]>
Commit | Line | Data |
---|---|---|
4ee9c684 | 1 | /* Optimization of PHI nodes by converting them into straightline code. |
d353bf18 | 2 | Copyright (C) 2004-2015 Free Software Foundation, Inc. |
4ee9c684 | 3 | |
4 | This file is part of GCC. | |
20e5647c | 5 | |
4ee9c684 | 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 | |
8c4c00c1 | 8 | Free Software Foundation; either version 3, or (at your option) any |
4ee9c684 | 9 | later version. |
20e5647c | 10 | |
4ee9c684 | 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. | |
20e5647c | 15 | |
4ee9c684 | 16 | You should have received a copy of the GNU General Public License |
8c4c00c1 | 17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
4ee9c684 | 19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
9ef16211 | 23 | #include "backend.h" |
4ee9c684 | 24 | #include "tree.h" |
9ef16211 | 25 | #include "gimple.h" |
26 | #include "rtl.h" | |
27 | #include "ssa.h" | |
28 | #include "alias.h" | |
b20a8bb4 | 29 | #include "fold-const.h" |
9ed99284 | 30 | #include "stor-layout.h" |
0beac6fc | 31 | #include "flags.h" |
4ee9c684 | 32 | #include "tm_p.h" |
94ea8568 | 33 | #include "cfganal.h" |
bc61cadb | 34 | #include "internal-fn.h" |
a8783bee | 35 | #include "gimplify.h" |
dcf1a1ec | 36 | #include "gimple-iterator.h" |
e795d6e1 | 37 | #include "gimplify-me.h" |
073c1fd5 | 38 | #include "tree-cfg.h" |
d53441c8 | 39 | #include "insn-config.h" |
40 | #include "expmed.h" | |
41 | #include "dojump.h" | |
42 | #include "explow.h" | |
43 | #include "calls.h" | |
44 | #include "emit-rtl.h" | |
45 | #include "varasm.h" | |
46 | #include "stmt.h" | |
9ed99284 | 47 | #include "expr.h" |
073c1fd5 | 48 | #include "tree-dfa.h" |
4ee9c684 | 49 | #include "tree-pass.h" |
4ee9c684 | 50 | #include "langhooks.h" |
e6d0e152 | 51 | #include "domwalk.h" |
ec611e12 | 52 | #include "cfgloop.h" |
53 | #include "tree-data-ref.h" | |
239e9670 | 54 | #include "gimple-pretty-print.h" |
34517c64 | 55 | #include "insn-codes.h" |
239e9670 | 56 | #include "optabs.h" |
f6568ea4 | 57 | #include "tree-scalar-evolution.h" |
b6814ca0 | 58 | #include "tree-inline.h" |
239e9670 | 59 | |
239e9670 | 60 | static unsigned int tree_ssa_phiopt_worker (bool, bool); |
a4844041 | 61 | static bool conditional_replacement (basic_block, basic_block, |
1a91d914 | 62 | edge, edge, gphi *, tree, tree); |
fb9912ea | 63 | static int value_replacement (basic_block, basic_block, |
64 | edge, edge, gimple, tree, tree); | |
a4844041 | 65 | static bool minmax_replacement (basic_block, basic_block, |
75a70cf9 | 66 | edge, edge, gimple, tree, tree); |
a4844041 | 67 | static bool abs_replacement (basic_block, basic_block, |
75a70cf9 | 68 | edge, edge, gimple, tree, tree); |
e6d0e152 | 69 | static bool cond_store_replacement (basic_block, basic_block, edge, edge, |
431205b7 | 70 | hash_set<tree> *); |
91cf53d5 | 71 | static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block); |
431205b7 | 72 | static hash_set<tree> * get_non_trapping (); |
75a70cf9 | 73 | static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree); |
239e9670 | 74 | static void hoist_adjacent_loads (basic_block, basic_block, |
75 | basic_block, basic_block); | |
76 | static bool gate_hoist_loads (void); | |
902929aa | 77 | |
e6d0e152 | 78 | /* This pass tries to transform conditional stores into unconditional |
79 | ones, enabling further simplifications with the simpler then and else | |
80 | blocks. In particular it replaces this: | |
81 | ||
82 | bb0: | |
83 | if (cond) goto bb2; else goto bb1; | |
84 | bb1: | |
91cf53d5 | 85 | *p = RHS; |
e6d0e152 | 86 | bb2: |
87 | ||
88 | with | |
89 | ||
90 | bb0: | |
91 | if (cond) goto bb1; else goto bb2; | |
92 | bb1: | |
93 | condtmp' = *p; | |
94 | bb2: | |
95 | condtmp = PHI <RHS, condtmp'> | |
91cf53d5 | 96 | *p = condtmp; |
e6d0e152 | 97 | |
98 | This transformation can only be done under several constraints, | |
91cf53d5 | 99 | documented below. It also replaces: |
100 | ||
101 | bb0: | |
102 | if (cond) goto bb2; else goto bb1; | |
103 | bb1: | |
104 | *p = RHS1; | |
105 | goto bb3; | |
106 | bb2: | |
107 | *p = RHS2; | |
108 | bb3: | |
109 | ||
110 | with | |
111 | ||
112 | bb0: | |
113 | if (cond) goto bb3; else goto bb1; | |
114 | bb1: | |
115 | bb3: | |
116 | condtmp = PHI <RHS1, RHS2> | |
117 | *p = condtmp; */ | |
e6d0e152 | 118 | |
119 | static unsigned int | |
120 | tree_ssa_cs_elim (void) | |
121 | { | |
f6568ea4 | 122 | unsigned todo; |
123 | /* ??? We are not interested in loop related info, but the following | |
124 | will create it, ICEing as we didn't init loops with pre-headers. | |
125 | An interfacing issue of find_data_references_in_bb. */ | |
126 | loop_optimizer_init (LOOPS_NORMAL); | |
127 | scev_initialize (); | |
128 | todo = tree_ssa_phiopt_worker (true, false); | |
129 | scev_finalize (); | |
130 | loop_optimizer_finalize (); | |
131 | return todo; | |
e6d0e152 | 132 | } |
133 | ||
c3597b05 | 134 | /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */ |
135 | ||
1a91d914 | 136 | static gphi * |
c3597b05 | 137 | single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1) |
138 | { | |
139 | gimple_stmt_iterator i; | |
1a91d914 | 140 | gphi *phi = NULL; |
c3597b05 | 141 | if (gimple_seq_singleton_p (seq)) |
1a91d914 | 142 | return as_a <gphi *> (gsi_stmt (gsi_start (seq))); |
c3597b05 | 143 | for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i)) |
144 | { | |
1a91d914 | 145 | gphi *p = as_a <gphi *> (gsi_stmt (i)); |
c3597b05 | 146 | /* If the PHI arguments are equal then we can skip this PHI. */ |
147 | if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx), | |
148 | gimple_phi_arg_def (p, e1->dest_idx))) | |
149 | continue; | |
150 | ||
151 | /* If we already have a PHI that has the two edge arguments are | |
152 | different, then return it is not a singleton for these PHIs. */ | |
153 | if (phi) | |
154 | return NULL; | |
155 | ||
156 | phi = p; | |
157 | } | |
158 | return phi; | |
159 | } | |
160 | ||
e6d0e152 | 161 | /* The core routine of conditional store replacement and normal |
162 | phi optimizations. Both share much of the infrastructure in how | |
163 | to match applicable basic block patterns. DO_STORE_ELIM is true | |
239e9670 | 164 | when we want to do conditional store replacement, false otherwise. |
f32420fb | 165 | DO_HOIST_LOADS is true when we want to hoist adjacent loads out |
239e9670 | 166 | of diamond control flow patterns, false otherwise. */ |
e6d0e152 | 167 | static unsigned int |
239e9670 | 168 | tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads) |
4ee9c684 | 169 | { |
170 | basic_block bb; | |
194899bf | 171 | basic_block *bb_order; |
172 | unsigned n, i; | |
1e4b21e3 | 173 | bool cfgchanged = false; |
431205b7 | 174 | hash_set<tree> *nontrap = 0; |
e6d0e152 | 175 | |
176 | if (do_store_elim) | |
03d37e4e | 177 | /* Calculate the set of non-trapping memory accesses. */ |
178 | nontrap = get_non_trapping (); | |
194899bf | 179 | |
180 | /* Search every basic block for COND_EXPR we may be able to optimize. | |
181 | ||
182 | We walk the blocks in order that guarantees that a block with | |
183 | a single predecessor is processed before the predecessor. | |
184 | This ensures that we collapse inner ifs before visiting the | |
185 | outer ones, and also that we do not try to visit a removed | |
186 | block. */ | |
ba4d2b2f | 187 | bb_order = single_pred_before_succ_order (); |
a28770e1 | 188 | n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS; |
4ee9c684 | 189 | |
48e1416a | 190 | for (i = 0; i < n; i++) |
4ee9c684 | 191 | { |
1a91d914 | 192 | gimple cond_stmt; |
193 | gphi *phi; | |
33784d89 | 194 | basic_block bb1, bb2; |
195 | edge e1, e2; | |
194899bf | 196 | tree arg0, arg1; |
197 | ||
198 | bb = bb_order[i]; | |
20e5647c | 199 | |
75a70cf9 | 200 | cond_stmt = last_stmt (bb); |
201 | /* Check to see if the last statement is a GIMPLE_COND. */ | |
202 | if (!cond_stmt | |
203 | || gimple_code (cond_stmt) != GIMPLE_COND) | |
33784d89 | 204 | continue; |
20e5647c | 205 | |
33784d89 | 206 | e1 = EDGE_SUCC (bb, 0); |
207 | bb1 = e1->dest; | |
208 | e2 = EDGE_SUCC (bb, 1); | |
209 | bb2 = e2->dest; | |
20e5647c | 210 | |
33784d89 | 211 | /* We cannot do the optimization on abnormal edges. */ |
212 | if ((e1->flags & EDGE_ABNORMAL) != 0 | |
213 | || (e2->flags & EDGE_ABNORMAL) != 0) | |
214 | continue; | |
20e5647c | 215 | |
33784d89 | 216 | /* If either bb1's succ or bb2 or bb2's succ is non NULL. */ |
ea091dfd | 217 | if (EDGE_COUNT (bb1->succs) == 0 |
33784d89 | 218 | || bb2 == NULL |
ea091dfd | 219 | || EDGE_COUNT (bb2->succs) == 0) |
33784d89 | 220 | continue; |
20e5647c | 221 | |
33784d89 | 222 | /* Find the bb which is the fall through to the other. */ |
223 | if (EDGE_SUCC (bb1, 0)->dest == bb2) | |
224 | ; | |
225 | else if (EDGE_SUCC (bb2, 0)->dest == bb1) | |
226 | { | |
a4f59596 | 227 | std::swap (bb1, bb2); |
228 | std::swap (e1, e2); | |
33784d89 | 229 | } |
91cf53d5 | 230 | else if (do_store_elim |
231 | && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest) | |
232 | { | |
233 | basic_block bb3 = EDGE_SUCC (bb1, 0)->dest; | |
234 | ||
235 | if (!single_succ_p (bb1) | |
236 | || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0 | |
237 | || !single_succ_p (bb2) | |
238 | || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0 | |
239 | || EDGE_COUNT (bb3->preds) != 2) | |
240 | continue; | |
241 | if (cond_if_else_store_replacement (bb1, bb2, bb3)) | |
242 | cfgchanged = true; | |
243 | continue; | |
244 | } | |
239e9670 | 245 | else if (do_hoist_loads |
246 | && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest) | |
247 | { | |
248 | basic_block bb3 = EDGE_SUCC (bb1, 0)->dest; | |
249 | ||
250 | if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt))) | |
251 | && single_succ_p (bb1) | |
252 | && single_succ_p (bb2) | |
253 | && single_pred_p (bb1) | |
254 | && single_pred_p (bb2) | |
255 | && EDGE_COUNT (bb->succs) == 2 | |
256 | && EDGE_COUNT (bb3->preds) == 2 | |
257 | /* If one edge or the other is dominant, a conditional move | |
258 | is likely to perform worse than the well-predicted branch. */ | |
259 | && !predictable_edge_p (EDGE_SUCC (bb, 0)) | |
260 | && !predictable_edge_p (EDGE_SUCC (bb, 1))) | |
261 | hoist_adjacent_loads (bb, bb1, bb2, bb3); | |
262 | continue; | |
263 | } | |
33784d89 | 264 | else |
f32420fb | 265 | continue; |
20e5647c | 266 | |
33784d89 | 267 | e1 = EDGE_SUCC (bb1, 0); |
20e5647c | 268 | |
33784d89 | 269 | /* Make sure that bb1 is just a fall through. */ |
db5ba14c | 270 | if (!single_succ_p (bb1) |
33784d89 | 271 | || (e1->flags & EDGE_FALLTHRU) == 0) |
272 | continue; | |
20e5647c | 273 | |
3472707f | 274 | /* Also make sure that bb1 only have one predecessor and that it |
275 | is bb. */ | |
ea091dfd | 276 | if (!single_pred_p (bb1) |
277 | || single_pred (bb1) != bb) | |
33784d89 | 278 | continue; |
20e5647c | 279 | |
e6d0e152 | 280 | if (do_store_elim) |
281 | { | |
282 | /* bb1 is the middle block, bb2 the join block, bb the split block, | |
283 | e1 the fallthrough edge from bb1 to bb2. We can't do the | |
