1 | /* |
2 | * Copyright (c) 2011, 2018, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "opto/loopnode.hpp" |
27 | #include "opto/addnode.hpp" |
28 | #include "opto/callnode.hpp" |
29 | #include "opto/connode.hpp" |
30 | #include "opto/convertnode.hpp" |
31 | #include "opto/loopnode.hpp" |
32 | #include "opto/matcher.hpp" |
33 | #include "opto/mulnode.hpp" |
34 | #include "opto/opaquenode.hpp" |
35 | #include "opto/rootnode.hpp" |
36 | #include "opto/subnode.hpp" |
37 | #include <fenv.h> |
38 | #include <math.h> |
39 | |
40 | /* |
41 | * The general idea of Loop Predication is to insert a predicate on the entry |
42 | * path to a loop, and raise a uncommon trap if the check of the condition fails. |
43 | * The condition checks are promoted from inside the loop body, and thus |
44 | * the checks inside the loop could be eliminated. Currently, loop predication |
45 | * optimization has been applied to remove array range check and loop invariant |
46 | * checks (such as null checks). |
47 | */ |
48 | |
49 | //-------------------------------register_control------------------------- |
50 | void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) { |
51 | assert(n->is_CFG(), "must be control node" ); |
52 | _igvn.register_new_node_with_optimizer(n); |
53 | loop->_body.push(n); |
54 | set_loop(n, loop); |
55 | // When called from beautify_loops() idom is not constructed yet. |
56 | if (_idom != NULL) { |
57 | set_idom(n, pred, dom_depth(pred)); |
58 | } |
59 | } |
60 | |
61 | //------------------------------create_new_if_for_predicate------------------------ |
62 | // create a new if above the uct_if_pattern for the predicate to be promoted. |
63 | // |
64 | // before after |
65 | // ---------- ---------- |
66 | // ctrl ctrl |
67 | // | | |
68 | // | | |
69 | // v v |
70 | // iff new_iff |
71 | // / \ / \ |
72 | // / \ / \ |
73 | // v v v v |
74 | // uncommon_proj cont_proj if_uct if_cont |
75 | // \ | | | | |
76 | // \ | | | | |
77 | // v v v | v |
78 | // rgn loop | iff |
79 | // | | / \ |
80 | // | | / \ |
81 | // v | v v |
82 | // uncommon_trap | uncommon_proj cont_proj |
83 | // \ \ | | |
84 | // \ \ | | |
85 | // v v v v |
86 | // rgn loop |
87 | // | |
88 | // | |
89 | // v |
90 | // uncommon_trap |
91 | // |
92 | // |
93 | // We will create a region to guard the uct call if there is no one there. |
94 | // The true projection (if_cont) of the new_iff is returned. |
95 | // This code is also used to clone predicates to cloned loops. |
96 | ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, |
97 | Deoptimization::DeoptReason reason, |
98 | int opcode) { |
99 | assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!" ); |
100 | IfNode* iff = cont_proj->in(0)->as_If(); |
101 | |
102 | ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); |
103 | Node *rgn = uncommon_proj->unique_ctrl_out(); |
104 | assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct" ); |
105 | |
106 | uint proj_index = 1; // region's edge corresponding to uncommon_proj |
107 | if (!rgn->is_Region()) { // create a region to guard the call |
108 | assert(rgn->is_Call(), "must be call uct" ); |
109 | CallNode* call = rgn->as_Call(); |
110 | IdealLoopTree* loop = get_loop(call); |
111 | rgn = new RegionNode(1); |
112 | rgn->add_req(uncommon_proj); |
113 | register_control(rgn, loop, uncommon_proj); |
114 | _igvn.replace_input_of(call, 0, rgn); |
115 | // When called from beautify_loops() idom is not constructed yet. |
116 | if (_idom != NULL) { |
117 | set_idom(call, rgn, dom_depth(rgn)); |
118 | } |
119 | for (DUIterator_Fast imax, i = uncommon_proj->fast_outs(imax); i < imax; i++) { |
120 | Node* n = uncommon_proj->fast_out(i); |
121 | if (n->is_Load() || n->is_Store()) { |
122 | _igvn.replace_input_of(n, 0, rgn); |
123 | --i; --imax; |
124 | } |
125 | } |
126 | } else { |
127 | // Find region's edge corresponding to uncommon_proj |
128 | for (; proj_index < rgn->req(); proj_index++) |
129 | if (rgn->in(proj_index) == uncommon_proj) break; |
130 | assert(proj_index < rgn->req(), "sanity" ); |
131 | } |
132 | |
133 | Node* entry = iff->in(0); |
134 | if (new_entry != NULL) { |
135 | // Clonning the predicate to new location. |
136 | entry = new_entry; |
137 | } |
138 | // Create new_iff |
139 | IdealLoopTree* lp = get_loop(entry); |
140 | IfNode* new_iff = NULL; |
141 | if (opcode == Op_If) { |
142 | new_iff = new IfNode(entry, iff->in(1), iff->_prob, iff->_fcnt); |
143 | } else { |
144 | assert(opcode == Op_RangeCheck, "no other if variant here" ); |
145 | new_iff = new RangeCheckNode(entry, iff->in(1), iff->_prob, iff->_fcnt); |
146 | } |
147 | register_control(new_iff, lp, entry); |
148 | Node *if_cont = new IfTrueNode(new_iff); |
149 | Node *if_uct = new IfFalseNode(new_iff); |
150 | if (cont_proj->is_IfFalse()) { |
151 | // Swap |
152 | Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; |
153 | } |
154 | register_control(if_cont, lp, new_iff); |
155 | register_control(if_uct, get_loop(rgn), new_iff); |
156 | |
157 | // if_uct to rgn |
158 | _igvn.hash_delete(rgn); |
159 | rgn->add_req(if_uct); |
160 | // When called from beautify_loops() idom is not constructed yet. |
161 | if (_idom != NULL) { |
162 | Node* ridom = idom(rgn); |
163 | Node* nrdom = dom_lca_internal(ridom, new_iff); |
164 | set_idom(rgn, nrdom, dom_depth(rgn)); |
165 | } |
166 | |
167 | // If rgn has phis add new edges which has the same |
