|
|
282 {
|
|
283 bitmap_clear (m_visited_blocks);
|
|
284 }
|
|
285
|
|
286 backprop::~backprop ()
|
|
287 {
|
|
288 BITMAP_FREE (m_worklist_names);
|
|
289 m_info_pool.release ();
|
|
290 }
|
|
291
|
|
292 /* Return usage information for general operand OP, or null if none. */
|
|
293
|
|
294 const usage_info *
|
|
295 backprop::lookup_operand (tree op)
|
|
296 {
|
|
297 if (op && TREE_CODE (op) == SSA_NAME)
|
|
298 {
|
|
299 usage_info **slot = m_info_map.get (op);
|
|
300 if (slot)
|
|
301 return *slot;
|
|
302 }
|
|
303 return NULL;
|
|
304 }
|
|
305
|
|
306 /* Add SSA name VAR to the worklist, if it isn't on the worklist already. */
|
|
307
|
|
308 void
|
|
309 backprop::push_to_worklist (tree var)
|
|
310 {
|
|
311 if (!bitmap_set_bit (m_worklist_names, SSA_NAME_VERSION (var)))
|
|
312 return;
|
|
313 m_worklist.safe_push (var);
|
|
314 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
315 {
|
|
316 fprintf (dump_file, "[WORKLIST] Pushing ");
|
|
317 print_generic_expr (dump_file, var);
|
|
318 fprintf (dump_file, "\n");
|
|
319 }
|
|
320 }
|
|
321
|
|
322 /* Remove and return the next SSA name from the worklist. The worklist
|
|
323 is known to be nonempty. */
|
|
324
|
|
325 tree
|
|
326 backprop::pop_from_worklist ()
|
|
327 {
|
|
328 tree var = m_worklist.pop ();
|
|
329 bitmap_clear_bit (m_worklist_names, SSA_NAME_VERSION (var));
|
|
330 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
331 {
|
|
332 fprintf (dump_file, "[WORKLIST] Popping ");
|
|
333 print_generic_expr (dump_file, var);
|
|
334 fprintf (dump_file, "\n");
|
|
335 }
|
|
336 return var;
|
|
337 }
|
|
338
|
|
339 /* Make INFO describe all uses of RHS in CALL, which is a call to a
|
|
340 built-in function. */
|
|
341
|
|
342 void
|
|
343 backprop::process_builtin_call_use (gcall *call, tree rhs, usage_info *info)
|
|
344 {
|
|
345 combined_fn fn = gimple_call_combined_fn (call);
|
|
346 tree lhs = gimple_call_lhs (call);
|
|
347 switch (fn)
|
|
348 {
|
|
349 case CFN_LAST:
|
|
350 break;
|
|
351
|
|
352 CASE_CFN_COS:
|
|
353 CASE_CFN_COSH:
|
|
354 CASE_CFN_CCOS:
|
|
355 CASE_CFN_CCOSH:
|
|
356 CASE_CFN_HYPOT:
|
|
357 /* The signs of all inputs are ignored. */
|
|
358 info->flags.ignore_sign = true;
|
|
359 break;
|
|
360
|
|
361 CASE_CFN_COPYSIGN:
|
131
|
362 CASE_CFN_COPYSIGN_FN:
|
111
|
363 /* The sign of the first input is ignored. */
|
|
364 if (rhs != gimple_call_arg (call, 1))
|
|
365 info->flags.ignore_sign = true;
|
|
366 break;
|
|
367
|
|
368 CASE_CFN_POW:
|
|
369 {
|
|
370 /* The sign of the first input is ignored as long as the second
|
|
371 input is an even real. */
|
|
372 tree power = gimple_call_arg (call, 1);
|
|
373 HOST_WIDE_INT n;
|
|
374 if (TREE_CODE (power) == REAL_CST
|
|
375 && real_isinteger (&TREE_REAL_CST (power), &n)
|
|
376 && (n & 1) == 0)
|
|
377 info->flags.ignore_sign = true;
|
|
378 break;
|
|
379 }
|
|
380
|
|
381 CASE_CFN_FMA:
|
131
|
382 CASE_CFN_FMA_FN:
|
|
383 case CFN_FMS:
|
|
384 case CFN_FNMA:
|
|
385 case CFN_FNMS:
|
111
|
386 /* In X * X + Y, where Y is distinct from X, the sign of X doesn't
|
|
387 matter. */
|
|
388 if (gimple_call_arg (call, 0) == rhs
|
|
389 && gimple_call_arg (call, 1) == rhs
|
|
390 && gimple_call_arg (call, 2) != rhs)
|
|
391 info->flags.ignore_sign = true;
|
|
392 break;
|
|
393
|
|
394 default:
|
|
395 if (negate_mathfn_p (fn))
|
|
396 {
|
|
397 /* The sign of the (single) input doesn't matter provided
|
|