284 | optimization if the join block has more than two predecessors. */ | |
285 | if (EDGE_COUNT (bb2->preds) > 2) | |
286 | continue; | |
287 | if (cond_store_replacement (bb1, bb2, e1, e2, nontrap)) | |
288 | cfgchanged = true; | |
289 | } | |
290 | else | |
291 | { | |
75a70cf9 | 292 | gimple_seq phis = phi_nodes (bb2); |
2109076a | 293 | gimple_stmt_iterator gsi; |
fb9912ea | 294 | bool candorest = true; |
c3597b05 | 295 | |
fb9912ea | 296 | /* Value replacement can work with more than one PHI |
297 | so try that first. */ | |
298 | for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi)) | |
299 | { | |
1a91d914 | 300 | phi = as_a <gphi *> (gsi_stmt (gsi)); |
fb9912ea | 301 | arg0 = gimple_phi_arg_def (phi, e1->dest_idx); |
302 | arg1 = gimple_phi_arg_def (phi, e2->dest_idx); | |
303 | if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2) | |
304 | { | |
305 | candorest = false; | |
306 | cfgchanged = true; | |
307 | break; | |
308 | } | |
309 | } | |
e6d0e152 | 310 | |
fb9912ea | 311 | if (!candorest) |
312 | continue; | |
f32420fb | 313 | |
c3597b05 | 314 | phi = single_non_singleton_phi_for_edges (phis, e1, e2); |
2109076a | 315 | if (!phi) |
e6d0e152 | 316 | continue; |
317 | ||
75a70cf9 | 318 | arg0 = gimple_phi_arg_def (phi, e1->dest_idx); |
319 | arg1 = gimple_phi_arg_def (phi, e2->dest_idx); | |
e6d0e152 | 320 | |
321 | /* Something is wrong if we cannot find the arguments in the PHI | |
322 | node. */ | |
323 | gcc_assert (arg0 != NULL && arg1 != NULL); | |
324 | ||
325 | /* Do the replacement of conditional if it can be done. */ | |
326 | if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) | |
327 | cfgchanged = true; | |
e6d0e152 | 328 | else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) |
329 | cfgchanged = true; | |
330 | else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1)) | |
331 | cfgchanged = true; | |
332 | } | |
194899bf | 333 | } |
334 | ||
335 | free (bb_order); | |
48e1416a | 336 | |
e6d0e152 | 337 | if (do_store_elim) |
431205b7 | 338 | delete nontrap; |
e6d0e152 | 339 | /* If the CFG has changed, we should cleanup the CFG. */ |
340 | if (cfgchanged && do_store_elim) | |
341 | { | |
342 | /* In cond-store replacement we have added some loads on edges | |
343 | and new VOPS (as we moved the store, and created a load). */ | |
75a70cf9 | 344 | gsi_commit_edge_inserts (); |
e6d0e152 | 345 | return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals; |
346 | } | |
347 | else if (cfgchanged) | |
348 | return TODO_cleanup_cfg; | |
349 | return 0; | |
194899bf | 350 | } |
351 | ||
fccee353 | 352 | /* Replace PHI node element whose edge is E in block BB with variable NEW. |
33784d89 | 353 | Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK |
902929aa | 354 | is known to have two edges, one of which must reach BB). */ |
355 | ||
356 | static void | |
a4844041 | 357 | replace_phi_edge_with_variable (basic_block cond_block, |
75a70cf9 | 358 | edge e, gimple phi, tree new_tree) |
902929aa | 359 | { |
75a70cf9 | 360 | basic_block bb = gimple_bb (phi); |
0e1a77e1 | 361 | basic_block block_to_remove; |
75a70cf9 | 362 | gimple_stmt_iterator gsi; |
33784d89 | 363 | |
20e5647c | 364 | /* Change the PHI argument to new. */ |
f0d6e81c | 365 | SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree); |
0e1a77e1 | 366 | |
0e1a77e1 | 367 | /* Remove the empty basic block. */ |
cd665a06 | 368 | if (EDGE_SUCC (cond_block, 0)->dest == bb) |
902929aa | 369 | { |
cd665a06 | 370 | EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU; |
371 | EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); | |
81c5be57 | 372 | EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE; |
373 | EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count; | |
0e1a77e1 | 374 | |
cd665a06 | 375 | block_to_remove = EDGE_SUCC (cond_block, 1)->dest; |
902929aa | 376 | } |
377 | else | |
378 | { | |
cd665a06 | 379 | EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU; |
380 | EDGE_SUCC (cond_block, 1)->flags | |
902929aa | 381 | &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE); |
81c5be57 | 382 | EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE; |
383 | EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count; | |
0e1a77e1 | 384 | |
cd665a06 | 385 | block_to_remove = EDGE_SUCC (cond_block, 0)->dest; |
902929aa | 386 | } |
0e1a77e1 | 387 | delete_basic_block (block_to_remove); |
20e5647c | 388 | |
902929aa | 389 | /* Eliminate the COND_EXPR at the end of COND_BLOCK. */ |
75a70cf9 | 390 | gsi = gsi_last_bb (cond_block); |
391 | gsi_remove (&gsi, true); | |
20e5647c | 392 | |
902929aa | 393 | if (dump_file && (dump_flags & TDF_DETAILS)) |
394 | fprintf (dump_file, | |
395 | "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n", | |
396 | cond_block->index, | |
397 | bb->index); | |
398 | } | |
399 | ||
400 | /* The function conditional_replacement does the main work of doing the | |
401 | conditional replacement. Return true if the replacement is done. | |
402 | Otherwise return false. | |
403 | BB is the basic block where the replacement is going to be done on. ARG0 | |
dac49aa5 | 404 | is argument 0 from PHI. Likewise for ARG1. */ |
902929aa | 405 | |
406 | static bool | |
33784d89 | 407 | conditional_replacement (basic_block cond_bb, basic_block middle_bb, |
1a91d914 | 408 | edge e0, edge e1, gphi *phi, |
33784d89 | 409 | tree arg0, tree arg1) |
902929aa | 410 | { |
411 | tree result; | |
1a91d914 | 412 | gimple stmt; |
413 | gassign *new_stmt; | |
75a70cf9 | 414 | tree cond; |
415 | gimple_stmt_iterator gsi; | |
902929aa | 416 | edge true_edge, false_edge; |
75a70cf9 | 417 | tree new_var, new_var2; |
678919fd | 418 | bool neg; |
902929aa | 419 | |
435e1a75 | 420 | /* FIXME: Gimplification of complex type is too hard for now. */ |
47b88316 | 421 | /* We aren't prepared to handle vectors either (and it is a question |
422 | if it would be worthwhile anyway). */ | |
423 | if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0)) | |
424 | || POINTER_TYPE_P (TREE_TYPE (arg0))) | |
425 | || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1)) | |
426 | || POINTER_TYPE_P (TREE_TYPE (arg1)))) | |
435e1a75 | 427 | return false; |
428 | ||
678919fd | 429 | /* The PHI arguments have the constants 0 and 1, or 0 and -1, then |
430 | convert it to the conditional. */ | |
902929aa | 431 | if ((integer_zerop (arg0) && integer_onep (arg1)) |
432 | || (integer_zerop (arg1) && integer_onep (arg0))) | |
678919fd | 433 | neg = false; |
434 | else if ((integer_zerop (arg0) && integer_all_onesp (arg1)) | |
435 | || (integer_zerop (arg1) && integer_all_onesp (arg0))) | |
436 | neg = true; | |
902929aa | 437 | else |
438 | return false; | |
20e5647c | 439 | |
33784d89 | 440 | if (!empty_block_p (middle_bb)) |
902929aa | 441 | return false; |
20e5647c | 442 | |
75a70cf9 | 443 | /* At this point we know we have a GIMPLE_COND with two successors. |
2ab0a163 | 444 | One successor is BB, the other successor is an empty block which |
445 | falls through into BB. | |
20e5647c | 446 | |
2ab0a163 | 447 | There is a single PHI node at the join point (BB) and its arguments |
678919fd | 448 | are constants (0, 1) or (0, -1). |
20e5647c | 449 | |
2ab0a163 | 450 | So, given the condition COND, and the two PHI arguments, we can |
20e5647c | 451 | rewrite this PHI into non-branching code: |
452 | ||
2ab0a163 | 453 | dest = (COND) or dest = COND' |
20e5647c | 454 | |
2ab0a163 | 455 | We use the condition as-is if the argument associated with the |
456 | true edge has the value one or the argument associated with the | |
457 | false edge as the value zero. Note that those conditions are not | |
75a70cf9 | 458 | the same since only one of the outgoing edges from the GIMPLE_COND |
2ab0a163 | 459 | will directly reach BB and thus be associated with an argument. */ |
ae5a4794 | 460 | |
75a70cf9 | 461 | stmt = last_stmt (cond_bb); |
462 | result = PHI_RESULT (phi); | |
b2a02a0e | 463 | |
75a70cf9 | 464 | /* To handle special cases like floating point comparison, it is easier and |
465 | less error-prone to build a tree and gimplify it on the fly though it is | |
466 | less efficient. */ | |
6f9714b3 | 467 | cond = fold_build2_loc (gimple_location (stmt), |
468 | gimple_cond_code (stmt), boolean_type_node, | |
469 | gimple_cond_lhs (stmt), gimple_cond_rhs (stmt)); | |
4ee9c684 | 470 | |
75a70cf9 | 471 | /* We need to know which is the true edge and which is the false |
472 | edge so that we know when to invert the condition below. */ | |
473 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); | |
474 | if ((e0 == true_edge && integer_zerop (arg0)) | |
678919fd | 475 | || (e0 == false_edge && !integer_zerop (arg0)) |
75a70cf9 | 476 | || (e1 == true_edge && integer_zerop (arg1)) |
678919fd | 477 | || (e1 == false_edge && !integer_zerop (arg1))) |
6f9714b3 | 478 | cond = fold_build1_loc (gimple_location (stmt), |
678919fd | 479 | TRUTH_NOT_EXPR, TREE_TYPE (cond), cond); |
480 | ||
481 | if (neg) | |
482 | { | |
483 | cond = fold_convert_loc (gimple_location (stmt), | |
484 | TREE_TYPE (result), cond); | |
485 | cond = fold_build1_loc (gimple_location (stmt), | |
486 | NEGATE_EXPR, TREE_TYPE (cond), cond); | |
487 | } | |
75a70cf9 | 488 | |
489 | /* Insert our new statements at the end of conditional block before the | |
490 | COND_STMT. */ | |
491 | gsi = gsi_for_stmt (stmt); | |
492 | new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true, | |
493 | GSI_SAME_STMT); | |
494 | ||
495 | if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var))) | |
496 | { | |
efbcb6de | 497 | source_location locus_0, locus_1; |
498 | ||
f9e245b2 | 499 | new_var2 = make_ssa_name (TREE_TYPE (result)); |
e9cf809e | 500 | new_stmt = gimple_build_assign (new_var2, CONVERT_EXPR, new_var); |
75a70cf9 | 501 | gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT); |
502 | new_var = new_var2; | |
efbcb6de | 503 | |
504 | /* Set the locus to the first argument, unless is doesn't have one. */ | |
505 | locus_0 = gimple_phi_arg_location (phi, 0); | |
506 | locus_1 = gimple_phi_arg_location (phi, 1); | |
507 | if (locus_0 == UNKNOWN_LOCATION) | |
508 | locus_0 = locus_1; | |
509 | gimple_set_location (new_stmt, locus_0); | |
4ee9c684 | 510 | } |
20e5647c | 511 | |
75a70cf9 | 512 | replace_phi_edge_with_variable (cond_bb, e1, phi, new_var); |
902929aa | 513 | |
4ee9c684 | 514 | /* Note that we optimized this PHI. */ |
515 | return true; | |
516 | } | |
517 | ||
17b9476e | 518 | /* Update *ARG which is defined in STMT so that it contains the |
519 | computed value if that seems profitable. Return true if the | |
520 | statement is made dead by that rewriting. */ | |
521 | ||
522 | static bool | |
523 | jump_function_from_stmt (tree *arg, gimple stmt) | |
524 | { | |
525 | enum tree_code code = gimple_assign_rhs_code (stmt); | |
526 | if (code == ADDR_EXPR) | |
527 | { | |
528 | /* For arg = &p->i transform it to p, if possible. */ | |
529 | tree rhs1 = gimple_assign_rhs1 (stmt); | |
530 | HOST_WIDE_INT offset; | |
531 | tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0), | |
532 | &offset); | |
533 | if (tem | |