168 | // value as on original uncommon_proj pass. |
169 | assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last" ); |
170 | bool has_phi = false; |
171 | for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) { |
172 | Node* use = rgn->fast_out(i); |
173 | if (use->is_Phi() && use->outcnt() > 0) { |
174 | assert(use->in(0) == rgn, "" ); |
175 | _igvn.rehash_node_delayed(use); |
176 | use->add_req(use->in(proj_index)); |
177 | has_phi = true; |
178 | } |
179 | } |
180 | assert(!has_phi || rgn->req() > 3, "no phis when region is created" ); |
181 | |
182 | if (new_entry == NULL) { |
183 | // Attach if_cont to iff |
184 | _igvn.replace_input_of(iff, 0, if_cont); |
185 | if (_idom != NULL) { |
186 | set_idom(iff, if_cont, dom_depth(iff)); |
187 | } |
188 | } |
189 | return if_cont->as_Proj(); |
190 | } |
191 | |
192 | //------------------------------create_new_if_for_predicate------------------------ |
193 | // Create a new if below new_entry for the predicate to be cloned (IGVN optimization) |
194 | ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, |
195 | Deoptimization::DeoptReason reason, |
196 | int opcode) { |
197 | assert(new_entry != 0, "only used for clone predicate" ); |
198 | assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!" ); |
199 | IfNode* iff = cont_proj->in(0)->as_If(); |
200 | |
201 | ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); |
202 | Node *rgn = uncommon_proj->unique_ctrl_out(); |
203 | assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct" ); |
204 | |
205 | uint proj_index = 1; // region's edge corresponding to uncommon_proj |
206 | if (!rgn->is_Region()) { // create a region to guard the call |
207 | assert(rgn->is_Call(), "must be call uct" ); |
208 | CallNode* call = rgn->as_Call(); |
209 | rgn = new RegionNode(1); |
210 | register_new_node_with_optimizer(rgn); |
211 | rgn->add_req(uncommon_proj); |
212 | replace_input_of(call, 0, rgn); |
213 | } else { |
214 | // Find region's edge corresponding to uncommon_proj |
215 | for (; proj_index < rgn->req(); proj_index++) |
216 | if (rgn->in(proj_index) == uncommon_proj) break; |
217 | assert(proj_index < rgn->req(), "sanity" ); |
218 | } |
219 | |
220 | // Create new_iff in new location. |
221 | IfNode* new_iff = NULL; |
222 | if (opcode == Op_If) { |
223 | new_iff = new IfNode(new_entry, iff->in(1), iff->_prob, iff->_fcnt); |
224 | } else { |
225 | assert(opcode == Op_RangeCheck, "no other if variant here" ); |
226 | new_iff = new RangeCheckNode(new_entry, iff->in(1), iff->_prob, iff->_fcnt); |
227 | } |
228 | |
229 | register_new_node_with_optimizer(new_iff); |
230 | Node *if_cont = new IfTrueNode(new_iff); |
231 | Node *if_uct = new IfFalseNode(new_iff); |
232 | if (cont_proj->is_IfFalse()) { |
233 | // Swap |
234 | Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; |
235 | } |
236 | register_new_node_with_optimizer(if_cont); |
237 | register_new_node_with_optimizer(if_uct); |
238 | |
239 | // if_uct to rgn |
240 | hash_delete(rgn); |
241 | rgn->add_req(if_uct); |
242 | |
243 | // If rgn has phis add corresponding new edges which has the same |
244 | // value as on original uncommon_proj pass. |
245 | assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last" ); |
246 | bool has_phi = false; |
247 | for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) { |
248 | Node* use = rgn->fast_out(i); |
249 | if (use->is_Phi() && use->outcnt() > 0) { |
250 | rehash_node_delayed(use); |
251 | use->add_req(use->in(proj_index)); |
252 | has_phi = true; |
253 | } |
254 | } |
255 | assert(!has_phi || rgn->req() > 3, "no phis when region is created" ); |
256 | |
257 | return if_cont->as_Proj(); |
258 | } |
259 | |
260 | //--------------------------clone_predicate----------------------- |
261 | ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry, |
262 | Deoptimization::DeoptReason reason, |
263 | PhaseIdealLoop* loop_phase, |
264 | PhaseIterGVN* igvn) { |
265 | ProjNode* new_predicate_proj; |
266 | if (loop_phase != NULL) { |
267 | new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason, Op_If); |
268 | } else { |
269 | new_predicate_proj = igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason, Op_If); |
270 | } |
271 | IfNode* iff = new_predicate_proj->in(0)->as_If(); |
272 | Node* ctrl = iff->in(0); |
273 | |
274 | // Match original condition since predicate's projections could be swapped. |
275 | assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be" ); |
276 | Node* opq = new Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1)); |
277 | igvn->C->add_predicate_opaq(opq); |
278 | |
279 | Node* bol = new Conv2BNode(opq); |
280 | if (loop_phase != NULL) { |
281 | loop_phase->register_new_node(opq, ctrl); |
282 | loop_phase->register_new_node(bol, ctrl); |
283 | } else { |
284 | igvn->register_new_node_with_optimizer(opq); |
285 | igvn->register_new_node_with_optimizer(bol); |
286 | } |
287 | igvn->hash_delete(iff); |
288 | iff->set_req(1, bol); |
289 | return new_predicate_proj; |
290 | } |
291 | |
292 | |
293 | //--------------------------clone_loop_predicates----------------------- |
294 | // Interface from IGVN |
295 | Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) { |
296 | return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, clone_limit_check, NULL, this); |
297 | } |
298 | |
299 | // Interface from PhaseIdealLoop |
300 | Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) { |