398 that the sign of the output doesn't matter. */
|
|
399 const usage_info *lhs_info = lookup_operand (lhs);
|
|
400 if (lhs_info)
|
|
401 info->flags.ignore_sign = lhs_info->flags.ignore_sign;
|
|
402 }
|
|
403 break;
|
|
404 }
|
|
405 }
|
|
406
|
|
407 /* Make INFO describe all uses of RHS in ASSIGN. */
|
|
408
|
|
409 void
|
|
410 backprop::process_assign_use (gassign *assign, tree rhs, usage_info *info)
|
|
411 {
|
|
412 tree lhs = gimple_assign_lhs (assign);
|
|
413 switch (gimple_assign_rhs_code (assign))
|
|
414 {
|
|
415 case ABS_EXPR:
|
131
|
416 case ABSU_EXPR:
|
111
|
417 /* The sign of the input doesn't matter. */
|
|
418 info->flags.ignore_sign = true;
|
|
419 break;
|
|
420
|
|
421 case COND_EXPR:
|
|
422 /* For A = B ? C : D, propagate information about all uses of A
|
|
423 to C and D. */
|
|
424 if (rhs != gimple_assign_rhs1 (assign))
|
|
425 {
|
|
426 const usage_info *lhs_info = lookup_operand (lhs);
|
|
427 if (lhs_info)
|
|
428 *info = *lhs_info;
|
|
429 }
|
|
430 break;
|
|
431
|
|
432 case MULT_EXPR:
|
|
433 /* In X * X, the sign of X doesn't matter. */
|
|
434 if (gimple_assign_rhs1 (assign) == rhs
|
|
435 && gimple_assign_rhs2 (assign) == rhs)
|
|
436 info->flags.ignore_sign = true;
|
|
437 /* Fall through. */
|
|
438
|
|
439 case NEGATE_EXPR:
|
|
440 case RDIV_EXPR:
|
|
441 /* If the sign of the result doesn't matter, the sign of the inputs
|
|
442 doesn't matter either. */
|
|
443 if (FLOAT_TYPE_P (TREE_TYPE (rhs)))
|
|
444 {
|
|
445 const usage_info *lhs_info = lookup_operand (lhs);
|
|
446 if (lhs_info)
|
|
447 info->flags.ignore_sign = lhs_info->flags.ignore_sign;
|
|
448 }
|
|
449 break;
|
|
450
|
|
451 default:
|
|
452 break;
|
|
453 }
|
|
454 }
|
|
455
|
|
456 /* Make INFO describe the uses of PHI's result. */
|
|
457
|
|
458 void
|
|
459 backprop::process_phi_use (gphi *phi, usage_info *info)
|
|
460 {
|
|
461 tree result = gimple_phi_result (phi);
|
|
462 if (const usage_info *result_info = lookup_operand (result))
|
|
463 *info = *result_info;
|
|
464 }
|
|
465
|
|
466 /* Make INFO describe all uses of RHS in STMT. */
|
|
467
|
|
468 void
|
|
469 backprop::process_use (gimple *stmt, tree rhs, usage_info *info)
|
|
470 {
|
|
471 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
472 {
|
|
473 fprintf (dump_file, "[USE] ");
|
|
474 print_generic_expr (dump_file, rhs);
|
|
475 fprintf (dump_file, " in ");
|
|
476 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
|
|
477 }
|
|
478
|
|
479 if (gcall *call = dyn_cast <gcall *> (stmt))
|
|
480 process_builtin_call_use (call, rhs, info);
|
|
481 else if (gassign *assign = dyn_cast <gassign *> (stmt))
|
|
482 process_assign_use (assign, rhs, info);
|
|
483 else if (gphi *phi = dyn_cast <gphi *> (stmt))
|
|
484 process_phi_use (phi, info);
|
|
485
|
|
486 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
487 dump_usage_info (dump_file, rhs, info);
|
|
488 }
|
|
489
|
|
490 /* Make INFO describe all uses of VAR, returning true if the result
|
|
491 is useful. If the uses include phis that haven't been processed yet,
|
|
492 make the most optimistic assumption possible, so that we aim for
|
|
493 a maximum rather than a minimum fixed point. */
|
|
494
|
|
495 bool
|
|
496 backprop::intersect_uses (tree var, usage_info *info)
|
|
497 {
|
|
498 imm_use_iterator iter;
|
131
|
499 use_operand_p use_p;
|
111
|
500 *info = usage_info::intersection_identity ();
|
131
|
501 FOR_EACH_IMM_USE_FAST (use_p, iter, var)