534 | && TREE_CODE (tem) == MEM_REF | |
796b6678 | 535 | && (mem_ref_offset (tem) + offset) == 0) |
17b9476e | 536 | { |
537 | *arg = TREE_OPERAND (tem, 0); | |
538 | return true; | |
539 | } | |
540 | } | |
541 | /* TODO: Much like IPA-CP jump-functions we want to handle constant | |
542 | additions symbolically here, and we'd need to update the comparison | |
543 | code that compares the arg + cst tuples in our caller. For now the | |
544 | code above exactly handles the VEC_BASE pattern from vec.h. */ | |
545 | return false; | |
546 | } | |
547 | ||
f32420fb | 548 | /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional |
549 | of the form SSA_NAME NE 0. | |
550 | ||
551 | If RHS is fed by a simple EQ_EXPR comparison of two values, see if | |
552 | the two input values of the EQ_EXPR match arg0 and arg1. | |
553 | ||
554 | If so update *code and return TRUE. Otherwise return FALSE. */ | |
555 | ||
556 | static bool | |
557 | rhs_is_fed_for_value_replacement (const_tree arg0, const_tree arg1, | |
558 | enum tree_code *code, const_tree rhs) | |
559 | { | |
560 | /* Obviously if RHS is not an SSA_NAME, we can't look at the defining | |
561 | statement. */ | |
562 | if (TREE_CODE (rhs) == SSA_NAME) | |
563 | { | |
564 | gimple def1 = SSA_NAME_DEF_STMT (rhs); | |
565 | ||
566 | /* Verify the defining statement has an EQ_EXPR on the RHS. */ | |
567 | if (is_gimple_assign (def1) && gimple_assign_rhs_code (def1) == EQ_EXPR) | |
568 | { | |
569 | /* Finally verify the source operands of the EQ_EXPR are equal | |
570 | to arg0 and arg1. */ | |
571 | tree op0 = gimple_assign_rhs1 (def1); | |
572 | tree op1 = gimple_assign_rhs2 (def1); | |
573 | if ((operand_equal_for_phi_arg_p (arg0, op0) | |
574 | && operand_equal_for_phi_arg_p (arg1, op1)) | |
575 | || (operand_equal_for_phi_arg_p (arg0, op1) | |
576 | && operand_equal_for_phi_arg_p (arg1, op0))) | |
577 | { | |
578 | /* We will perform the optimization. */ | |
579 | *code = gimple_assign_rhs_code (def1); | |
580 | return true; | |
581 | } | |
582 | } | |
583 | } | |
584 | return false; | |
585 | } | |
586 | ||
587 | /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND. | |
588 | ||
589 | Also return TRUE if arg0/arg1 are equal to the source arguments of a | |
590 | an EQ comparison feeding a BIT_AND_EXPR which feeds COND. | |
591 | ||
592 | Return FALSE otherwise. */ | |
593 | ||
594 | static bool | |
595 | operand_equal_for_value_replacement (const_tree arg0, const_tree arg1, | |
596 | enum tree_code *code, gimple cond) | |
597 | { | |
598 | gimple def; | |
599 | tree lhs = gimple_cond_lhs (cond); | |
600 | tree rhs = gimple_cond_rhs (cond); | |
601 | ||
602 | if ((operand_equal_for_phi_arg_p (arg0, lhs) | |
603 | && operand_equal_for_phi_arg_p (arg1, rhs)) | |
604 | || (operand_equal_for_phi_arg_p (arg1, lhs) | |
605 | && operand_equal_for_phi_arg_p (arg0, rhs))) | |
606 | return true; | |
607 | ||
608 | /* Now handle more complex case where we have an EQ comparison | |
609 | which feeds a BIT_AND_EXPR which feeds COND. | |
610 | ||
611 | First verify that COND is of the form SSA_NAME NE 0. */ | |
612 | if (*code != NE_EXPR || !integer_zerop (rhs) | |
613 | || TREE_CODE (lhs) != SSA_NAME) | |
614 | return false; | |
615 | ||
616 | /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */ | |
617 | def = SSA_NAME_DEF_STMT (lhs); | |
618 | if (!is_gimple_assign (def) || gimple_assign_rhs_code (def) != BIT_AND_EXPR) | |
619 | return false; | |
620 | ||
621 | /* Now verify arg0/arg1 correspond to the source arguments of an | |
622 | EQ comparison feeding the BIT_AND_EXPR. */ | |
623 | ||
624 | tree tmp = gimple_assign_rhs1 (def); | |
625 | if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp)) | |
626 | return true; | |
627 | ||
628 | tmp = gimple_assign_rhs2 (def); | |
629 | if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp)) | |
630 | return true; | |
631 | ||
632 | return false; | |
633 | } | |
634 | ||
b6814ca0 | 635 | /* Returns true if ARG is a neutral element for operation CODE |
636 | on the RIGHT side. */ | |
637 | ||
638 | static bool | |
639 | neutral_element_p (tree_code code, tree arg, bool right) | |
640 | { | |
641 | switch (code) | |
642 | { | |
643 | case PLUS_EXPR: | |
644 | case BIT_IOR_EXPR: | |
645 | case BIT_XOR_EXPR: | |
646 | return integer_zerop (arg); | |
647 | ||
648 | case LROTATE_EXPR: | |
649 | case RROTATE_EXPR: | |
650 | case LSHIFT_EXPR: | |
651 | case RSHIFT_EXPR: | |
652 | case MINUS_EXPR: | |
653 | case POINTER_PLUS_EXPR: | |
654 | return right && integer_zerop (arg); | |
655 | ||
656 | case MULT_EXPR: | |
657 | return integer_onep (arg); | |
658 | ||
659 | case TRUNC_DIV_EXPR: | |
660 | case CEIL_DIV_EXPR: | |
661 | case FLOOR_DIV_EXPR: | |
662 | case ROUND_DIV_EXPR: | |
663 | case EXACT_DIV_EXPR: | |
664 | return right && integer_onep (arg); | |
665 | ||
666 | case BIT_AND_EXPR: | |
667 | return integer_all_onesp (arg); | |
668 | ||
669 | default: | |
670 | return false; | |
671 | } | |
672 | } | |
673 | ||
674 | /* Returns true if ARG is an absorbing element for operation CODE. */ | |
675 | ||
676 | static bool | |
677 | absorbing_element_p (tree_code code, tree arg) | |
678 | { | |
679 | switch (code) | |
680 | { | |
681 | case BIT_IOR_EXPR: | |
682 | return integer_all_onesp (arg); | |
683 | ||
684 | case MULT_EXPR: | |
685 | case BIT_AND_EXPR: | |
686 | return integer_zerop (arg); | |
687 | ||
688 | default: | |
689 | return false; | |
690 | } | |
691 | } | |
692 | ||
0beac6fc | 693 | /* The function value_replacement does the main work of doing the value |
fb9912ea | 694 | replacement. Return non-zero if the replacement is done. Otherwise return |
695 | 0. If we remove the middle basic block, return 2. | |
0beac6fc | 696 | BB is the basic block where the replacement is going to be done on. ARG0 |
dac49aa5 | 697 | is argument 0 from the PHI. Likewise for ARG1. */ |
0beac6fc | 698 | |
fb9912ea | 699 | static int |
33784d89 | 700 | value_replacement (basic_block cond_bb, basic_block middle_bb, |
75a70cf9 | 701 | edge e0, edge e1, gimple phi, |
33784d89 | 702 | tree arg0, tree arg1) |
0beac6fc | 703 | { |
17b9476e | 704 | gimple_stmt_iterator gsi; |
75a70cf9 | 705 | gimple cond; |
0beac6fc | 706 | edge true_edge, false_edge; |
75a70cf9 | 707 | enum tree_code code; |
fb9912ea | 708 | bool emtpy_or_with_defined_p = true; |
0beac6fc | 709 | |
710 | /* If the type says honor signed zeros we cannot do this | |
dac49aa5 | 711 | optimization. */ |
fe994837 | 712 | if (HONOR_SIGNED_ZEROS (arg1)) |
fb9912ea | 713 | return 0; |
0beac6fc | 714 | |
fb9912ea | 715 | /* If there is a statement in MIDDLE_BB that defines one of the PHI |
716 | arguments, then adjust arg0 or arg1. */ | |
b6814ca0 | 717 | gsi = gsi_start_nondebug_after_labels_bb (middle_bb); |
fb9912ea | 718 | while (!gsi_end_p (gsi)) |
17b9476e | 719 | { |
fb9912ea | 720 | gimple stmt = gsi_stmt (gsi); |
721 | tree lhs; | |
722 | gsi_next_nondebug (&gsi); | |
723 | if (!is_gimple_assign (stmt)) | |
17b9476e | 724 | { |
fb9912ea | 725 | emtpy_or_with_defined_p = false; |
726 | continue; | |
17b9476e | 727 | } |
fb9912ea | 728 | /* Now try to adjust arg0 or arg1 according to the computation |
729 | in the statement. */ | |
730 | lhs = gimple_assign_lhs (stmt); | |
731 | if (!(lhs == arg0 | |
732 | && jump_function_from_stmt (&arg0, stmt)) | |
733 | || (lhs == arg1 | |
734 | && jump_function_from_stmt (&arg1, stmt))) | |
735 | emtpy_or_with_defined_p = false; | |
17b9476e | 736 | } |
0beac6fc | 737 | |
75a70cf9 | 738 | cond = last_stmt (cond_bb); |
739 | code = gimple_cond_code (cond); | |
0beac6fc | 740 | |
741 | /* This transformation is only valid for equality comparisons. */ | |
75a70cf9 | 742 | if (code != NE_EXPR && code != EQ_EXPR) |
fb9912ea | 743 | return 0; |
0beac6fc | 744 | |
745 | /* We need to know which is the true edge and which is the false | |
746 | edge so that we know if have abs or negative abs. */ | |
33784d89 | 747 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |
0beac6fc | 748 | |
749 | /* At this point we know we have a COND_EXPR with two successors. | |
750 | One successor is BB, the other successor is an empty block which | |
751 | falls through into BB. | |
752 | ||
753 | The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR. | |
754 | ||
755 | There is a single PHI node at the join point (BB) with two arguments. | |
756 | ||
757 | We now need to verify that the two arguments in the PHI node match | |
758 | the two arguments to the equality comparison. */ | |
20e5647c | 759 | |
f32420fb | 760 | if (operand_equal_for_value_replacement (arg0, arg1, &code, cond)) |
0beac6fc | 761 | { |
762 | edge e; | |
763 | tree arg; | |
764 | ||
50737d20 | 765 | /* For NE_EXPR, we want to build an assignment result = arg where |
766 | arg is the PHI argument associated with the true edge. For | |
767 | EQ_EXPR we want the PHI argument associated with the false edge. */ | |
75a70cf9 | 768 | e = (code == NE_EXPR ? true_edge : false_edge); |
50737d20 | 769 | |
770 | /* Unfortunately, E may not reach BB (it may instead have gone to | |
771 | OTHER_BLOCK). If that is the case, then we want the single outgoing | |
772 | edge from OTHER_BLOCK which reaches BB and represents the desired | |
773 | path from COND_BLOCK. */ | |
33784d89 | 774 | if (e->dest == middle_bb) |
ea091dfd | 775 | e = single_succ_edge (e->dest); |
50737d20 | 776 | |
777 | /* Now we know the incoming edge to BB that has the argument for the | |
778 | RHS of our new assignment statement. */ | |
33784d89 | 779 | if (e0 == e) |
0beac6fc | 780 | arg = arg0; |
781 | else | |
782 | arg = arg1; | |
783 | ||
fb9912ea | 784 | /* If the middle basic block was empty or is defining the |
c3597b05 | 785 | PHI arguments and this is a single phi where the args are different |
786 | for the edges e0 and e1 then we can remove the middle basic block. */ | |
fb9912ea | 787 | if (emtpy_or_with_defined_p |
c3597b05 | 788 | && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), |
ce75c7c2 | 789 | e0, e1) == phi) |
fb9912ea | 790 | { |
791 | replace_phi_edge_with_variable (cond_bb, e1, phi, arg); | |
792 | /* Note that we optimized this PHI. */ | |
793 | return 2; | |
794 | } | |
795 | else | |
796 | { | |
797 | /* Replace the PHI arguments with arg. */ | |
798 | SET_PHI_ARG_DEF (phi, e0->dest_idx, arg); | |
799 | SET_PHI_ARG_DEF (phi, e1->dest_idx, arg); | |
800 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
801 | { | |
802 | fprintf (dump_file, "PHI "); | |
803 | print_generic_expr (dump_file, gimple_phi_result (phi), 0); | |
c3597b05 | 804 | fprintf (dump_file, " reduced for COND_EXPR in block %d to ", |
805 | cond_bb->index); | |
fb9912ea | 806 | print_generic_expr (dump_file, arg, 0); |
807 | fprintf (dump_file, ".\n"); | |
808 | } | |
809 | return 1; | |
810 | } | |
0beac6fc | 811 | |
0beac6fc | 812 | } |
b6814ca0 | 813 | |
814 | /* Now optimize (x != 0) ? x + y : y to just y. | |
815 | The following condition is too restrictive, there can easily be another | |
816 | stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */ | |
817 | gimple assign = last_and_only_stmt (middle_bb); | |