301 | return clone_loop_predicates(old_entry, new_entry, clone_limit_check, this, &this->_igvn); |
302 | } |
303 | |
304 | void PhaseIdealLoop::clone_loop_predicates_fix_mem(ProjNode* dom_proj , ProjNode* proj, |
305 | PhaseIdealLoop* loop_phase, |
306 | PhaseIterGVN* igvn) { |
307 | Compile* C = NULL; |
308 | if (loop_phase != NULL) { |
309 | igvn = &loop_phase->igvn(); |
310 | } |
311 | C = igvn->C; |
312 | ProjNode* other_dom_proj = dom_proj->in(0)->as_Multi()->proj_out(1-dom_proj->_con); |
313 | Node* dom_r = other_dom_proj->unique_ctrl_out(); |
314 | if (dom_r->is_Region()) { |
315 | assert(dom_r->unique_ctrl_out()->is_Call(), "unc expected" ); |
316 | ProjNode* other_proj = proj->in(0)->as_Multi()->proj_out(1-proj->_con); |
317 | Node* r = other_proj->unique_ctrl_out(); |
318 | assert(r->is_Region() && r->unique_ctrl_out()->is_Call(), "cloned predicate should have caused region to be added" ); |
319 | for (DUIterator_Fast imax, i = dom_r->fast_outs(imax); i < imax; i++) { |
320 | Node* dom_use = dom_r->fast_out(i); |
321 | if (dom_use->is_Phi() && dom_use->bottom_type() == Type::MEMORY) { |
322 | assert(dom_use->in(0) == dom_r, "" ); |
323 | Node* phi = NULL; |
324 | for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) { |
325 | Node* use = r->fast_out(j); |
326 | if (use->is_Phi() && use->bottom_type() == Type::MEMORY && |
327 | use->adr_type() == dom_use->adr_type()) { |
328 | assert(use->in(0) == r, "" ); |
329 | assert(phi == NULL, "only one phi" ); |
330 | phi = use; |
331 | } |
332 | } |
333 | if (phi == NULL) { |
334 | const TypePtr* adr_type = dom_use->adr_type(); |
335 | int alias = C->get_alias_index(adr_type); |
336 | Node* call = r->unique_ctrl_out(); |
337 | Node* mem = call->in(TypeFunc::Memory); |
338 | MergeMemNode* mm = NULL; |
339 | if (mem->is_MergeMem()) { |
340 | mm = mem->clone()->as_MergeMem(); |
341 | if (adr_type == TypePtr::BOTTOM) { |
342 | mem = mem->as_MergeMem()->base_memory(); |
343 | } else { |
344 | mem = mem->as_MergeMem()->memory_at(alias); |
345 | } |
346 | } else { |
347 | mm = MergeMemNode::make(mem); |
348 | } |
349 | phi = PhiNode::make(r, mem, Type::MEMORY, adr_type); |
350 | if (adr_type == TypePtr::BOTTOM) { |
351 | mm->set_base_memory(phi); |
352 | } else { |
353 | mm->set_memory_at(alias, phi); |
354 | } |
355 | if (loop_phase != NULL) { |
356 | loop_phase->register_new_node(mm, r); |
357 | loop_phase->register_new_node(phi, r); |
358 | } else { |
359 | igvn->register_new_node_with_optimizer(mm); |
360 | igvn->register_new_node_with_optimizer(phi); |
361 | } |
362 | igvn->replace_input_of(call, TypeFunc::Memory, mm); |
363 | } |
364 | igvn->replace_input_of(phi, r->find_edge(other_proj), dom_use->in(dom_r->find_edge(other_dom_proj))); |
365 | } |
366 | } |
367 | } |
368 | } |
369 | |
370 | |
371 | // Clone loop predicates to cloned loops (peeled, unswitched, split_if). |
372 | Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, |
373 | bool clone_limit_check, |
374 | PhaseIdealLoop* loop_phase, |
375 | PhaseIterGVN* igvn) { |
376 | #ifdef ASSERT |
377 | if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) { |
378 | if (new_entry != NULL) |
379 | new_entry->dump(); |
380 | assert(false, "not IfTrue, IfFalse, Region or SafePoint" ); |
381 | } |
382 | #endif |
383 | // Search original predicates |
384 | Node* entry = old_entry; |
385 | ProjNode* limit_check_proj = NULL; |
386 | limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); |
387 | if (limit_check_proj != NULL) { |
388 | entry = skip_loop_predicates(entry); |
389 | } |
390 | ProjNode* profile_predicate_proj = NULL; |
391 | ProjNode* predicate_proj = NULL; |
392 | if (UseProfiledLoopPredicate) { |
393 | profile_predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate); |
394 | if (profile_predicate_proj != NULL) { |
395 | entry = skip_loop_predicates(entry); |
396 | } |
397 | } |
398 | if (UseLoopPredicate) { |
399 | predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); |
400 | } |
401 | if (predicate_proj != NULL) { // right pattern that can be used by loop predication |
402 | // clone predicate |
403 | ProjNode* proj = clone_predicate(predicate_proj, new_entry, |
404 | Deoptimization::Reason_predicate, |
405 | loop_phase, igvn); |
406 | assert(proj != NULL, "IfTrue or IfFalse after clone predicate" ); |
407 | new_entry = proj; |
408 | if (TraceLoopPredicate) { |
409 | tty->print("Loop Predicate cloned: " ); |
410 | debug_only( new_entry->in(0)->dump(); ); |
411 | } |
412 | if (profile_predicate_proj != NULL) { |
413 | // A node that produces memory may be out of loop and depend on |
414 | // a profiled predicates. In that case the memory state at the |
415 | // end of profiled predicates and at the end of predicates are |
416 | // not the same. The cloned predicates are dominated by the |
417 | // profiled predicates but may have the wrong memory |
418 | // state. Update it. |
419 | clone_loop_predicates_fix_mem(profile_predicate_proj, proj, loop_phase, igvn); |
420 | } |
421 | } |
422 | if (profile_predicate_proj != NULL) { // right pattern that can be used by loop predication |
423 | // clone predicate |
424 | new_entry = clone_predicate(profile_predicate_proj, new_entry, |
425 | Deoptimization::Reason_profile_predicate, |
426 | loop_phase, igvn); |
427 | assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate" ); |
428 | if (TraceLoopPredicate) { |