|
111
|
502 {
|
131
|
503 gimple *stmt = USE_STMT (use_p);
|
111
|
504 if (is_gimple_debug (stmt))
|
|
505 continue;
|
131
|
506 gphi *phi = dyn_cast <gphi *> (stmt);
|
|
507 if (phi
|
|
508 && !bitmap_bit_p (m_visited_blocks, gimple_bb (phi)->index)
|
|
509 && !bitmap_bit_p (m_visited_phis,
|
|
510 SSA_NAME_VERSION (gimple_phi_result (phi))))
|
111
|
511 {
|
|
512 /* Skip unprocessed phis. */
|
|
513 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
514 {
|
|
515 fprintf (dump_file, "[BACKEDGE] ");
|
|
516 print_generic_expr (dump_file, var);
|
|
517 fprintf (dump_file, " in ");
|
131
|
518 print_gimple_stmt (dump_file, phi, 0, TDF_SLIM);
|
111
|
519 }
|
|
520 }
|
|
521 else
|
|
522 {
|
|
523 usage_info subinfo;
|
|
524 process_use (stmt, var, &subinfo);
|
|
525 *info &= subinfo;
|
|
526 if (!info->is_useful ())
|
131
|
527 return false;
|
111
|
528 }
|
|
529 }
|
|
530 return true;
|
|
531 }
|
|
532
|
|
533 /* Queue for reconsideration any input of STMT that has information
|
|
534 associated with it. This is used if that information might be
|
|
535 too optimistic. */
|
|
536
|
|
537 void
|
|
538 backprop::reprocess_inputs (gimple *stmt)
|
|
539 {
|
|
540 use_operand_p use_p;
|
|
541 ssa_op_iter oi;
|
|
542 FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, oi, SSA_OP_USE)
|
|
543 {
|
|
544 tree var = get_use_from_ptr (use_p);
|
|
545 if (lookup_operand (var))
|
|
546 push_to_worklist (var);
|
|
547 }
|
|
548 }
|
|
549
|
|
550 /* Say that we're recording INFO for SSA name VAR, or that we're deleting
|
|
551 existing information if INFO is null. INTRO describes the change. */
|
|
552
|
|
553 static void
|
|
554 dump_var_info (tree var, usage_info *info, const char *intro)
|
|
555 {
|
|
556 fprintf (dump_file, "[DEF] %s for ", intro);
|
|
557 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (var), 0, TDF_SLIM);
|
|
558 if (info)
|
|
559 dump_usage_info (dump_file, var, info);
|
|
560 }
|
|
561
|
|
562 /* Process all uses of VAR and record or update the result in
|
|
563 M_INFO_MAP and M_VARS. */
|
|
564
|
|
565 void
|
|
566 backprop::process_var (tree var)
|
|
567 {
|
|
568 if (has_zero_uses (var))
|
|
569 return;
|
|
570
|
|
571 usage_info info;
|
|
572 intersect_uses (var, &info);
|
|
573
|
|
574 gimple *stmt = SSA_NAME_DEF_STMT (var);
|
|
575 if (info.is_useful ())
|
|
576 {
|
|
577 bool existed;
|
|
578 usage_info *&map_info = m_info_map.get_or_insert (var, &existed);
|
|
579 if (!existed)
|
|
580 {
|
|
581 /* Recording information about VAR for the first time. */
|
|
582 map_info = m_info_pool.allocate ();
|
|
583 *map_info = info;
|
|
584 m_vars.safe_push (var_info_pair (var, map_info));
|
|
585 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
586 dump_var_info (var, map_info, "Recording new information");
|
|
587
|
|
588 /* If STMT is a phi, reprocess any backedge uses. This is a
|
|
589 no-op for other uses, which won't have any information
|
|
590 associated with them. */
|
|
591 if (is_a <gphi *> (stmt))
|
|
592 reprocess_inputs (stmt);
|
|
593 }
|
|
594 else if (info != *map_info)
|
|
595 {
|
|
596 /* Recording information that is less optimistic than before. */
|
|
597 gcc_checking_assert ((info & *map_info) == info);
|
|
598 *map_info = info;
|
|
599 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
600 dump_var_info (var, map_info, "Updating information");
|
|
601 reprocess_inputs (stmt);
|
|
602 }
|
|
603 }
|
|
604 else