818 | if (!assign || gimple_code (assign) != GIMPLE_ASSIGN | |
819 | || gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS | |
820 | || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0)) | |
821 | && !POINTER_TYPE_P (TREE_TYPE (arg0)))) | |
822 | return 0; | |
823 | ||
6c96fe34 | 824 | /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */ |
825 | if (!gimple_seq_empty_p (phi_nodes (middle_bb))) | |
826 | return 0; | |
827 | ||
43fbec4a | 828 | /* Only transform if it removes the condition. */ |
829 | if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, e1)) | |
830 | return 0; | |
831 | ||
b6814ca0 | 832 | /* Size-wise, this is always profitable. */ |
833 | if (optimize_bb_for_speed_p (cond_bb) | |
834 | /* The special case is useless if it has a low probability. */ | |
835 | && profile_status_for_fn (cfun) != PROFILE_ABSENT | |
836 | && EDGE_PRED (middle_bb, 0)->probability < PROB_EVEN | |
837 | /* If assign is cheap, there is no point avoiding it. */ | |
838 | && estimate_num_insns (assign, &eni_time_weights) | |
839 | >= 3 * estimate_num_insns (cond, &eni_time_weights)) | |
840 | return 0; | |
841 | ||
842 | tree lhs = gimple_assign_lhs (assign); | |
843 | tree rhs1 = gimple_assign_rhs1 (assign); | |
844 | tree rhs2 = gimple_assign_rhs2 (assign); | |
845 | enum tree_code code_def = gimple_assign_rhs_code (assign); | |
846 | tree cond_lhs = gimple_cond_lhs (cond); | |
847 | tree cond_rhs = gimple_cond_rhs (cond); | |
848 | ||
849 | if (((code == NE_EXPR && e1 == false_edge) | |
850 | || (code == EQ_EXPR && e1 == true_edge)) | |
851 | && arg0 == lhs | |
852 | && ((arg1 == rhs1 | |
853 | && operand_equal_for_phi_arg_p (rhs2, cond_lhs) | |
854 | && neutral_element_p (code_def, cond_rhs, true)) | |
855 | || (arg1 == rhs2 | |
856 | && operand_equal_for_phi_arg_p (rhs1, cond_lhs) | |
857 | && neutral_element_p (code_def, cond_rhs, false)) | |
858 | || (operand_equal_for_phi_arg_p (arg1, cond_rhs) | |
859 | && (operand_equal_for_phi_arg_p (rhs2, cond_lhs) | |
860 | || operand_equal_for_phi_arg_p (rhs1, cond_lhs)) | |
861 | && absorbing_element_p (code_def, cond_rhs)))) | |
862 | { | |
863 | gsi = gsi_for_stmt (cond); | |
18c06fb8 | 864 | if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))) |
865 | { | |
866 | /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6 | |
867 | def-stmt in: | |
868 | if (n_5 != 0) | |
869 | goto <bb 3>; | |
870 | else | |
871 | goto <bb 4>; | |
872 | ||
873 | <bb 3>: | |
874 | # RANGE [0, 4294967294] | |
875 | u_6 = n_5 + 4294967295; | |
876 | ||
877 | <bb 4>: | |
878 | # u_3 = PHI <u_6(3), 4294967295(2)> */ | |
879 | SSA_NAME_RANGE_INFO (lhs) = NULL; | |
880 | SSA_NAME_ANTI_RANGE_P (lhs) = 0; | |
881 | /* If available, we can use VR of phi result at least. */ | |
882 | tree phires = gimple_phi_result (phi); | |
883 | struct range_info_def *phires_range_info | |
884 | = SSA_NAME_RANGE_INFO (phires); | |
885 | if (phires_range_info) | |
886 | duplicate_ssa_name_range_info (lhs, SSA_NAME_RANGE_TYPE (phires), | |
887 | phires_range_info); | |
888 | } | |
b6814ca0 | 889 | gimple_stmt_iterator gsi_from = gsi_for_stmt (assign); |
890 | gsi_move_before (&gsi_from, &gsi); | |
891 | replace_phi_edge_with_variable (cond_bb, e1, phi, lhs); | |
892 | return 2; | |
893 | } | |
894 | ||
fb9912ea | 895 | return 0; |
0beac6fc | 896 | } |
897 | ||
194899bf | 898 | /* The function minmax_replacement does the main work of doing the minmax |
899 | replacement. Return true if the replacement is done. Otherwise return | |
900 | false. | |
901 | BB is the basic block where the replacement is going to be done on. ARG0 | |
902 | is argument 0 from the PHI. Likewise for ARG1. */ | |
903 | ||
904 | static bool | |
905 | minmax_replacement (basic_block cond_bb, basic_block middle_bb, | |
75a70cf9 | 906 | edge e0, edge e1, gimple phi, |
194899bf | 907 | tree arg0, tree arg1) |
908 | { | |
909 | tree result, type; | |
1a91d914 | 910 | gcond *cond; |
911 | gassign *new_stmt; | |
194899bf | 912 | edge true_edge, false_edge; |
913 | enum tree_code cmp, minmax, ass_code; | |
914 | tree smaller, larger, arg_true, arg_false; | |
75a70cf9 | 915 | gimple_stmt_iterator gsi, gsi_from; |
194899bf | 916 | |
917 | type = TREE_TYPE (PHI_RESULT (phi)); | |
918 | ||
919 | /* The optimization may be unsafe due to NaNs. */ | |
93633022 | 920 | if (HONOR_NANS (type)) |
194899bf | 921 | return false; |
922 | ||
1a91d914 | 923 | cond = as_a <gcond *> (last_stmt (cond_bb)); |
75a70cf9 | 924 | cmp = gimple_cond_code (cond); |
194899bf | 925 | |
926 | /* This transformation is only valid for order comparisons. Record which | |
927 | operand is smaller/larger if the result of the comparison is true. */ | |
928 | if (cmp == LT_EXPR || cmp == LE_EXPR) | |
929 | { | |
75a70cf9 | 930 | smaller = gimple_cond_lhs (cond); |
931 | larger = gimple_cond_rhs (cond); | |
194899bf | 932 | } |
933 | else if (cmp == GT_EXPR || cmp == GE_EXPR) | |
934 | { | |
75a70cf9 | 935 | smaller = gimple_cond_rhs (cond); |
936 | larger = gimple_cond_lhs (cond); | |
194899bf | 937 | } |
938 | else | |
939 | return false; | |
940 | ||
941 | /* We need to know which is the true edge and which is the false | |
942 | edge so that we know if have abs or negative abs. */ | |
943 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); | |
944 | ||
945 | /* Forward the edges over the middle basic block. */ | |
946 | if (true_edge->dest == middle_bb) | |
947 | true_edge = EDGE_SUCC (true_edge->dest, 0); | |
948 | if (false_edge->dest == middle_bb) | |
949 | false_edge = EDGE_SUCC (false_edge->dest, 0); | |
950 | ||
951 | if (true_edge == e0) | |
952 | { | |
953 | gcc_assert (false_edge == e1); | |
954 | arg_true = arg0; | |
955 | arg_false = arg1; | |
956 | } | |
957 | else | |
958 | { | |
959 | gcc_assert (false_edge == e0); | |
960 | gcc_assert (true_edge == e1); | |
961 | arg_true = arg1; | |
962 | arg_false = arg0; | |
963 | } | |
964 | ||
965 | if (empty_block_p (middle_bb)) | |
966 | { | |
967 | if (operand_equal_for_phi_arg_p (arg_true, smaller) | |
968 | && operand_equal_for_phi_arg_p (arg_false, larger)) | |
969 | { | |
970 | /* Case | |
48e1416a | 971 | |
194899bf | 972 | if (smaller < larger) |
973 | rslt = smaller; | |
974 | else | |
975 | rslt = larger; */ | |
976 | minmax = MIN_EXPR; | |
977 | } | |
978 | else if (operand_equal_for_phi_arg_p (arg_false, smaller) | |
979 | && operand_equal_for_phi_arg_p (arg_true, larger)) | |
980 | minmax = MAX_EXPR; | |
981 | else | |
982 | return false; | |
983 | } | |
984 | else | |
985 | { | |
986 | /* Recognize the following case, assuming d <= u: | |
987 | ||
988 | if (a <= u) | |
989 | b = MAX (a, d); | |
990 | x = PHI <b, u> | |
991 | ||
992 | This is equivalent to | |
993 | ||
994 | b = MAX (a, d); | |
995 | x = MIN (b, u); */ | |
996 | ||
75a70cf9 | 997 | gimple assign = last_and_only_stmt (middle_bb); |
998 | tree lhs, op0, op1, bound; | |
194899bf | 999 | |
1000 | if (!assign | |
75a70cf9 | 1001 | || gimple_code (assign) != GIMPLE_ASSIGN) |
194899bf | 1002 | return false; |
1003 | ||
75a70cf9 | 1004 | lhs = gimple_assign_lhs (assign); |
1005 | ass_code = gimple_assign_rhs_code (assign); | |
194899bf | 1006 | if (ass_code != MAX_EXPR && ass_code != MIN_EXPR) |
1007 | return false; | |
75a70cf9 | 1008 | op0 = gimple_assign_rhs1 (assign); |
1009 | op1 = gimple_assign_rhs2 (assign); | |
194899bf | 1010 | |
1011 | if (true_edge->src == middle_bb) | |
1012 | { | |
1013 | /* We got here if the condition is true, i.e., SMALLER < LARGER. */ | |
1014 | if (!operand_equal_for_phi_arg_p (lhs, arg_true)) | |
1015 | return false; | |
1016 | ||
1017 | if (operand_equal_for_phi_arg_p (arg_false, larger)) | |
1018 | { | |
1019 | /* Case | |
1020 | ||
1021 | if (smaller < larger) | |
1022 | { | |
1023 | r' = MAX_EXPR (smaller, bound) | |
1024 | } | |
1025 | r = PHI <r', larger> --> to be turned to MIN_EXPR. */ | |
1026 | if (ass_code != MAX_EXPR) | |
1027 | return false; | |
1028 | ||
1029 | minmax = MIN_EXPR; | |
1030 | if (operand_equal_for_phi_arg_p (op0, smaller)) | |
1031 | bound = op1; | |
1032 | else if (operand_equal_for_phi_arg_p (op1, smaller)) | |
1033 | bound = op0; | |
1034 | else | |
1035 | return false; | |
1036 | ||
1037 | /* We need BOUND <= LARGER. */ | |
49d00087 | 1038 | if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, |
1039 | bound, larger))) | |
194899bf | 1040 | return false; |
1041 | } | |
1042 | else if (operand_equal_for_phi_arg_p (arg_false, smaller)) | |
1043 | { | |
1044 | /* Case | |
1045 | ||
1046 | if (smaller < larger) | |
1047 | { | |
1048 | r' = MIN_EXPR (larger, bound) | |
1049 | } | |
1050 | r = PHI <r', smaller> --> to be turned to MAX_EXPR. */ | |
1051 | if (ass_code != MIN_EXPR) | |
1052 | return false; | |
1053 | ||
1054 | minmax = MAX_EXPR; | |
1055 | if (operand_equal_for_phi_arg_p (op0, larger)) | |
1056 | bound = op1; | |
1057 | else if (operand_equal_for_phi_arg_p (op1, larger)) | |
1058 | bound = op0; | |
1059 | else | |
1060 | return false; | |
1061 | ||
1062 | /* We need BOUND >= SMALLER. */ | |
49d00087 | 1063 | if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, |
1064 | bound, smaller))) | |
194899bf | 1065 | return false; |
1066 | } | |
1067 | else | |
1068 | return false; | |
1069 | } | |
1070 | else | |
1071 | { | |
1072 | /* We got here if the condition is false, i.e., SMALLER > LARGER. */ | |
1073 | if (!operand_equal_for_phi_arg_p (lhs, arg_false)) | |
1074 | return false; | |
1075 | ||
1076 | if (operand_equal_for_phi_arg_p (arg_true, larger)) | |
1077 | { | |
1078 | /* Case | |
1079 | ||
1080 | if (smaller > larger) | |
1081 | { | |
1082 | r' = MIN_EXPR (smaller, bound) | |
1083 | } | |
1084 | r = PHI <r', larger> --> to be turned to MAX_EXPR. */ | |
1085 | if (ass_code != MIN_EXPR) | |
1086 | return false; | |
1087 | ||
1088 | minmax = MAX_EXPR; | |
1089 | if (operand_equal_for_phi_arg_p (op0, smaller)) | |
1090 | bound = op1; | |
1091 | else if (operand_equal_for_phi_arg_p (op1, smaller)) | |
1092 | bound = op0; | |
1093 | else | |
1094 | return false; | |
1095 | ||
1096 | /* We need BOUND >= LARGER. */ | |
49d00087 | 1097 | if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node, |
1098 | bound, larger))) | |
194899bf | 1099 | return false; |
1100 | } | |
1101 | else if (operand_equal_for_phi_arg_p (arg_true, smaller)) | |
1102 | { | |
1103 | /* Case | |
1104 | ||
1105 | if (smaller > larger) | |
1106 | { | |
1107 | r' = MAX_EXPR (larger, bound) | |
1108 | } | |
1109 | r = PHI <r', smaller> --> to be turned to MIN_EXPR. */ | |
1110 | if (ass_code != MAX_EXPR) | |
1111 | return false; | |
1112 | ||
1113 | minmax = MIN_EXPR; | |
1114 | if (operand_equal_for_phi_arg_p (op0, larger)) | |
1115 | bound = op1; | |
1116 | else if (operand_equal_for_phi_arg_p (op1, larger)) | |
1117 | bound = op0; | |
1118 | else | |
1119 | return false; | |
1120 | ||
1121 | /* We need BOUND <= SMALLER. */ | |
49d00087 | 1122 | if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node, |
1123 | bound, smaller))) | |
194899bf | 1124 | return false; |
1125 | } | |
1126 | else | |
1127 | return false; | |