429 | tty->print("Loop Predicate cloned: " ); |
430 | debug_only( new_entry->in(0)->dump(); ); |
431 | } |
432 | } |
433 | if (limit_check_proj != NULL && clone_limit_check) { |
434 | // Clone loop limit check last to insert it before loop. |
435 | // Don't clone a limit check which was already finalized |
436 | // for this counted loop (only one limit check is needed). |
437 | new_entry = clone_predicate(limit_check_proj, new_entry, |
438 | Deoptimization::Reason_loop_limit_check, |
439 | loop_phase, igvn); |
440 | assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check" ); |
441 | if (TraceLoopLimitCheck) { |
442 | tty->print("Loop Limit Check cloned: " ); |
443 | debug_only( new_entry->in(0)->dump(); ) |
444 | } |
445 | } |
446 | return new_entry; |
447 | } |
448 | |
449 | //--------------------------skip_loop_predicates------------------------------ |
450 | // Skip related predicates. |
451 | Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) { |
452 | IfNode* iff = entry->in(0)->as_If(); |
453 | ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con); |
454 | Node* rgn = uncommon_proj->unique_ctrl_out(); |
455 | assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct" ); |
456 | entry = entry->in(0)->in(0); |
457 | while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) { |
458 | uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con); |
459 | if (uncommon_proj->unique_ctrl_out() != rgn) |
460 | break; |
461 | entry = entry->in(0)->in(0); |
462 | } |
463 | return entry; |
464 | } |
465 | |
466 | Node* PhaseIdealLoop::skip_all_loop_predicates(Node* entry) { |
467 | Node* predicate = NULL; |
468 | predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); |
469 | if (predicate != NULL) { |
470 | entry = skip_loop_predicates(entry); |
471 | } |
472 | if (UseProfiledLoopPredicate) { |
473 | predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate); |
474 | if (predicate != NULL) { // right pattern that can be used by loop predication |
475 | entry = skip_loop_predicates(entry); |
476 | } |
477 | } |
478 | if (UseLoopPredicate) { |
479 | predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); |
480 | if (predicate != NULL) { // right pattern that can be used by loop predication |
481 | entry = skip_loop_predicates(entry); |
482 | } |
483 | } |
484 | return entry; |
485 | } |
486 | |
487 | //--------------------------find_predicate_insertion_point------------------- |
488 | // Find a good location to insert a predicate |
489 | ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) { |
490 | if (start_c == NULL || !start_c->is_Proj()) |
491 | return NULL; |
492 | if (start_c->as_Proj()->is_uncommon_trap_if_pattern(reason)) { |
493 | return start_c->as_Proj(); |
494 | } |
495 | return NULL; |
496 | } |
497 | |
498 | //--------------------------find_predicate------------------------------------ |
499 | // Find a predicate |
500 | Node* PhaseIdealLoop::find_predicate(Node* entry) { |
501 | Node* predicate = NULL; |
502 | predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); |
503 | if (predicate != NULL) { // right pattern that can be used by loop predication |
504 | return entry; |
505 | } |
506 | if (UseLoopPredicate) { |
507 | predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); |
508 | if (predicate != NULL) { // right pattern that can be used by loop predication |
509 | return entry; |
510 | } |
511 | } |
512 | if (UseProfiledLoopPredicate) { |
513 | predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate); |
514 | if (predicate != NULL) { // right pattern that can be used by loop predication |
515 | return entry; |
516 | } |
517 | } |
518 | return NULL; |
519 | } |
520 | |
521 | //------------------------------Invariance----------------------------------- |
522 | // Helper class for loop_predication_impl to compute invariance on the fly and |
523 | // clone invariants. |
524 | class Invariance : public StackObj { |
525 | VectorSet _visited, _invariant; |
526 | Node_Stack _stack; |
527 | VectorSet _clone_visited; |
528 | Node_List _old_new; // map of old to new (clone) |
529 | IdealLoopTree* _lpt; |
530 | PhaseIdealLoop* _phase; |
531 | |
532 | // Helper function to set up the invariance for invariance computation |
533 | // If n is a known invariant, set up directly. Otherwise, look up the |
534 | // the possibility to push n onto the stack for further processing. |
535 | void visit(Node* use, Node* n) { |
536 | if (_lpt->is_invariant(n)) { // known invariant |
537 | _invariant.set(n->_idx); |
538 | } else if (!n->is_CFG()) { |
539 | Node *n_ctrl = _phase->ctrl_or_self(n); |
540 | Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG |
541 | if (_phase->is_dominator(n_ctrl, u_ctrl)) { |
542 | _stack.push(n, n->in(0) == NULL ? 1 : 0); |
543 | } |
544 | } |
545 | } |
546 | |
547 | // Compute invariance for "the_node" and (possibly) all its inputs recursively |
548 | // on the fly |
549 | void compute_invariance(Node* n) { |
550 | assert(_visited.test(n->_idx), "must be" ); |
551 | visit(n, n); |
552 | while (_stack.is_nonempty()) { |
553 | Node* n = _stack.node(); |
554 | uint idx = _stack.index(); |
555 | if (idx == n->req()) { // all inputs are processed |
556 | _stack.pop(); |
557 | // n is invariant if it's inputs are all invariant |
558 | bool all_inputs_invariant = true; |
559 | for (uint i = 0; i < n->req(); i++) { |
560 | Node* in = n->in(i); |
561 | if (in == NULL) continue; |