|
|
605 {
|
|
606 if (usage_info **slot = m_info_map.get (var))
|
|
607 {
|
|
608 /* Removing previously-recorded information. */
|
|
609 **slot = info;
|
|
610 m_info_map.remove (var);
|
|
611 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
612 dump_var_info (var, NULL, "Deleting information");
|
|
613 reprocess_inputs (stmt);
|
|
614 }
|
|
615 else
|
|
616 {
|
|
617 /* If STMT is a phi, remove any information recorded for
|
|
618 its arguments. */
|
|
619 if (is_a <gphi *> (stmt))
|
|
620 reprocess_inputs (stmt);
|
|
621 }
|
|
622 }
|
|
623 }
|
|
624
|
|
625 /* Process all statements and phis in BB, during the first post-order walk. */
|
|
626
|
|
627 void
|
|
628 backprop::process_block (basic_block bb)
|
|
629 {
|
|
630 for (gimple_stmt_iterator gsi = gsi_last_bb (bb); !gsi_end_p (gsi);
|
|
631 gsi_prev (&gsi))
|
|
632 {
|
|
633 tree lhs = gimple_get_lhs (gsi_stmt (gsi));
|
|
634 if (lhs && TREE_CODE (lhs) == SSA_NAME)
|
|
635 process_var (lhs);
|
|
636 }
|
|
637 for (gphi_iterator gpi = gsi_start_phis (bb); !gsi_end_p (gpi);
|
|
638 gsi_next (&gpi))
|
131
|
639 {
|
|
640 tree result = gimple_phi_result (gpi.phi ());
|
|
641 process_var (result);
|
|
642 bitmap_set_bit (m_visited_phis, SSA_NAME_VERSION (result));
|
|
643 }
|
|
644 bitmap_clear (m_visited_phis);
|
111
|
645 }
|
|
646
|
|
647 /* Delete the definition of VAR, which has no uses. */
|
|
648
|
|
649 static void
|
|
650 remove_unused_var (tree var)
|
|
651 {
|
|
652 gimple *stmt = SSA_NAME_DEF_STMT (var);
|
|
653 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
654 {
|
|
655 fprintf (dump_file, "Deleting ");
|
|
656 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
|
|
657 }
|
|
658 gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
|
|
659 gsi_remove (&gsi, true);
|
|
660 release_defs (stmt);
|
|
661 }
|
|
662
|
|
663 /* Note that we're replacing OLD_RHS with NEW_RHS in STMT. */
|
|
664
|
|
665 static void
|
|
666 note_replacement (gimple *stmt, tree old_rhs, tree new_rhs)
|
|
667 {
|
|
668 fprintf (dump_file, "Replacing use of ");
|
|
669 print_generic_expr (dump_file, old_rhs);
|
|
670 fprintf (dump_file, " with ");
|
|
671 print_generic_expr (dump_file, new_rhs);
|
|
672 fprintf (dump_file, " in ");
|
|
673 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
|
|
674 }
|
|
675
|
|
676 /* If RHS is an SSA name whose definition just changes the sign of a value,
|
|
677 return that other value, otherwise return null. */
|
|
678
|
|
679 static tree
|
|
680 strip_sign_op_1 (tree rhs)
|
|
681 {
|
|
682 if (TREE_CODE (rhs) != SSA_NAME)
|
|
683 return NULL_TREE;
|
|
684
|
|
685 gimple *def_stmt = SSA_NAME_DEF_STMT (rhs);
|
|
686 if (gassign *assign = dyn_cast <gassign *> (def_stmt))
|
|
687 switch (gimple_assign_rhs_code (assign))
|
|
688 {
|
|
689 case ABS_EXPR:
|
131
|
690 case ABSU_EXPR:
|
111
|
691 case NEGATE_EXPR:
|
|
692 return gimple_assign_rhs1 (assign);
|
|
693
|
|
694 default:
|
|
695 break;
|
|
696 }
|
|
697 else if (gcall *call = dyn_cast <gcall *> (def_stmt))
|
|
698 switch (gimple_call_combined_fn (call))
|
|
699 {
|
|
700 CASE_CFN_COPYSIGN:
|
131
|
701 CASE_CFN_COPYSIGN_FN:
|
111
|
702 return gimple_call_arg (call, 0);
|
|
703
|
|
704 default:
|
|
705 break;
|
|
706 }
|
|
707
|
|
708 return NULL_TREE;
|
|
709 }
|
|
710
|
|
711 /* If RHS is an SSA name whose definition just changes the sign of a value,
|
|
712 strip all such operations and return the ultimate input to them.