1128 | } | |
1129 | ||
1130 | /* Move the statement from the middle block. */ | |
75a70cf9 | 1131 | gsi = gsi_last_bb (cond_bb); |
445a6ba5 | 1132 | gsi_from = gsi_last_nondebug_bb (middle_bb); |
75a70cf9 | 1133 | gsi_move_before (&gsi_from, &gsi); |
194899bf | 1134 | } |
1135 | ||
1136 | /* Emit the statement to compute min/max. */ | |
1137 | result = duplicate_ssa_name (PHI_RESULT (phi), NULL); | |
e9cf809e | 1138 | new_stmt = gimple_build_assign (result, minmax, arg0, arg1); |
75a70cf9 | 1139 | gsi = gsi_last_bb (cond_bb); |
1140 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); | |
194899bf | 1141 | |
a4844041 | 1142 | replace_phi_edge_with_variable (cond_bb, e1, phi, result); |
194899bf | 1143 | return true; |
1144 | } | |
1145 | ||
70512b93 | 1146 | /* The function absolute_replacement does the main work of doing the absolute |
1147 | replacement. Return true if the replacement is done. Otherwise return | |
1148 | false. | |
1149 | bb is the basic block where the replacement is going to be done on. arg0 | |
f7f07c95 | 1150 | is argument 0 from the phi. Likewise for arg1. */ |
33784d89 | 1151 | |
70512b93 | 1152 | static bool |
33784d89 | 1153 | abs_replacement (basic_block cond_bb, basic_block middle_bb, |
a4844041 | 1154 | edge e0 ATTRIBUTE_UNUSED, edge e1, |
75a70cf9 | 1155 | gimple phi, tree arg0, tree arg1) |
70512b93 | 1156 | { |
1157 | tree result; | |
1a91d914 | 1158 | gassign *new_stmt; |
1159 | gimple cond; | |
75a70cf9 | 1160 | gimple_stmt_iterator gsi; |
70512b93 | 1161 | edge true_edge, false_edge; |
75a70cf9 | 1162 | gimple assign; |
70512b93 | 1163 | edge e; |
194899bf | 1164 | tree rhs, lhs; |
70512b93 | 1165 | bool negate; |
1166 | enum tree_code cond_code; | |
1167 | ||
1168 | /* If the type says honor signed zeros we cannot do this | |
dac49aa5 | 1169 | optimization. */ |
fe994837 | 1170 | if (HONOR_SIGNED_ZEROS (arg1)) |
70512b93 | 1171 | return false; |
1172 | ||
70512b93 | 1173 | /* OTHER_BLOCK must have only one executable statement which must have the |
1174 | form arg0 = -arg1 or arg1 = -arg0. */ | |
70512b93 | 1175 | |
194899bf | 1176 | assign = last_and_only_stmt (middle_bb); |
70512b93 | 1177 | /* If we did not find the proper negation assignment, then we can not |
1178 | optimize. */ | |
1179 | if (assign == NULL) | |
1180 | return false; | |
48e1416a | 1181 | |
194899bf | 1182 | /* If we got here, then we have found the only executable statement |
1183 | in OTHER_BLOCK. If it is anything other than arg = -arg1 or | |
1184 | arg1 = -arg0, then we can not optimize. */ | |
75a70cf9 | 1185 | if (gimple_code (assign) != GIMPLE_ASSIGN) |
194899bf | 1186 | return false; |
1187 | ||
75a70cf9 | 1188 | lhs = gimple_assign_lhs (assign); |
194899bf | 1189 | |
75a70cf9 | 1190 | if (gimple_assign_rhs_code (assign) != NEGATE_EXPR) |
194899bf | 1191 | return false; |
1192 | ||
75a70cf9 | 1193 | rhs = gimple_assign_rhs1 (assign); |
48e1416a | 1194 | |
194899bf | 1195 | /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */ |
1196 | if (!(lhs == arg0 && rhs == arg1) | |
1197 | && !(lhs == arg1 && rhs == arg0)) | |
1198 | return false; | |
70512b93 | 1199 | |
75a70cf9 | 1200 | cond = last_stmt (cond_bb); |
70512b93 | 1201 | result = PHI_RESULT (phi); |
1202 | ||
1203 | /* Only relationals comparing arg[01] against zero are interesting. */ | |
75a70cf9 | 1204 | cond_code = gimple_cond_code (cond); |
70512b93 | 1205 | if (cond_code != GT_EXPR && cond_code != GE_EXPR |
1206 | && cond_code != LT_EXPR && cond_code != LE_EXPR) | |
1207 | return false; | |
1208 | ||
dac49aa5 | 1209 | /* Make sure the conditional is arg[01] OP y. */ |
75a70cf9 | 1210 | if (gimple_cond_lhs (cond) != rhs) |
70512b93 | 1211 | return false; |
1212 | ||
75a70cf9 | 1213 | if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond))) |
1214 | ? real_zerop (gimple_cond_rhs (cond)) | |
1215 | : integer_zerop (gimple_cond_rhs (cond))) | |
70512b93 | 1216 | ; |
1217 | else | |
1218 | return false; | |
1219 | ||
1220 | /* We need to know which is the true edge and which is the false | |
1221 | edge so that we know if have abs or negative abs. */ | |
33784d89 | 1222 | extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge); |
70512b93 | 1223 | |
1224 | /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we | |
1225 | will need to negate the result. Similarly for LT_EXPR/LE_EXPR if | |
1226 | the false edge goes to OTHER_BLOCK. */ | |
1227 | if (cond_code == GT_EXPR || cond_code == GE_EXPR) | |
1228 | e = true_edge; | |
1229 | else | |
1230 | e = false_edge; | |
20e5647c | 1231 | |
33784d89 | 1232 | if (e->dest == middle_bb) |
70512b93 | 1233 | negate = true; |
1234 | else | |
1235 | negate = false; | |
20e5647c | 1236 | |
33784d89 | 1237 | result = duplicate_ssa_name (result, NULL); |
20e5647c | 1238 | |
70512b93 | 1239 | if (negate) |
f9e245b2 | 1240 | lhs = make_ssa_name (TREE_TYPE (result)); |
70512b93 | 1241 | else |
1242 | lhs = result; | |
1243 | ||
dac49aa5 | 1244 | /* Build the modify expression with abs expression. */ |
e9cf809e | 1245 | new_stmt = gimple_build_assign (lhs, ABS_EXPR, rhs); |
70512b93 | 1246 | |
75a70cf9 | 1247 | gsi = gsi_last_bb (cond_bb); |
1248 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); | |
70512b93 | 1249 | |
1250 | if (negate) | |
1251 | { | |
75a70cf9 | 1252 | /* Get the right GSI. We want to insert after the recently |
70512b93 | 1253 | added ABS_EXPR statement (which we know is the first statement |
1254 | in the block. */ | |
e9cf809e | 1255 | new_stmt = gimple_build_assign (result, NEGATE_EXPR, lhs); |
70512b93 | 1256 | |
75a70cf9 | 1257 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); |
70512b93 | 1258 | } |
20e5647c | 1259 | |
a4844041 | 1260 | replace_phi_edge_with_variable (cond_bb, e1, phi, result); |
70512b93 | 1261 | |
1262 | /* Note that we optimized this PHI. */ | |
1263 | return true; | |
1264 | } | |
1265 | ||
e6d0e152 | 1266 | /* Auxiliary functions to determine the set of memory accesses which |
1267 | can't trap because they are preceded by accesses to the same memory | |
182cf5a9 | 1268 | portion. We do that for MEM_REFs, so we only need to track |
e6d0e152 | 1269 | the SSA_NAME of the pointer indirectly referenced. The algorithm |
1270 | simply is a walk over all instructions in dominator order. When | |
182cf5a9 | 1271 | we see an MEM_REF we determine if we've already seen a same |
e6d0e152 | 1272 | ref anywhere up to the root of the dominator tree. If we do the |
af4f74fa | 1273 | current access can't trap. If we don't see any dominating access |
e6d0e152 | 1274 | the current access might trap, but might also make later accesses |
af4f74fa | 1275 | non-trapping, so we remember it. We need to be careful with loads |
1276 | or stores, for instance a load might not trap, while a store would, | |
1277 | so if we see a dominating read access this doesn't mean that a later | |
1278 | write access would not trap. Hence we also need to differentiate the | |
1279 | type of access(es) seen. | |
1280 | ||
1281 | ??? We currently are very conservative and assume that a load might | |
1282 | trap even if a store doesn't (write-only memory). This probably is | |
1283 | overly conservative. */ | |
e6d0e152 | 1284 | |
182cf5a9 | 1285 | /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF |
e6d0e152 | 1286 | through it was seen, which would constitute a no-trap region for |
1287 | same accesses. */ | |
1288 | struct name_to_bb | |
1289 | { | |
963aee26 | 1290 | unsigned int ssa_name_ver; |
42540642 | 1291 | unsigned int phase; |
963aee26 | 1292 | bool store; |
1293 | HOST_WIDE_INT offset, size; | |
e6d0e152 | 1294 | basic_block bb; |
1295 | }; | |
1296 | ||
d9dd21a8 | 1297 | /* Hashtable helpers. */ |
1298 | ||
298e7f9a | 1299 | struct ssa_names_hasher : free_ptr_hash <name_to_bb> |
d9dd21a8 | 1300 | { |
9969c043 | 1301 | static inline hashval_t hash (const name_to_bb *); |
1302 | static inline bool equal (const name_to_bb *, const name_to_bb *); | |
d9dd21a8 | 1303 | }; |
e6d0e152 | 1304 | |
42540642 | 1305 | /* Used for quick clearing of the hash-table when we see calls. |
1306 | Hash entries with phase < nt_call_phase are invalid. */ | |
1307 | static unsigned int nt_call_phase; | |
1308 | ||
963aee26 | 1309 | /* The hash function. */ |
d9dd21a8 | 1310 | |
1311 | inline hashval_t | |
9969c043 | 1312 | ssa_names_hasher::hash (const name_to_bb *n) |
e6d0e152 | 1313 | { |
963aee26 | 1314 | return n->ssa_name_ver ^ (((hashval_t) n->store) << 31) |
1315 | ^ (n->offset << 6) ^ (n->size << 3); | |
e6d0e152 | 1316 | } |
1317 | ||
963aee26 | 1318 | /* The equality function of *P1 and *P2. */ |
e6d0e152 | 1319 | |
d9dd21a8 | 1320 | inline bool |
9969c043 | 1321 | ssa_names_hasher::equal (const name_to_bb *n1, const name_to_bb *n2) |
d9dd21a8 | 1322 | { |
963aee26 | 1323 | return n1->ssa_name_ver == n2->ssa_name_ver |
1324 | && n1->store == n2->store | |
1325 | && n1->offset == n2->offset | |
1326 | && n1->size == n2->size; | |
e6d0e152 | 1327 | } |
1328 | ||
c1f445d2 | 1329 | class nontrapping_dom_walker : public dom_walker |
1330 | { | |
1331 | public: | |
431205b7 | 1332 | nontrapping_dom_walker (cdi_direction direction, hash_set<tree> *ps) |
c1f445d2 | 1333 | : dom_walker (direction), m_nontrapping (ps), m_seen_ssa_names (128) {} |
1334 | ||
1335 | virtual void before_dom_children (basic_block); | |
1336 | virtual void after_dom_children (basic_block); | |
1337 | ||
1338 | private: | |
1339 | ||
1340 | /* We see the expression EXP in basic block BB. If it's an interesting | |
1341 | expression (an MEM_REF through an SSA_NAME) possibly insert the | |
1342 | expression into the set NONTRAP or the hash table of seen expressions. | |
1343 | STORE is true if this expression is on the LHS, otherwise it's on | |
1344 | the RHS. */ | |
1345 | void add_or_mark_expr (basic_block, tree, bool); | |
1346 | ||
431205b7 | 1347 | hash_set<tree> *m_nontrapping; |
c1f445d2 | 1348 | |
1349 | /* The hash table for remembering what we've seen. */ | |
1350 | hash_table<ssa_names_hasher> m_seen_ssa_names; | |
1351 | }; | |
1352 | ||
1353 | /* Called by walk_dominator_tree, when entering the block BB. */ | |
1354 | void | |
1355 | nontrapping_dom_walker::before_dom_children (basic_block bb) | |
1356 | { | |
1357 | edge e; | |
1358 | edge_iterator ei; | |
1359 | gimple_stmt_iterator gsi; | |
1360 | ||
1361 | /* If we haven't seen all our predecessors, clear the hash-table. */ | |
1362 | FOR_EACH_EDGE (e, ei, bb->preds) | |
1363 | if ((((size_t)e->src->aux) & 2) == 0) | |
1364 | { | |
1365 | nt_call_phase++; | |
1366 | break; | |
1367 | } | |
1368 | ||
1369 | /* Mark this BB as being on the path to dominator root and as visited. */ | |
1370 | bb->aux = (void*)(1 | 2); | |
1371 | ||
1372 | /* And walk the statements in order. */ | |
1373 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1374 | { | |
1375 | gimple stmt = gsi_stmt (gsi); | |
1376 | ||
1377 | if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt)) | |
1378 | nt_call_phase++; | |
1379 | else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt)) | |
1380 | { | |
1381 | add_or_mark_expr (bb, gimple_assign_lhs (stmt), true); | |