562 | assert(_visited.test(in->_idx), "must have visited input" ); |
563 | if (!_invariant.test(in->_idx)) { // bad guy |
564 | all_inputs_invariant = false; |
565 | break; |
566 | } |
567 | } |
568 | if (all_inputs_invariant) { |
569 | // If n's control is a predicate that was moved out of the |
570 | // loop, it was marked invariant but n is only invariant if |
571 | // it depends only on that test. Otherwise, unless that test |
572 | // is out of the loop, it's not invariant. |
573 | if (n->is_CFG() || n->depends_only_on_test() || n->in(0) == NULL || !_phase->is_member(_lpt, n->in(0))) { |
574 | _invariant.set(n->_idx); // I am a invariant too |
575 | } |
576 | } |
577 | } else { // process next input |
578 | _stack.set_index(idx + 1); |
579 | Node* m = n->in(idx); |
580 | if (m != NULL && !_visited.test_set(m->_idx)) { |
581 | visit(n, m); |
582 | } |
583 | } |
584 | } |
585 | } |
586 | |
587 | // Helper function to set up _old_new map for clone_nodes. |
588 | // If n is a known invariant, set up directly ("clone" of n == n). |
589 | // Otherwise, push n onto the stack for real cloning. |
590 | void clone_visit(Node* n) { |
591 | assert(_invariant.test(n->_idx), "must be invariant" ); |
592 | if (_lpt->is_invariant(n)) { // known invariant |
593 | _old_new.map(n->_idx, n); |
594 | } else { // to be cloned |
595 | assert(!n->is_CFG(), "should not see CFG here" ); |
596 | _stack.push(n, n->in(0) == NULL ? 1 : 0); |
597 | } |
598 | } |
599 | |
600 | // Clone "n" and (possibly) all its inputs recursively |
601 | void clone_nodes(Node* n, Node* ctrl) { |
602 | clone_visit(n); |
603 | while (_stack.is_nonempty()) { |
604 | Node* n = _stack.node(); |
605 | uint idx = _stack.index(); |
606 | if (idx == n->req()) { // all inputs processed, clone n! |
607 | _stack.pop(); |
608 | // clone invariant node |
609 | Node* n_cl = n->clone(); |
610 | _old_new.map(n->_idx, n_cl); |
611 | _phase->register_new_node(n_cl, ctrl); |
612 | for (uint i = 0; i < n->req(); i++) { |
613 | Node* in = n_cl->in(i); |
614 | if (in == NULL) continue; |
615 | n_cl->set_req(i, _old_new[in->_idx]); |
616 | } |
617 | } else { // process next input |
618 | _stack.set_index(idx + 1); |
619 | Node* m = n->in(idx); |
620 | if (m != NULL && !_clone_visited.test_set(m->_idx)) { |
621 | clone_visit(m); // visit the input |
622 | } |
623 | } |
624 | } |
625 | } |
626 | |
627 | public: |
628 | Invariance(Arena* area, IdealLoopTree* lpt) : |
629 | _visited(area), _invariant(area), |
630 | _stack(area, 10 /* guess */), |
631 | _clone_visited(area), _old_new(area), |
632 | _lpt(lpt), _phase(lpt->_phase) |
633 | { |
634 | LoopNode* head = _lpt->_head->as_Loop(); |
635 | Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl); |
636 | if (entry->outcnt() != 1) { |
637 | // If a node is pinned between the predicates and the loop |
638 | // entry, we won't be able to move any node in the loop that |
639 | // depends on it above it in a predicate. Mark all those nodes |
640 | // as non loop invariatnt. |
641 | Unique_Node_List wq; |
642 | wq.push(entry); |
643 | for (uint next = 0; next < wq.size(); ++next) { |
644 | Node *n = wq.at(next); |
645 | for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { |
646 | Node* u = n->fast_out(i); |
647 | if (!u->is_CFG()) { |
648 | Node* c = _phase->get_ctrl(u); |
649 | if (_lpt->is_member(_phase->get_loop(c)) || _phase->is_dominator(c, head)) { |
650 | _visited.set(u->_idx); |
651 | wq.push(u); |
652 | } |
653 | } |
654 | } |
655 | } |
656 | } |
657 | } |
658 | |
659 | // Map old to n for invariance computation and clone |
660 | void map_ctrl(Node* old, Node* n) { |
661 | assert(old->is_CFG() && n->is_CFG(), "must be" ); |
662 | _old_new.map(old->_idx, n); // "clone" of old is n |
663 | _invariant.set(old->_idx); // old is invariant |
664 | _clone_visited.set(old->_idx); |
665 | } |
666 | |
667 | // Driver function to compute invariance |
668 | bool is_invariant(Node* n) { |
669 | if (!_visited.test_set(n->_idx)) |
670 | compute_invariance(n); |
671 | return (_invariant.test(n->_idx) != 0); |
672 | } |
673 | |
674 | // Driver function to clone invariant |
675 | Node* clone(Node* n, Node* ctrl) { |
676 | assert(ctrl->is_CFG(), "must be" ); |
677 | assert(_invariant.test(n->_idx), "must be an invariant" ); |
678 | if (!_clone_visited.test(n->_idx)) |
679 | clone_nodes(n, ctrl); |
680 | return _old_new[n->_idx]; |
681 | } |
682 | }; |
683 | |
684 | //------------------------------is_range_check_if ----------------------------------- |
685 | // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format |
686 | // Note: this function is particularly designed for loop predication. We require load_range |
687 | // and offset to be loop invariant computed on the fly by "invar" |
688 | bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const { |
689 | if (!is_loop_exit(iff)) { |
690 | return false; |
691 | } |
692 | if (!iff->in(1)->is_Bool()) { |
693 | return false; |
694 | } |
695 | const BoolNode *bol = iff->in(1)->as_Bool(); |
696 | if (bol->_test._test != BoolTest::lt) { |
697 | return false; |
698 | } |
699 | if (!bol->in(1)->is_Cmp()) { |
700 | return false; |
701 | } |
702 | const CmpNode *cmp = bol->in(1)->as_Cmp(); |
703 | if (cmp->Opcode() != Op_CmpU) { |
704 | return false; |
705 | } |
706 | Node* range = cmp->in(2); |
707 | if (range->Opcode() != Op_LoadRange && !iff->is_RangeCheck()) { |