|
|
713 Return null otherwise.
|
|
714
|
|
715 Although this could in principle lead to quadratic searching,
|
|
716 in practice a long sequence of sign manipulations should already
|
|
717 have been folded down. E.g. --x -> x, abs(-x) -> abs(x). We search
|
|
718 for more than one operation in order to catch cases like -abs(x). */
|
|
719
|
|
720 static tree
|
|
721 strip_sign_op (tree rhs)
|
|
722 {
|
|
723 tree new_rhs = strip_sign_op_1 (rhs);
|
|
724 if (!new_rhs)
|
|
725 return NULL_TREE;
|
|
726 while (tree next = strip_sign_op_1 (new_rhs))
|
|
727 new_rhs = next;
|
|
728 return new_rhs;
|
|
729 }
|
|
730
|
|
731 /* Start a change in the value of VAR that is suitable for all non-debug
|
|
732 uses of VAR. We need to make sure that debug statements continue to
|
|
733 use the original definition of VAR where possible, or are nullified
|
|
734 otherwise. */
|
|
735
|
|
736 void
|
|
737 backprop::prepare_change (tree var)
|
|
738 {
|
131
|
739 if (MAY_HAVE_DEBUG_BIND_STMTS)
|
111
|
740 insert_debug_temp_for_var_def (NULL, var);
|
|
741 reset_flow_sensitive_info (var);
|
|
742 }
|
|
743
|
|
744 /* STMT has been changed. Give the fold machinery a chance to simplify
|
|
745 and canonicalize it (e.g. by ensuring that commutative operands have
|
|
746 the right order), then record the updates. */
|
|
747
|
|
748 void
|
|
749 backprop::complete_change (gimple *stmt)
|
|
750 {
|
|
751 gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
|
|
752 if (fold_stmt (&gsi))
|
|
753 {
|
|
754 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
755 {
|
|
756 fprintf (dump_file, " which folds to: ");
|
|
757 print_gimple_stmt (dump_file, gsi_stmt (gsi), 0, TDF_SLIM);
|
|
758 }
|
|
759 }
|
|
760 update_stmt (gsi_stmt (gsi));
|
|
761 }
|
|
762
|
|
763 /* Optimize CALL, a call to a built-in function with lhs LHS, on the
|
|
764 basis that INFO describes all uses of LHS. */
|
|
765
|
|
766 void
|
|
767 backprop::optimize_builtin_call (gcall *call, tree lhs, const usage_info *info)
|
|
768 {
|
|
769 /* If we have an f such that -f(x) = f(-x), and if the sign of the result
|
|
770 doesn't matter, strip any sign operations from the input. */
|
|
771 if (info->flags.ignore_sign
|
|
772 && negate_mathfn_p (gimple_call_combined_fn (call)))
|
|
773 {
|
|
774 tree new_arg = strip_sign_op (gimple_call_arg (call, 0));
|
|
775 if (new_arg)
|
|
776 {
|
|
777 prepare_change (lhs);
|
|
778 gimple_call_set_arg (call, 0, new_arg);
|
|
779 complete_change (call);
|
|
780 }
|
|
781 }
|
|
782 }
|
|
783
|
|
784 /* Optimize ASSIGN, an assignment to LHS, by replacing rhs operand N
|
|
785 with RHS<N>, if RHS<N> is nonnull. This may change the value of LHS. */
|
|
786
|
|
787 void
|
|
788 backprop::replace_assign_rhs (gassign *assign, tree lhs, tree rhs1,
|
|
789 tree rhs2, tree rhs3)
|
|
790 {
|
|
791 if (!rhs1 && !rhs2 && !rhs3)
|
|
792 return;