1382 | add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), false); | |
1383 | } | |
1384 | } | |
1385 | } | |
1386 | ||
1387 | /* Called by walk_dominator_tree, when basic block BB is exited. */ | |
1388 | void | |
1389 | nontrapping_dom_walker::after_dom_children (basic_block bb) | |
1390 | { | |
1391 | /* This BB isn't on the path to dominator root anymore. */ | |
1392 | bb->aux = (void*)2; | |
1393 | } | |
d9dd21a8 | 1394 | |
f0b5f617 | 1395 | /* We see the expression EXP in basic block BB. If it's an interesting |
182cf5a9 | 1396 | expression (an MEM_REF through an SSA_NAME) possibly insert the |
af4f74fa | 1397 | expression into the set NONTRAP or the hash table of seen expressions. |
1398 | STORE is true if this expression is on the LHS, otherwise it's on | |
1399 | the RHS. */ | |
c1f445d2 | 1400 | void |
1401 | nontrapping_dom_walker::add_or_mark_expr (basic_block bb, tree exp, bool store) | |
e6d0e152 | 1402 | { |
963aee26 | 1403 | HOST_WIDE_INT size; |
1404 | ||
182cf5a9 | 1405 | if (TREE_CODE (exp) == MEM_REF |
963aee26 | 1406 | && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME |
e913b5cd | 1407 | && tree_fits_shwi_p (TREE_OPERAND (exp, 1)) |
963aee26 | 1408 | && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0) |
e6d0e152 | 1409 | { |
1410 | tree name = TREE_OPERAND (exp, 0); | |
1411 | struct name_to_bb map; | |
d9dd21a8 | 1412 | name_to_bb **slot; |
af4f74fa | 1413 | struct name_to_bb *n2bb; |
e6d0e152 | 1414 | basic_block found_bb = 0; |
1415 | ||
182cf5a9 | 1416 | /* Try to find the last seen MEM_REF through the same |
e6d0e152 | 1417 | SSA_NAME, which can trap. */ |
963aee26 | 1418 | map.ssa_name_ver = SSA_NAME_VERSION (name); |
42540642 | 1419 | map.phase = 0; |
e6d0e152 | 1420 | map.bb = 0; |
af4f74fa | 1421 | map.store = store; |
e913b5cd | 1422 | map.offset = tree_to_shwi (TREE_OPERAND (exp, 1)); |
963aee26 | 1423 | map.size = size; |
1424 | ||
c1f445d2 | 1425 | slot = m_seen_ssa_names.find_slot (&map, INSERT); |
d9dd21a8 | 1426 | n2bb = *slot; |
42540642 | 1427 | if (n2bb && n2bb->phase >= nt_call_phase) |
af4f74fa | 1428 | found_bb = n2bb->bb; |
e6d0e152 | 1429 | |
182cf5a9 | 1430 | /* If we've found a trapping MEM_REF, _and_ it dominates EXP |
e6d0e152 | 1431 | (it's in a basic block on the path from us to the dominator root) |
1432 | then we can't trap. */ | |
42540642 | 1433 | if (found_bb && (((size_t)found_bb->aux) & 1) == 1) |
e6d0e152 | 1434 | { |
431205b7 | 1435 | m_nontrapping->add (exp); |
e6d0e152 | 1436 | } |
1437 | else | |
1438 | { | |
1439 | /* EXP might trap, so insert it into the hash table. */ | |
af4f74fa | 1440 | if (n2bb) |
e6d0e152 | 1441 | { |
42540642 | 1442 | n2bb->phase = nt_call_phase; |
af4f74fa | 1443 | n2bb->bb = bb; |
e6d0e152 | 1444 | } |
1445 | else | |
1446 | { | |
af4f74fa | 1447 | n2bb = XNEW (struct name_to_bb); |
963aee26 | 1448 | n2bb->ssa_name_ver = SSA_NAME_VERSION (name); |
42540642 | 1449 | n2bb->phase = nt_call_phase; |
af4f74fa | 1450 | n2bb->bb = bb; |
1451 | n2bb->store = store; | |
963aee26 | 1452 | n2bb->offset = map.offset; |
1453 | n2bb->size = size; | |
af4f74fa | 1454 | *slot = n2bb; |
e6d0e152 | 1455 | } |
1456 | } | |
1457 | } | |
1458 | } | |
1459 | ||
e6d0e152 | 1460 | /* This is the entry point of gathering non trapping memory accesses. |
1461 | It will do a dominator walk over the whole function, and it will | |
1462 | make use of the bb->aux pointers. It returns a set of trees | |
182cf5a9 | 1463 | (the MEM_REFs itself) which can't trap. */ |
431205b7 | 1464 | static hash_set<tree> * |
e6d0e152 | 1465 | get_non_trapping (void) |
1466 | { | |
42540642 | 1467 | nt_call_phase = 0; |
431205b7 | 1468 | hash_set<tree> *nontrap = new hash_set<tree>; |
e6d0e152 | 1469 | /* We're going to do a dominator walk, so ensure that we have |
1470 | dominance information. */ | |
1471 | calculate_dominance_info (CDI_DOMINATORS); | |
1472 | ||
54c91640 | 1473 | nontrapping_dom_walker (CDI_DOMINATORS, nontrap) |
1474 | .walk (cfun->cfg->x_entry_block_ptr); | |
1475 | ||
42540642 | 1476 | clear_aux_for_blocks (); |
e6d0e152 | 1477 | return nontrap; |
1478 | } | |
1479 | ||
1480 | /* Do the main work of conditional store replacement. We already know | |
1481 | that the recognized pattern looks like so: | |
1482 | ||
1483 | split: | |
1484 | if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1) | |
1485 | MIDDLE_BB: | |
1486 | something | |
1487 | fallthrough (edge E0) | |
1488 | JOIN_BB: | |
1489 | some more | |
1490 | ||
1491 | We check that MIDDLE_BB contains only one store, that that store | |
1492 | doesn't trap (not via NOTRAP, but via checking if an access to the same | |
1493 | memory location dominates us) and that the store has a "simple" RHS. */ | |
1494 | ||
1495 | static bool | |
1496 | cond_store_replacement (basic_block middle_bb, basic_block join_bb, | |
431205b7 | 1497 | edge e0, edge e1, hash_set<tree> *nontrap) |
e6d0e152 | 1498 | { |
75a70cf9 | 1499 | gimple assign = last_and_only_stmt (middle_bb); |
03d37e4e | 1500 | tree lhs, rhs, name, name2; |
1a91d914 | 1501 | gphi *newphi; |
1502 | gassign *new_stmt; | |
75a70cf9 | 1503 | gimple_stmt_iterator gsi; |
efbcb6de | 1504 | source_location locus; |
e6d0e152 | 1505 | |
1506 | /* Check if middle_bb contains of only one store. */ | |
1507 | if (!assign | |
6cc085b6 | 1508 | || !gimple_assign_single_p (assign) |
1509 | || gimple_has_volatile_ops (assign)) | |
e6d0e152 | 1510 | return false; |
1511 | ||
efbcb6de | 1512 | locus = gimple_location (assign); |
75a70cf9 | 1513 | lhs = gimple_assign_lhs (assign); |
1514 | rhs = gimple_assign_rhs1 (assign); | |
182cf5a9 | 1515 | if (TREE_CODE (lhs) != MEM_REF |
91cf53d5 | 1516 | || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME |
3211fa0a | 1517 | || !is_gimple_reg_type (TREE_TYPE (lhs))) |
e6d0e152 | 1518 | return false; |
91cf53d5 | 1519 | |
e6d0e152 | 1520 | /* Prove that we can move the store down. We could also check |
1521 | TREE_THIS_NOTRAP here, but in that case we also could move stores, | |
1522 | whose value is not available readily, which we want to avoid. */ | |
431205b7 | 1523 | if (!nontrap->contains (lhs)) |
e6d0e152 | 1524 | return false; |
1525 | ||
1526 | /* Now we've checked the constraints, so do the transformation: | |
1527 | 1) Remove the single store. */ | |
75a70cf9 | 1528 | gsi = gsi_for_stmt (assign); |
3211fa0a | 1529 | unlink_stmt_vdef (assign); |
75a70cf9 | 1530 | gsi_remove (&gsi, true); |
91cf53d5 | 1531 | release_defs (assign); |
e6d0e152 | 1532 | |
03d37e4e | 1533 | /* 2) Insert a load from the memory of the store to the temporary |
e6d0e152 | 1534 | on the edge which did not contain the store. */ |
1535 | lhs = unshare_expr (lhs); | |
03d37e4e | 1536 | name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); |
1537 | new_stmt = gimple_build_assign (name, lhs); | |
efbcb6de | 1538 | gimple_set_location (new_stmt, locus); |
75a70cf9 | 1539 | gsi_insert_on_edge (e1, new_stmt); |
e6d0e152 | 1540 | |
03d37e4e | 1541 | /* 3) Create a PHI node at the join block, with one argument |
e6d0e152 | 1542 | holding the old RHS, and the other holding the temporary |
1543 | where we stored the old memory contents. */ | |
03d37e4e | 1544 | name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); |
1545 | newphi = create_phi_node (name2, join_bb); | |
60d535d2 | 1546 | add_phi_arg (newphi, rhs, e0, locus); |
1547 | add_phi_arg (newphi, name, e1, locus); | |
e6d0e152 | 1548 | |
1549 | lhs = unshare_expr (lhs); | |
75a70cf9 | 1550 | new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi)); |
e6d0e152 | 1551 | |
03d37e4e | 1552 | /* 4) Insert that PHI node. */ |
75a70cf9 | 1553 | gsi = gsi_after_labels (join_bb); |
1554 | if (gsi_end_p (gsi)) | |
e6d0e152 | 1555 | { |
75a70cf9 | 1556 | gsi = gsi_last_bb (join_bb); |
1557 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); | |
e6d0e152 | 1558 | } |
1559 | else | |
75a70cf9 | 1560 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); |
e6d0e152 | 1561 | |
1562 | return true; | |
1563 | } | |
4ee9c684 | 1564 | |
ec611e12 | 1565 | /* Do the main work of conditional store replacement. */ |
91cf53d5 | 1566 | |
1567 | static bool | |
ec611e12 | 1568 | cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb, |
1569 | basic_block join_bb, gimple then_assign, | |
1570 | gimple else_assign) | |
91cf53d5 | 1571 | { |
03d37e4e | 1572 | tree lhs_base, lhs, then_rhs, else_rhs, name; |
91cf53d5 | 1573 | source_location then_locus, else_locus; |
1574 | gimple_stmt_iterator gsi; | |
1a91d914 | 1575 | gphi *newphi; |
1576 | gassign *new_stmt; | |
91cf53d5 | 1577 | |
91cf53d5 | 1578 | if (then_assign == NULL |
1579 | || !gimple_assign_single_p (then_assign) | |
3c25489e | 1580 | || gimple_clobber_p (then_assign) |
6cc085b6 | 1581 | || gimple_has_volatile_ops (then_assign) |
91cf53d5 | 1582 | || else_assign == NULL |
3c25489e | 1583 | || !gimple_assign_single_p (else_assign) |
6cc085b6 | 1584 | || gimple_clobber_p (else_assign) |
1585 | || gimple_has_volatile_ops (else_assign)) | |
91cf53d5 | 1586 | return false; |
1587 | ||
1588 | lhs = gimple_assign_lhs (then_assign); | |
1589 | if (!is_gimple_reg_type (TREE_TYPE (lhs)) | |
1590 | || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0)) | |
1591 | return false; | |
1592 | ||
1593 | lhs_base = get_base_address (lhs); | |
1594 | if (lhs_base == NULL_TREE | |
1595 | || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF)) | |
1596 | return false; | |
1597 | ||
1598 | then_rhs = gimple_assign_rhs1 (then_assign); | |
1599 | else_rhs = gimple_assign_rhs1 (else_assign); | |
1600 | then_locus = gimple_location (then_assign); | |
1601 | else_locus = gimple_location (else_assign); | |
1602 | ||
1603 | /* Now we've checked the constraints, so do the transformation: | |
1604 | 1) Remove the stores. */ | |
1605 | gsi = gsi_for_stmt (then_assign); | |
1606 | unlink_stmt_vdef (then_assign); | |
1607 | gsi_remove (&gsi, true); | |
1608 | release_defs (then_assign); | |
1609 | ||
1610 | gsi = gsi_for_stmt (else_assign); | |
1611 | unlink_stmt_vdef (else_assign); | |
1612 | gsi_remove (&gsi, true); | |
1613 | release_defs (else_assign); | |
1614 | ||
03d37e4e | 1615 | /* 2) Create a PHI node at the join block, with one argument |
91cf53d5 | 1616 | holding the old RHS, and the other holding the temporary |
1617 | where we stored the old memory contents. */ | |
03d37e4e | 1618 | name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore"); |
1619 | newphi = create_phi_node (name, join_bb); | |
60d535d2 | 1620 | add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus); |
1621 | add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus); | |
91cf53d5 | 1622 | |
1623 | new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi)); | |
1624 | ||
03d37e4e | 1625 | /* 3) Insert that PHI node. */ |
91cf53d5 | 1626 | gsi = gsi_after_labels (join_bb); |
1627 | if (gsi_end_p (gsi)) | |
1628 | { | |
1629 | gsi = gsi_last_bb (join_bb); | |
1630 | gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT); | |
1631 | } | |
1632 | else | |
1633 | gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT); | |