708 | const TypeInt* tint = phase->_igvn.type(range)->isa_int(); |
709 | if (tint == NULL || tint->empty() || tint->_lo < 0) { |
710 | // Allow predication on positive values that aren't LoadRanges. |
711 | // This allows optimization of loops where the length of the |
712 | // array is a known value and doesn't need to be loaded back |
713 | // from the array. |
714 | return false; |
715 | } |
716 | } |
717 | if (!invar.is_invariant(range)) { |
718 | return false; |
719 | } |
720 | Node *iv = _head->as_CountedLoop()->phi(); |
721 | int scale = 0; |
722 | Node *offset = NULL; |
723 | if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) { |
724 | return false; |
725 | } |
726 | if (offset && !invar.is_invariant(offset)) { // offset must be invariant |
727 | return false; |
728 | } |
729 | return true; |
730 | } |
731 | |
732 | //------------------------------rc_predicate----------------------------------- |
733 | // Create a range check predicate |
734 | // |
735 | // for (i = init; i < limit; i += stride) { |
736 | // a[scale*i+offset] |
737 | // } |
738 | // |
739 | // Compute max(scale*i + offset) for init <= i < limit and build the predicate |
740 | // as "max(scale*i + offset) u< a.length". |
741 | // |
742 | // There are two cases for max(scale*i + offset): |
743 | // (1) stride*scale > 0 |
744 | // max(scale*i + offset) = scale*(limit-stride) + offset |
745 | // (2) stride*scale < 0 |
746 | // max(scale*i + offset) = scale*init + offset |
747 | BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl, |
748 | int scale, Node* offset, |
749 | Node* init, Node* limit, jint stride, |
750 | Node* range, bool upper, bool &overflow) { |
751 | jint con_limit = (limit != NULL && limit->is_Con()) ? limit->get_int() : 0; |
752 | jint con_init = init->is_Con() ? init->get_int() : 0; |
753 | jint con_offset = offset->is_Con() ? offset->get_int() : 0; |
754 | |
755 | stringStream* predString = NULL; |
756 | if (TraceLoopPredicate) { |
757 | predString = new stringStream(); |
758 | predString->print("rc_predicate " ); |
759 | } |
760 | |
761 | overflow = false; |
762 | Node* max_idx_expr = NULL; |
763 | const TypeInt* idx_type = TypeInt::INT; |
764 | if ((stride > 0) == (scale > 0) == upper) { |
765 | guarantee(limit != NULL, "sanity" ); |
766 | if (TraceLoopPredicate) { |
767 | if (limit->is_Con()) { |
768 | predString->print("(%d " , con_limit); |
769 | } else { |
770 | predString->print("(limit " ); |
771 | } |
772 | predString->print("- %d) " , stride); |
773 | } |
774 | // Check if (limit - stride) may overflow |
775 | const TypeInt* limit_type = _igvn.type(limit)->isa_int(); |
776 | jint limit_lo = limit_type->_lo; |
777 | jint limit_hi = limit_type->_hi; |
778 | if ((stride > 0 && (java_subtract(limit_lo, stride) < limit_lo)) || |
779 | (stride < 0 && (java_subtract(limit_hi, stride) > limit_hi))) { |
780 | // No overflow possible |
781 | ConINode* con_stride = _igvn.intcon(stride); |
782 | set_ctrl(con_stride, C->root()); |
783 | max_idx_expr = new SubINode(limit, con_stride); |
784 | idx_type = TypeInt::make(limit_lo - stride, limit_hi - stride, limit_type->_widen); |
785 | } else { |
786 | // May overflow |
787 | overflow = true; |
788 | limit = new ConvI2LNode(limit); |
789 | register_new_node(limit, ctrl); |
790 | ConLNode* con_stride = _igvn.longcon(stride); |
791 | set_ctrl(con_stride, C->root()); |
792 | max_idx_expr = new SubLNode(limit, con_stride); |
793 | } |
794 | register_new_node(max_idx_expr, ctrl); |
795 | } else { |
796 | if (TraceLoopPredicate) { |
797 | if (init->is_Con()) { |
798 | predString->print("%d " , con_init); |
799 | } else { |
800 | predString->print("init " ); |
801 | } |
802 | } |
803 | idx_type = _igvn.type(init)->isa_int(); |
804 | max_idx_expr = init; |
805 | } |
806 | |
807 | if (scale != 1) { |
808 | ConNode* con_scale = _igvn.intcon(scale); |
809 | set_ctrl(con_scale, C->root()); |
810 | if (TraceLoopPredicate) { |
811 | predString->print("* %d " , scale); |
812 | } |
813 | // Check if (scale * max_idx_expr) may overflow |
814 | const TypeInt* scale_type = TypeInt::make(scale); |
815 | MulINode* mul = new MulINode(max_idx_expr, con_scale); |
816 | idx_type = (TypeInt*)mul->mul_ring(idx_type, scale_type); |
817 | if (overflow || TypeInt::INT->higher_equal(idx_type)) { |
818 | // May overflow |
819 | mul->destruct(); |
820 | if (!overflow) { |
821 | max_idx_expr = new ConvI2LNode(max_idx_expr); |
822 | register_new_node(max_idx_expr, ctrl); |
823 | } |
824 | overflow = true; |
825 | con_scale = _igvn.longcon(scale); |
826 | set_ctrl(con_scale, C->root()); |
827 | max_idx_expr = new MulLNode(max_idx_expr, con_scale); |
828 | } else { |
829 | // No overflow possible |
830 | max_idx_expr = mul; |
831 | } |
832 | register_new_node(max_idx_expr, ctrl); |
833 | } |
834 | |
835 | if (offset && (!offset->is_Con() || con_offset != 0)){ |
836 | if (TraceLoopPredicate) { |
837 | if (offset->is_Con()) { |
838 | predString->print("+ %d " , con_offset); |
839 | } else { |
840 | predString->print("+ offset" ); |
841 | } |
842 | } |
843 | // Check if (max_idx_expr + offset) may overflow |
844 | const TypeInt* offset_type = _igvn.type(offset)->isa_int(); |
845 | jint lo = java_add(idx_type->_lo, offset_type->_lo); |
846 | jint hi = java_add(idx_type->_hi, offset_type->_hi); |
847 | if (overflow || (lo > hi) || |
848 | ((idx_type->_lo & offset_type->_lo) < 0 && lo >= 0) || |