|
|
793
|
|
794 prepare_change (lhs);
|
|
795 if (rhs1)
|
|
796 gimple_assign_set_rhs1 (assign, rhs1);
|
|
797 if (rhs2)
|
|
798 gimple_assign_set_rhs2 (assign, rhs2);
|
|
799 if (rhs3)
|
|
800 gimple_assign_set_rhs3 (assign, rhs3);
|
|
801 complete_change (assign);
|
|
802 }
|
|
803
|
|
804 /* Optimize ASSIGN, an assignment to LHS, on the basis that INFO
|
|
805 describes all uses of LHS. */
|
|
806
|
|
807 void
|
|
808 backprop::optimize_assign (gassign *assign, tree lhs, const usage_info *info)
|
|
809 {
|
|
810 switch (gimple_assign_rhs_code (assign))
|
|
811 {
|
|
812 case MULT_EXPR:
|
|
813 case RDIV_EXPR:
|
|
814 /* If the sign of the result doesn't matter, strip sign operations
|
|
815 from both inputs. */
|
|
816 if (info->flags.ignore_sign)
|
|
817 replace_assign_rhs (assign, lhs,
|
|
818 strip_sign_op (gimple_assign_rhs1 (assign)),
|
|
819 strip_sign_op (gimple_assign_rhs2 (assign)),
|
|
820 NULL_TREE);
|
|
821 break;
|
|
822
|
|
823 case COND_EXPR:
|
|
824 /* If the sign of A ? B : C doesn't matter, strip sign operations
|
|
825 from both B and C. */
|
|
826 if (info->flags.ignore_sign)
|
|
827 replace_assign_rhs (assign, lhs,
|
|
828 NULL_TREE,
|
|
829 strip_sign_op (gimple_assign_rhs2 (assign)),
|
|
830 strip_sign_op (gimple_assign_rhs3 (assign)));
|
|
831 break;
|
|
832
|
|
833 default:
|
|
834 break;
|
|
835 }
|
|
836 }
|
|
837
|
|
838 /* Optimize PHI, which defines VAR, on the basis that INFO describes all
|
|
839 uses of the result. */
|
|
840
|
|
841 void
|
|
842 backprop::optimize_phi (gphi *phi, tree var, const usage_info *info)
|
|
843 {
|
|
844 /* If the sign of the result doesn't matter, try to strip sign operations
|
|
845 from arguments. */
|
|
846 if (info->flags.ignore_sign)
|
|
847 {
|
|
848 basic_block bb = gimple_bb (phi);
|
|
849 use_operand_p use;
|
|
850 ssa_op_iter oi;
|
|
851 bool replaced = false;
|
|
852 FOR_EACH_PHI_ARG (use, phi, oi, SSA_OP_USE)
|
|
853 {
|
|
854 /* Propagating along abnormal edges is delicate, punt for now. */
|
|
855 const int index = PHI_ARG_INDEX_FROM_USE (use);
|
|
856 if (EDGE_PRED (bb, index)->flags & EDGE_ABNORMAL)
|
|
857 continue;
|
|
858
|
|
859 tree new_arg = strip_sign_op (USE_FROM_PTR (use));
|
|
860 if (new_arg)
|
|
861 {
|
|
862 if (!replaced)
|
|
863 prepare_change (var);
|
|
864 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
865 note_replacement (phi, USE_FROM_PTR (use), new_arg);
|
|
866 replace_exp (use, new_arg);
|
|
867 replaced = true;
|
|
868 }
|
|
869 }
|
|
870 }
|
|
871 }
|
|
872
|
|
873 void
|
|
874 backprop::execute ()
|
|
875 {
|
|
876 /* Phase 1: Traverse the function, making optimistic assumptions
|
|
877 about any phi whose definition we haven't seen. */
|
|
878 int *postorder = XNEWVEC (int, n_basic_blocks_for_fn (m_fn));
|
|
879 unsigned int postorder_num = post_order_compute (postorder, false, false);
|
|