1634 | ||
1635 | return true; | |
1636 | } | |
1637 | ||
ec611e12 | 1638 | /* Conditional store replacement. We already know |
1639 | that the recognized pattern looks like so: | |
1640 | ||
1641 | split: | |
1642 | if (cond) goto THEN_BB; else goto ELSE_BB (edge E1) | |
1643 | THEN_BB: | |
1644 | ... | |
1645 | X = Y; | |
1646 | ... | |
1647 | goto JOIN_BB; | |
1648 | ELSE_BB: | |
1649 | ... | |
1650 | X = Z; | |
1651 | ... | |
1652 | fallthrough (edge E0) | |
1653 | JOIN_BB: | |
1654 | some more | |
1655 | ||
1656 | We check that it is safe to sink the store to JOIN_BB by verifying that | |
1657 | there are no read-after-write or write-after-write dependencies in | |
1658 | THEN_BB and ELSE_BB. */ | |
1659 | ||
1660 | static bool | |
1661 | cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb, | |
1662 | basic_block join_bb) | |
1663 | { | |
1664 | gimple then_assign = last_and_only_stmt (then_bb); | |
1665 | gimple else_assign = last_and_only_stmt (else_bb); | |
f1f41a6c | 1666 | vec<data_reference_p> then_datarefs, else_datarefs; |
1667 | vec<ddr_p> then_ddrs, else_ddrs; | |
ec611e12 | 1668 | gimple then_store, else_store; |
1669 | bool found, ok = false, res; | |
1670 | struct data_dependence_relation *ddr; | |
1671 | data_reference_p then_dr, else_dr; | |
1672 | int i, j; | |
1673 | tree then_lhs, else_lhs; | |
ec611e12 | 1674 | basic_block blocks[3]; |
1675 | ||
1676 | if (MAX_STORES_TO_SINK == 0) | |
1677 | return false; | |
1678 | ||
1679 | /* Handle the case with single statement in THEN_BB and ELSE_BB. */ | |
1680 | if (then_assign && else_assign) | |
1681 | return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb, | |
1682 | then_assign, else_assign); | |
1683 | ||
1684 | /* Find data references. */ | |
f1f41a6c | 1685 | then_datarefs.create (1); |
1686 | else_datarefs.create (1); | |
ec611e12 | 1687 | if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs) |
1688 | == chrec_dont_know) | |
f1f41a6c | 1689 | || !then_datarefs.length () |
ec611e12 | 1690 | || (find_data_references_in_bb (NULL, else_bb, &else_datarefs) |
c71d3c24 | 1691 | == chrec_dont_know) |
f1f41a6c | 1692 | || !else_datarefs.length ()) |
ec611e12 | 1693 | { |
1694 | free_data_refs (then_datarefs); | |
1695 | free_data_refs (else_datarefs); | |
1696 | return false; | |
1697 | } | |
1698 | ||
1699 | /* Find pairs of stores with equal LHS. */ | |
4997014d | 1700 | auto_vec<gimple, 1> then_stores, else_stores; |
f1f41a6c | 1701 | FOR_EACH_VEC_ELT (then_datarefs, i, then_dr) |
ec611e12 | 1702 | { |
1703 | if (DR_IS_READ (then_dr)) | |
1704 | continue; | |
1705 | ||
1706 | then_store = DR_STMT (then_dr); | |
728dcc71 | 1707 | then_lhs = gimple_get_lhs (then_store); |
c71d3c24 | 1708 | if (then_lhs == NULL_TREE) |
1709 | continue; | |
ec611e12 | 1710 | found = false; |
1711 | ||
f1f41a6c | 1712 | FOR_EACH_VEC_ELT (else_datarefs, j, else_dr) |
ec611e12 | 1713 | { |
1714 | if (DR_IS_READ (else_dr)) | |
1715 | continue; | |
1716 | ||
1717 | else_store = DR_STMT (else_dr); | |
728dcc71 | 1718 | else_lhs = gimple_get_lhs (else_store); |
c71d3c24 | 1719 | if (else_lhs == NULL_TREE) |
1720 | continue; | |
ec611e12 | 1721 | |
1722 | if (operand_equal_p (then_lhs, else_lhs, 0)) | |
1723 | { | |
1724 | found = true; | |
1725 | break; | |
1726 | } | |
1727 | } | |
1728 | ||
1729 | if (!found) | |
1730 | continue; | |
1731 | ||
f1f41a6c | 1732 | then_stores.safe_push (then_store); |
1733 | else_stores.safe_push (else_store); | |
ec611e12 | 1734 | } |
1735 | ||
1736 | /* No pairs of stores found. */ | |
f1f41a6c | 1737 | if (!then_stores.length () |
1738 | || then_stores.length () > (unsigned) MAX_STORES_TO_SINK) | |
ec611e12 | 1739 | { |
1740 | free_data_refs (then_datarefs); | |
1741 | free_data_refs (else_datarefs); | |
ec611e12 | 1742 | return false; |
1743 | } | |
1744 | ||
1745 | /* Compute and check data dependencies in both basic blocks. */ | |
f1f41a6c | 1746 | then_ddrs.create (1); |
1747 | else_ddrs.create (1); | |
1748 | if (!compute_all_dependences (then_datarefs, &then_ddrs, | |
1e094109 | 1749 | vNULL, false) |
f1f41a6c | 1750 | || !compute_all_dependences (else_datarefs, &else_ddrs, |
1e094109 | 1751 | vNULL, false)) |
8b3fb720 | 1752 | { |
1753 | free_dependence_relations (then_ddrs); | |
1754 | free_dependence_relations (else_ddrs); | |
1755 | free_data_refs (then_datarefs); | |
1756 | free_data_refs (else_datarefs); | |
8b3fb720 | 1757 | return false; |
1758 | } | |
ec611e12 | 1759 | blocks[0] = then_bb; |
1760 | blocks[1] = else_bb; | |
1761 | blocks[2] = join_bb; | |
1762 | renumber_gimple_stmt_uids_in_blocks (blocks, 3); | |
1763 | ||
1764 | /* Check that there are no read-after-write or write-after-write dependencies | |
1765 | in THEN_BB. */ | |
f1f41a6c | 1766 | FOR_EACH_VEC_ELT (then_ddrs, i, ddr) |
ec611e12 | 1767 | { |
1768 | struct data_reference *dra = DDR_A (ddr); | |
1769 | struct data_reference *drb = DDR_B (ddr); | |
1770 | ||
1771 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known | |
1772 | && ((DR_IS_READ (dra) && DR_IS_WRITE (drb) | |
1773 | && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb))) | |
1774 | || (DR_IS_READ (drb) && DR_IS_WRITE (dra) | |
1775 | && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra))) | |
1776 | || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb)))) | |
1777 | { | |
1778 | free_dependence_relations (then_ddrs); | |
1779 | free_dependence_relations (else_ddrs); | |
2473bfb7 | 1780 | free_data_refs (then_datarefs); |
1781 | free_data_refs (else_datarefs); | |
ec611e12 | 1782 | return false; |
1783 | } | |
1784 | } | |
1785 | ||
1786 | /* Check that there are no read-after-write or write-after-write dependencies | |
1787 | in ELSE_BB. */ | |
f1f41a6c | 1788 | FOR_EACH_VEC_ELT (else_ddrs, i, ddr) |
ec611e12 | 1789 | { |
1790 | struct data_reference *dra = DDR_A (ddr); | |
1791 | struct data_reference *drb = DDR_B (ddr); | |
1792 | ||
1793 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known | |
1794 | && ((DR_IS_READ (dra) && DR_IS_WRITE (drb) | |
1795 | && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb))) | |
1796 | || (DR_IS_READ (drb) && DR_IS_WRITE (dra) | |
1797 | && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra))) | |
1798 | || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb)))) | |
1799 | { | |
1800 | free_dependence_relations (then_ddrs); | |
1801 | free_dependence_relations (else_ddrs); | |
2473bfb7 | 1802 | free_data_refs (then_datarefs); |
1803 | free_data_refs (else_datarefs); | |
ec611e12 | 1804 | return false; |
1805 | } | |
1806 | } | |
1807 | ||
1808 | /* Sink stores with same LHS. */ | |
f1f41a6c | 1809 | FOR_EACH_VEC_ELT (then_stores, i, then_store) |
ec611e12 | 1810 | { |
f1f41a6c | 1811 | else_store = else_stores[i]; |
ec611e12 | 1812 | res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb, |
1813 | then_store, else_store); | |
1814 | ok = ok || res; | |
1815 | } | |
1816 | ||
1817 | free_dependence_relations (then_ddrs); | |
1818 | free_dependence_relations (else_ddrs); | |
2473bfb7 | 1819 | free_data_refs (then_datarefs); |
1820 | free_data_refs (else_datarefs); | |
ec611e12 | 1821 | |
1822 | return ok; | |
1823 | } | |
1824 | ||
239e9670 | 1825 | /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */ |
1826 | ||
1827 | static bool | |
1828 | local_mem_dependence (gimple stmt, basic_block bb) | |
1829 | { | |
1830 | tree vuse = gimple_vuse (stmt); | |
1831 | gimple def; | |
1832 | ||
1833 | if (!vuse) | |
1834 | return false; | |
1835 | ||
1836 | def = SSA_NAME_DEF_STMT (vuse); | |
1837 | return (def && gimple_bb (def) == bb); | |
1838 | } | |
1839 | ||
1840 | /* Given a "diamond" control-flow pattern where BB0 tests a condition, | |
1841 | BB1 and BB2 are "then" and "else" blocks dependent on this test, | |
f32420fb | 1842 | and BB3 rejoins control flow following BB1 and BB2, look for |
239e9670 | 1843 | opportunities to hoist loads as follows. If BB3 contains a PHI of |
1844 | two loads, one each occurring in BB1 and BB2, and the loads are | |
1845 | provably of adjacent fields in the same structure, then move both | |
1846 | loads into BB0. Of course this can only be done if there are no | |
1847 | dependencies preventing such motion. | |
1848 | ||
1849 | One of the hoisted loads will always be speculative, so the | |
1850 | transformation is currently conservative: | |
1851 | ||
1852 | - The fields must be strictly adjacent. | |
1853 | - The two fields must occupy a single memory block that is | |
1854 | guaranteed to not cross a page boundary. | |
1855 | ||
1856 | The last is difficult to prove, as such memory blocks should be | |
1857 | aligned on the minimum of the stack alignment boundary and the | |
1858 | alignment guaranteed by heap allocation interfaces. Thus we rely | |
1859 | on a parameter for the alignment value. | |
1860 | ||
1861 | Provided a good value is used for the last case, the first | |
1862 | restriction could possibly be relaxed. */ | |
1863 | ||
1864 | static void | |
1865 | hoist_adjacent_loads (basic_block bb0, basic_block bb1, | |
1866 | basic_block bb2, basic_block bb3) | |
1867 | { | |
1868 | int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE); | |
1869 | unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT); | |
1a91d914 | 1870 | gphi_iterator gsi; |
239e9670 | 1871 | |
1872 | /* Walk the phis in bb3 looking for an opportunity. We are looking | |
1873 | for phis of two SSA names, one each of which is defined in bb1 and | |
1874 | bb2. */ | |
1875 | for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1876 | { | |
1a91d914 | 1877 | gphi *phi_stmt = gsi.phi (); |
dfcf26a5 | 1878 | gimple def1, def2; |
1879 | tree arg1, arg2, ref1, ref2, field1, field2; | |
239e9670 | 1880 | tree tree_offset1, tree_offset2, tree_size2, next; |
1881 | int offset1, offset2, size2; | |
1882 | unsigned align1; | |
1883 | gimple_stmt_iterator gsi2; | |
1884 | basic_block bb_for_def1, bb_for_def2; | |
1885 | ||
7c782c9b | 1886 | if (gimple_phi_num_args (phi_stmt) != 2 |
1887 | || virtual_operand_p (gimple_phi_result (phi_stmt))) | |
239e9670 | 1888 | continue; |
1889 | ||
1890 | arg1 = gimple_phi_arg_def (phi_stmt, 0); | |
1891 | arg2 = gimple_phi_arg_def (phi_stmt, 1); | |
f32420fb | 1892 | |
239e9670 | 1893 | if (TREE_CODE (arg1) != SSA_NAME |
1894 | || TREE_CODE (arg2) != SSA_NAME | |
1895 | || SSA_NAME_IS_DEFAULT_DEF (arg1) | |
7c782c9b | 1896 | || SSA_NAME_IS_DEFAULT_DEF (arg2)) |
239e9670 | 1897 | continue; |
1898 | ||
1899 | def1 = SSA_NAME_DEF_STMT (arg1); | |
1900 | def2 = SSA_NAME_DEF_STMT (arg2); | |
1901 | ||
1902 | if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2) | |
1903 | && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2)) | |
1904 | continue; | |
1905 | ||
1906 | /* Check the mode of the arguments to be sure a conditional move | |
1907 | can be generated for it. */ | |
935611bc | 1908 | if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1))) |
1909 | == CODE_FOR_nothing) | |
239e9670 | 1910 | continue; |
1911 | ||
1912 | /* Both statements must be assignments whose RHS is a COMPONENT_REF. */ | |