849 | ((~(idx_type->_hi | offset_type->_hi)) < 0 && hi < 0)) { |
850 | // May overflow |
851 | if (!overflow) { |
852 | max_idx_expr = new ConvI2LNode(max_idx_expr); |
853 | register_new_node(max_idx_expr, ctrl); |
854 | } |
855 | overflow = true; |
856 | offset = new ConvI2LNode(offset); |
857 | register_new_node(offset, ctrl); |
858 | max_idx_expr = new AddLNode(max_idx_expr, offset); |
859 | } else { |
860 | // No overflow possible |
861 | max_idx_expr = new AddINode(max_idx_expr, offset); |
862 | } |
863 | register_new_node(max_idx_expr, ctrl); |
864 | } |
865 | |
866 | CmpNode* cmp = NULL; |
867 | if (overflow) { |
868 | // Integer expressions may overflow, do long comparison |
869 | range = new ConvI2LNode(range); |
870 | register_new_node(range, ctrl); |
871 | cmp = new CmpULNode(max_idx_expr, range); |
872 | } else { |
873 | cmp = new CmpUNode(max_idx_expr, range); |
874 | } |
875 | register_new_node(cmp, ctrl); |
876 | BoolNode* bol = new BoolNode(cmp, BoolTest::lt); |
877 | register_new_node(bol, ctrl); |
878 | |
879 | if (TraceLoopPredicate) { |
880 | predString->print_cr("<u range" ); |
881 | tty->print("%s" , predString->as_string()); |
882 | } |
883 | return bol; |
884 | } |
885 | |
886 | // Should loop predication look not only in the path from tail to head |
887 | // but also in branches of the loop body? |
888 | bool PhaseIdealLoop::loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt) { |
889 | if (!UseProfiledLoopPredicate) { |
890 | return false; |
891 | } |
892 | |
893 | if (predicate_proj == NULL) { |
894 | return false; |
895 | } |
896 | |
897 | LoopNode* head = loop->_head->as_Loop(); |
898 | bool follow_branches = true; |
899 | IdealLoopTree* l = loop->_child; |
900 | // For leaf loops and loops with a single inner loop |
901 | while (l != NULL && follow_branches) { |
902 | IdealLoopTree* child = l; |
903 | if (child->_child != NULL && |
904 | child->_head->is_OuterStripMinedLoop()) { |
905 | assert(child->_child->_next == NULL, "only one inner loop for strip mined loop" ); |
906 | assert(child->_child->_head->is_CountedLoop() && child->_child->_head->as_CountedLoop()->is_strip_mined(), "inner loop should be strip mined" ); |
907 | child = child->_child; |
908 | } |
909 | if (child->_child != NULL || child->_irreducible) { |
910 | follow_branches = false; |
911 | } |
912 | l = l->_next; |
913 | } |
914 | if (follow_branches) { |
915 | loop->compute_profile_trip_cnt(this); |
916 | if (head->is_profile_trip_failed()) { |
917 | follow_branches = false; |
918 | } else { |
919 | loop_trip_cnt = head->profile_trip_cnt(); |
920 | if (head->is_CountedLoop()) { |
921 | CountedLoopNode* cl = head->as_CountedLoop(); |
922 | if (cl->phi() != NULL) { |
923 | const TypeInt* t = _igvn.type(cl->phi())->is_int(); |
924 | float worst_case_trip_cnt = ((float)t->_hi - t->_lo) / ABS(cl->stride_con()); |
925 | if (worst_case_trip_cnt < loop_trip_cnt) { |
926 | loop_trip_cnt = worst_case_trip_cnt; |
927 | } |
928 | } |
929 | } |
930 | } |
931 | } |
932 | return follow_branches; |
933 | } |
934 | |
935 | // Compute probability of reaching some CFG node from a fixed |
936 | // dominating CFG node |
937 | class PathFrequency { |
938 | private: |
939 | Node* _dom; // frequencies are computed relative to this node |
940 | Node_Stack _stack; |
941 | GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions |
942 | GrowableArray<float> _freqs; // cache frequencies |
943 | PhaseIdealLoop* _phase; |
944 | |
945 | void set_rounding(int mode) { |
946 | // fesetround is broken on windows |
947 | NOT_WINDOWS(fesetround(mode);) |
948 | } |
949 | |
950 | void check_frequency(float f) { |
951 | NOT_WINDOWS(assert(f <= 1 && f >= 0, "Incorrect frequency" );) |
952 | } |
953 | |
954 | public: |
955 | PathFrequency(Node* dom, PhaseIdealLoop* phase) |
956 | : _dom(dom), _stack(0), _phase(phase) { |
957 | } |
958 | |
959 | float to(Node* n) { |
960 | // post order walk on the CFG graph from n to _dom |
961 | set_rounding(FE_TOWARDZERO); // make sure rounding doesn't push frequency above 1 |
962 | IdealLoopTree* loop = _phase->get_loop(_dom); |
963 | Node* c = n; |
964 | for (;;) { |
965 | assert(_phase->get_loop(c) == loop, "have to be in the same loop" ); |
966 | if (c == _dom || _freqs.at_grow(c->_idx, -1) >= 0) { |
967 | float f = c == _dom ? 1 : _freqs.at(c->_idx); |
968 | Node* prev = c; |
969 | while (_stack.size() > 0 && prev == c) { |
970 | Node* n = _stack.node(); |
971 | if (!n->is_Region()) { |
972 | if (_phase->get_loop(n) != _phase->get_loop(n->in(0))) { |
973 | // Found an inner loop: compute frequency of reaching this |
974 | // exit from the loop head by looking at the number of |
975 | // times each loop exit was taken |
976 | IdealLoopTree* inner_loop = _phase->get_loop(n->in(0)); |
977 | LoopNode* inner_head = inner_loop->_head->as_Loop(); |
978 | assert(_phase->get_loop(n) == loop, "only 1 inner loop" ); |
979 | if (inner_head->is_OuterStripMinedLoop()) { |
980 | inner_head->verify_strip_mined(1); |
981 | if (n->in(0) == inner_head->in(LoopNode::LoopBackControl)->in(0)) { |
982 | n = n->in(0)->in(0)->in(0); |
983 | } |
984 | inner_loop = inner_loop->_child; |
985 | inner_head = inner_loop->_head->as_Loop(); |
986 | inner_head->verify_strip_mined(1); |
987 | } |
988 | set_rounding(FE_UPWARD); // make sure rounding doesn't push frequency above 1 |
989 | float loop_exit_cnt = 0.0f; |
990 | for (uint i = 0; i < inner_loop->_body.size(); i++) { |