880 for (unsigned int i = 0; i < postorder_num; ++i)
|
|
881 {
|
|
882 process_block (BASIC_BLOCK_FOR_FN (m_fn, postorder[i]));
|
|
883 bitmap_set_bit (m_visited_blocks, postorder[i]);
|
|
884 }
|
|
885 XDELETEVEC (postorder);
|
|
886
|
|
887 /* Phase 2: Use the initial (perhaps overly optimistic) information
|
|
888 to create a maximal fixed point solution. */
|
|
889 while (!m_worklist.is_empty ())
|
|
890 process_var (pop_from_worklist ());
|
|
891
|
|
892 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
893 fprintf (dump_file, "\n");
|
|
894
|
|
895 /* Phase 3: Do a reverse post-order walk, using information about
|
|
896 the uses of SSA names to optimize their definitions. */
|
|
897 for (unsigned int i = m_vars.length (); i-- > 0;)
|
|
898 {
|
|
899 usage_info *info = m_vars[i].second;
|
|
900 if (info->is_useful ())
|
|
901 {
|
|
902 tree var = m_vars[i].first;
|
|
903 gimple *stmt = SSA_NAME_DEF_STMT (var);
|
|
904 if (gcall *call = dyn_cast <gcall *> (stmt))
|
|
905 optimize_builtin_call (call, var, info);
|
|
906 else if (gassign *assign = dyn_cast <gassign *> (stmt))
|
|
907 optimize_assign (assign, var, info);
|
|
908 else if (gphi *phi = dyn_cast <gphi *> (stmt))
|
|
909 optimize_phi (phi, var, info);
|
|
910 }
|
|
911 }
|
|
912
|
|
913 /* Phase 4: Do a post-order walk, deleting statements that are no
|
|
914 longer needed. */
|
|
915 for (unsigned int i = 0; i < m_vars.length (); ++i)
|
|
916 {
|
|
917 tree var = m_vars[i].first;
|
|
918 if (has_zero_uses (var))
|
|
919 remove_unused_var (var);
|
|
920 }
|
|
921
|
|
922 if (dump_file && (dump_flags & TDF_DETAILS))
|
|
923 fprintf (dump_file, "\n");
|
|
924 }
|
|
925
|
|
926 const pass_data pass_data_backprop =
|
|
927 {
|
|
928 GIMPLE_PASS, /* type */
|
|
929 "backprop", /* name */
|
|
930 OPTGROUP_NONE, /* optinfo_flags */
|
|
931 TV_TREE_BACKPROP, /* tv_id */
|
|
932 ( PROP_cfg | PROP_ssa ), /* properties_required */
|
|
933 0, /* properties_provided */
|
|
934 0, /* properties_destroyed */
|
|
935 0, /* todo_flags_start */
|
|
936 0, /* todo_flags_finish */
|
|
937 };
|
|
938
|
|
939 class pass_backprop : public gimple_opt_pass
|
|
940 {
|
|
941 public:
|
|
942 pass_backprop (gcc::context *ctxt)
|
|
943 : gimple_opt_pass (pass_data_backprop, ctxt)
|
|
944 {}
|
|
945
|
|
946 /* opt_pass methods: */
|
|
947 opt_pass * clone () { return new pass_backprop (m_ctxt); }
|
|
948 virtual bool gate (function *) { return flag_ssa_backprop; }
|
|
949 virtual unsigned int execute (function *);
|
|
950
|
|
951 }; // class pass_backprop
|
|
952
|
|
953 unsigned int
|
|
954 pass_backprop::execute (function *fn)
|
|
955 {
|
|
956 backprop (fn).execute ();
|
|
957 return 0;
|
|
958 }
|
|
959
|
|
960 } // anon namespace
|
|
961
|
|
962 gimple_opt_pass *
|
|
963 make_pass_backprop (gcc::context *ctxt)
|
|
964 {
|
|
965 return new pass_backprop (ctxt);
|
|
966 }
|