1913 | if (!gimple_assign_single_p (def1) | |
6cc085b6 | 1914 | || !gimple_assign_single_p (def2) |
1915 | || gimple_has_volatile_ops (def1) | |
1916 | || gimple_has_volatile_ops (def2)) | |
239e9670 | 1917 | continue; |
1918 | ||
1919 | ref1 = gimple_assign_rhs1 (def1); | |
1920 | ref2 = gimple_assign_rhs1 (def2); | |
1921 | ||
1922 | if (TREE_CODE (ref1) != COMPONENT_REF | |
1923 | || TREE_CODE (ref2) != COMPONENT_REF) | |
1924 | continue; | |
1925 | ||
1926 | /* The zeroth operand of the two component references must be | |
1927 | identical. It is not sufficient to compare get_base_address of | |
1928 | the two references, because this could allow for different | |
1929 | elements of the same array in the two trees. It is not safe to | |
1930 | assume that the existence of one array element implies the | |
1931 | existence of a different one. */ | |
1932 | if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0)) | |
1933 | continue; | |
1934 | ||
1935 | field1 = TREE_OPERAND (ref1, 1); | |
1936 | field2 = TREE_OPERAND (ref2, 1); | |
1937 | ||
1938 | /* Check for field adjacency, and ensure field1 comes first. */ | |
1939 | for (next = DECL_CHAIN (field1); | |
1940 | next && TREE_CODE (next) != FIELD_DECL; | |
1941 | next = DECL_CHAIN (next)) | |
1942 | ; | |
1943 | ||
1944 | if (next != field2) | |
1945 | { | |
1946 | for (next = DECL_CHAIN (field2); | |
1947 | next && TREE_CODE (next) != FIELD_DECL; | |
1948 | next = DECL_CHAIN (next)) | |
1949 | ; | |
1950 | ||
1951 | if (next != field1) | |
1952 | continue; | |
1953 | ||
dfcf26a5 | 1954 | std::swap (field1, field2); |
1955 | std::swap (def1, def2); | |
239e9670 | 1956 | } |
1957 | ||
7c74ee50 | 1958 | bb_for_def1 = gimple_bb (def1); |
1959 | bb_for_def2 = gimple_bb (def2); | |
1960 | ||
239e9670 | 1961 | /* Check for proper alignment of the first field. */ |
1962 | tree_offset1 = bit_position (field1); | |
1963 | tree_offset2 = bit_position (field2); | |
1964 | tree_size2 = DECL_SIZE (field2); | |
1965 | ||
e913b5cd | 1966 | if (!tree_fits_uhwi_p (tree_offset1) |
1967 | || !tree_fits_uhwi_p (tree_offset2) | |
1968 | || !tree_fits_uhwi_p (tree_size2)) | |
239e9670 | 1969 | continue; |
1970 | ||
e913b5cd | 1971 | offset1 = tree_to_uhwi (tree_offset1); |
1972 | offset2 = tree_to_uhwi (tree_offset2); | |
1973 | size2 = tree_to_uhwi (tree_size2); | |
239e9670 | 1974 | align1 = DECL_ALIGN (field1) % param_align_bits; |
1975 | ||
1976 | if (offset1 % BITS_PER_UNIT != 0) | |
1977 | continue; | |
1978 | ||
1979 | /* For profitability, the two field references should fit within | |
1980 | a single cache line. */ | |
1981 | if (align1 + offset2 - offset1 + size2 > param_align_bits) | |
1982 | continue; | |
1983 | ||
1984 | /* The two expressions cannot be dependent upon vdefs defined | |
1985 | in bb1/bb2. */ | |
1986 | if (local_mem_dependence (def1, bb_for_def1) | |
1987 | || local_mem_dependence (def2, bb_for_def2)) | |
1988 | continue; | |
1989 | ||
1990 | /* The conditions are satisfied; hoist the loads from bb1 and bb2 into | |
1991 | bb0. We hoist the first one first so that a cache miss is handled | |
1992 | efficiently regardless of hardware cache-fill policy. */ | |
1993 | gsi2 = gsi_for_stmt (def1); | |
1994 | gsi_move_to_bb_end (&gsi2, bb0); | |
1995 | gsi2 = gsi_for_stmt (def2); | |
1996 | gsi_move_to_bb_end (&gsi2, bb0); | |
1997 | ||
1998 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1999 | { | |
2000 | fprintf (dump_file, | |
2001 | "\nHoisting adjacent loads from %d and %d into %d: \n", | |
2002 | bb_for_def1->index, bb_for_def2->index, bb0->index); | |
2003 | print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS); | |
2004 | print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS); | |
2005 | } | |
2006 | } | |
2007 | } | |
2008 | ||
2009 | /* Determine whether we should attempt to hoist adjacent loads out of | |
2010 | diamond patterns in pass_phiopt. Always hoist loads if | |
2011 | -fhoist-adjacent-loads is specified and the target machine has | |
6f0ddab1 | 2012 | both a conditional move instruction and a defined cache line size. */ |
239e9670 | 2013 | |
2014 | static bool | |
2015 | gate_hoist_loads (void) | |
2016 | { | |
6f0ddab1 | 2017 | return (flag_hoist_adjacent_loads == 1 |
2018 | && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE) | |
2019 | && HAVE_conditional_move); | |
239e9670 | 2020 | } |
2021 | ||
65b0537f | 2022 | /* This pass tries to replaces an if-then-else block with an |
2023 | assignment. We have four kinds of transformations. Some of these | |
2024 | transformations are also performed by the ifcvt RTL optimizer. | |
2025 | ||
2026 | Conditional Replacement | |
2027 | ----------------------- | |
2028 | ||
2029 | This transformation, implemented in conditional_replacement, | |
2030 | replaces | |
2031 | ||
2032 | bb0: | |
2033 | if (cond) goto bb2; else goto bb1; | |
2034 | bb1: | |
2035 | bb2: | |
2036 | x = PHI <0 (bb1), 1 (bb0), ...>; | |
2037 | ||
2038 | with | |
2039 | ||
2040 | bb0: | |
2041 | x' = cond; | |
2042 | goto bb2; | |
2043 | bb2: | |
2044 | x = PHI <x' (bb0), ...>; | |
2045 | ||
2046 | We remove bb1 as it becomes unreachable. This occurs often due to | |
2047 | gimplification of conditionals. | |
2048 | ||
2049 | Value Replacement | |
2050 | ----------------- | |
2051 | ||
2052 | This transformation, implemented in value_replacement, replaces | |
2053 | ||
2054 | bb0: | |
2055 | if (a != b) goto bb2; else goto bb1; | |
2056 | bb1: | |
2057 | bb2: | |
2058 | x = PHI <a (bb1), b (bb0), ...>; | |
2059 | ||
2060 | with | |
2061 | ||
2062 | bb0: | |
2063 | bb2: | |
2064 | x = PHI <b (bb0), ...>; | |
2065 | ||
2066 | This opportunity can sometimes occur as a result of other | |
2067 | optimizations. | |
2068 | ||
2069 | ||
2070 | Another case caught by value replacement looks like this: | |
2071 | ||
2072 | bb0: | |
2073 | t1 = a == CONST; | |
2074 | t2 = b > c; | |
2075 | t3 = t1 & t2; | |
2076 | if (t3 != 0) goto bb1; else goto bb2; | |
2077 | bb1: | |
2078 | bb2: | |
2079 | x = PHI (CONST, a) | |
2080 | ||
2081 | Gets replaced with: | |
2082 | bb0: | |
2083 | bb2: | |
2084 | t1 = a == CONST; | |
2085 | t2 = b > c; | |
2086 | t3 = t1 & t2; | |
2087 | x = a; | |
2088 | ||
2089 | ABS Replacement | |
2090 | --------------- | |
2091 | ||
2092 | This transformation, implemented in abs_replacement, replaces | |
2093 | ||
2094 | bb0: | |
2095 | if (a >= 0) goto bb2; else goto bb1; | |
2096 | bb1: | |
2097 | x = -a; | |
2098 | bb2: | |
2099 | x = PHI <x (bb1), a (bb0), ...>; | |
2100 | ||
2101 | with | |
2102 | ||
2103 | bb0: | |
2104 | x' = ABS_EXPR< a >; | |
2105 | bb2: | |
2106 | x = PHI <x' (bb0), ...>; | |
2107 | ||
2108 | MIN/MAX Replacement | |
2109 | ------------------- | |
2110 | ||
2111 | This transformation, minmax_replacement replaces | |
2112 | ||
2113 | bb0: | |
2114 | if (a <= b) goto bb2; else goto bb1; | |
2115 | bb1: | |
2116 | bb2: | |
2117 | x = PHI <b (bb1), a (bb0), ...>; | |
2118 | ||
2119 | with | |
2120 | ||
2121 | bb0: | |
2122 | x' = MIN_EXPR (a, b) | |
2123 | bb2: | |
2124 | x = PHI <x' (bb0), ...>; | |
2125 | ||
2126 | A similar transformation is done for MAX_EXPR. | |
2127 | ||
2128 | ||
2129 | This pass also performs a fifth transformation of a slightly different | |
2130 | flavor. | |
2131 | ||
2132 | Adjacent Load Hoisting | |
2133 | ---------------------- | |
2134 | ||
2135 | This transformation replaces | |
2136 | ||
2137 | bb0: | |
2138 | if (...) goto bb2; else goto bb1; | |
2139 | bb1: | |
2140 | x1 = (<expr>).field1; | |
2141 | goto bb3; | |
2142 | bb2: | |
2143 | x2 = (<expr>).field2; | |
2144 | bb3: | |
2145 | # x = PHI <x1, x2>; | |
2146 | ||
2147 | with | |
2148 | ||
2149 | bb0: | |
2150 | x1 = (<expr>).field1; | |
2151 | x2 = (<expr>).field2; | |
2152 | if (...) goto bb2; else goto bb1; | |
2153 | bb1: | |
2154 | goto bb3; | |
2155 | bb2: | |
2156 | bb3: | |
2157 | # x = PHI <x1, x2>; | |
2158 | ||
2159 | The purpose of this transformation is to enable generation of conditional | |
2160 | move instructions such as Intel CMOVE or PowerPC ISEL. Because one of | |
2161 | the loads is speculative, the transformation is restricted to very | |
2162 | specific cases to avoid introducing a page fault. We are looking for | |
2163 | the common idiom: | |
2164 | ||
2165 | if (...) | |
2166 | x = y->left; | |
2167 | else | |
2168 | x = y->right; | |
2169 | ||
2170 | where left and right are typically adjacent pointers in a tree structure. */ | |
20e5647c | 2171 | |
cbe8bda8 | 2172 | namespace { |
2173 | ||
2174 | const pass_data pass_data_phiopt = | |
4ee9c684 | 2175 | { |
cbe8bda8 | 2176 | GIMPLE_PASS, /* type */ |
2177 | "phiopt", /* name */ | |
2178 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 2179 | TV_TREE_PHIOPT, /* tv_id */ |
2180 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2181 | 0, /* properties_provided */ | |
2182 | 0, /* properties_destroyed */ | |
2183 | 0, /* todo_flags_start */ | |
8b88439e | 2184 | 0, /* todo_flags_finish */ |
4ee9c684 | 2185 | }; |
e6d0e152 | 2186 | |
cbe8bda8 | 2187 | class pass_phiopt : public gimple_opt_pass |
2188 | { | |
2189 | public: | |
9af5ce0c | 2190 | pass_phiopt (gcc::context *ctxt) |
2191 | : gimple_opt_pass (pass_data_phiopt, ctxt) | |
cbe8bda8 | 2192 | {} |
2193 | ||
2194 | /* opt_pass methods: */ | |
ae84f584 | 2195 | opt_pass * clone () { return new pass_phiopt (m_ctxt); } |
3dac50cc | 2196 | virtual bool gate (function *) { return flag_ssa_phiopt; } |
65b0537f | 2197 | virtual unsigned int execute (function *) |
2198 | { | |
2199 | return tree_ssa_phiopt_worker (false, gate_hoist_loads ()); | |
2200 | } | |
cbe8bda8 | 2201 | |
2202 | }; // class pass_phiopt | |
2203 | ||
2204 | } // anon namespace | |
2205 | ||
2206 | gimple_opt_pass * | |
2207 | make_pass_phiopt (gcc::context *ctxt) | |
2208 | { | |
2209 | return new pass_phiopt (ctxt); | |
2210 | } | |
2211 | ||
cbe8bda8 | 2212 | namespace { |
2213 | ||
2214 | const pass_data pass_data_cselim = | |
e6d0e152 | 2215 | { |
cbe8bda8 | 2216 | GIMPLE_PASS, /* type */ |
2217 | "cselim", /* name */ | |
2218 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 2219 | TV_TREE_PHIOPT, /* tv_id */ |
2220 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2221 | 0, /* properties_provided */ | |
2222 | 0, /* properties_destroyed */ | |
2223 | 0, /* todo_flags_start */ | |
8b88439e | 2224 | 0, /* todo_flags_finish */ |
e6d0e152 | 2225 | }; |
cbe8bda8 | 2226 | |
2227 | class pass_cselim : public gimple_opt_pass | |
2228 | { | |
2229 | public: | |
9af5ce0c | 2230 | pass_cselim (gcc::context *ctxt) |
2231 | : gimple_opt_pass (pass_data_cselim, ctxt) | |
cbe8bda8 | 2232 | {} |
2233 | ||
2234 | /* opt_pass methods: */ | |
31315c24 | 2235 | virtual bool gate (function *) { return flag_tree_cselim; } |
65b0537f | 2236 | virtual unsigned int execute (function *) { return tree_ssa_cs_elim (); } |
cbe8bda8 | 2237 | |
2238 | }; // class pass_cselim | |
2239 | ||
2240 | } // anon namespace | |
2241 | ||
2242 | gimple_opt_pass * | |
2243 | make_pass_cselim (gcc::context *ctxt) | |
2244 | { | |
2245 | return new pass_cselim (ctxt); | |
2246 | } |