991 | Node *n = inner_loop->_body[i]; |
992 | float c = inner_loop->compute_profile_trip_cnt_helper(n); |
993 | loop_exit_cnt += c; |
994 | } |
995 | set_rounding(FE_TOWARDZERO); |
996 | float cnt = -1; |
997 | if (n->in(0)->is_If()) { |
998 | IfNode* iff = n->in(0)->as_If(); |
999 | float p = n->in(0)->as_If()->_prob; |
1000 | if (n->Opcode() == Op_IfFalse) { |
1001 | p = 1 - p; |
1002 | } |
1003 | if (p > PROB_MIN) { |
1004 | cnt = p * iff->_fcnt; |
1005 | } else { |
1006 | cnt = 0; |
1007 | } |
1008 | } else { |
1009 | assert(n->in(0)->is_Jump(), "unsupported node kind" ); |
1010 | JumpNode* jmp = n->in(0)->as_Jump(); |
1011 | float p = n->in(0)->as_Jump()->_probs[n->as_JumpProj()->_con]; |
1012 | cnt = p * jmp->_fcnt; |
1013 | } |
1014 | float this_exit_f = cnt > 0 ? cnt / loop_exit_cnt : 0; |
1015 | check_frequency(this_exit_f); |
1016 | f = f * this_exit_f; |
1017 | check_frequency(f); |
1018 | } else { |
1019 | float p = -1; |
1020 | if (n->in(0)->is_If()) { |
1021 | p = n->in(0)->as_If()->_prob; |
1022 | if (n->Opcode() == Op_IfFalse) { |
1023 | p = 1 - p; |
1024 | } |
1025 | } else { |
1026 | assert(n->in(0)->is_Jump(), "unsupported node kind" ); |
1027 | p = n->in(0)->as_Jump()->_probs[n->as_JumpProj()->_con]; |
1028 | } |
1029 | f = f * p; |
1030 | check_frequency(f); |
1031 | } |
1032 | _freqs.at_put_grow(n->_idx, (float)f, -1); |
1033 | _stack.pop(); |
1034 | } else { |
1035 | float prev_f = _freqs_stack.pop(); |
1036 | float new_f = f; |
1037 | f = new_f + prev_f; |
1038 | check_frequency(f); |
1039 | uint i = _stack.index(); |
1040 | if (i < n->req()) { |
1041 | c = n->in(i); |
1042 | _stack.set_index(i+1); |
1043 | _freqs_stack.push(f); |
1044 | } else { |
1045 | _freqs.at_put_grow(n->_idx, f, -1); |
1046 | _stack.pop(); |
1047 | } |
1048 | } |
1049 | } |
1050 | if (_stack.size() == 0) { |
1051 | set_rounding(FE_TONEAREST); |
1052 | check_frequency(f); |
1053 | return f; |
1054 | } |
1055 | } else if (c->is_Loop()) { |
1056 | ShouldNotReachHere(); |
1057 | c = c->in(LoopNode::EntryControl); |
1058 | } else if (c->is_Region()) { |
1059 | _freqs_stack.push(0); |
1060 | _stack.push(c, 2); |
1061 | c = c->in(1); |
1062 | } else { |
1063 | if (c->is_IfProj()) { |
1064 | IfNode* iff = c->in(0)->as_If(); |
1065 | if (iff->_prob == PROB_UNKNOWN) { |
1066 | // assume never taken |
1067 | _freqs.at_put_grow(c->_idx, 0, -1); |
1068 | } else if (_phase->get_loop(c) != _phase->get_loop(iff)) { |
1069 | if (iff->_fcnt == COUNT_UNKNOWN) { |
1070 | // assume never taken |
1071 | _freqs.at_put_grow(c->_idx, 0, -1); |
1072 | } else { |
1073 | // skip over loop |
1074 | _stack.push(c, 1); |
1075 | c = _phase->get_loop(c->in(0))->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); |
1076 | } |
1077 | } else { |
1078 | _stack.push(c, 1); |
1079 | c = iff; |
1080 | } |
1081 | } else if (c->is_JumpProj()) { |
1082 | JumpNode* jmp = c->in(0)->as_Jump(); |
1083 | if (_phase->get_loop(c) != _phase->get_loop(jmp)) { |
1084 | if (jmp->_fcnt == COUNT_UNKNOWN) { |
1085 | // assume never taken |
1086 | _freqs.at_put_grow(c->_idx, 0, -1); |
1087 | } else { |
1088 | // skip over loop |
1089 | _stack.push(c, 1); |
1090 | c = _phase->get_loop(c->in(0))->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); |
1091 | } |
1092 | } else { |
1093 | _stack.push(c, 1); |
1094 | c = jmp; |
1095 | } |
1096 | } else if (c->Opcode() == Op_CatchProj && |
1097 | c->in(0)->Opcode() == Op_Catch && |
1098 | c->in(0)->in(0)->is_Proj() && |
1099 | c->in(0)->in(0)->in(0)->is_Call()) { |
1100 | // assume exceptions are never thrown |
1101 | uint con = c->as_Proj()->_con; |
1102 | if (con == CatchProjNode::fall_through_index) { |
1103 | Node* call = c->in(0)->in(0)->in(0)->in(0); |
1104 | if (_phase->get_loop(call) != _phase->get_loop(c)) { |
1105 | _freqs.at_put_grow(c->_idx, 0, -1); |
1106 | } else { |
1107 | c = call; |
1108 | } |
1109 | } else { |
1110 | assert(con >= CatchProjNode::catch_all_index, "what else?" ); |
1111 | _freqs.at_put_grow(c->_idx, 0, -1); |
1112 | } |
1113 | } else if (c->unique_ctrl_out() == NULL && !c->is_If() && !c->is_Jump()) { |
1114 | ShouldNotReachHere(); |
1115 | } else { |
1116 | c = c->in(0); |
1117 | } |
1118 | } |
1119 | } |
1120 | ShouldNotReachHere(); |
1121 | return -1; |
1122 | } |
1123 | }; |
1124 | |
1125 | void PhaseIdealLoop::loop_predication_follow_branches(Node *n, IdealLoopTree *loop, float loop_trip_cnt, |
1126 | PathFrequency& pf, Node_Stack& stack, VectorSet& seen, |
1127 | Node_List& if_proj_list) { |
1128 | assert(n->is_Region(), "start from a region" ); |
1129 | Node* tail = loop->tail(); |
1130 | stack.push(n, 1); |
1131 | do { |
1132 | Node* c = stack.node(); |
1133 | assert(c->is_Region() || c->is_IfProj(), "only region here" ); |
1134 | uint i = stack.index(); |
1135 | |
1136 | if (i < c->req()) { |
1137 | stack.set_index(i+1); |
1138 | Node* in = c->in(i); |
1139 | while (!is_dominator(in, tail) && !seen.test_set(in->_idx)) { |
1140 | IdealLoopTree* in_loop = get_loop(in); |
1141 | if (in_loop != loop) { |
1142 | in = in_loop->_head->in(LoopNode::EntryControl); |
1143 | } else if (in->is_Region()) { |
1144 | stack.push(in, 1); |
1145 | break; |
1146 | } else if (in->is_IfProj() && |
1147 | in->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && |
1148 | (in->in(0)->Opcode() == Op_If || |
1149 | in->in(0)->Opcode() == Op_RangeCheck)) { |
1150 | if (pf.to(in) * loop_trip_cnt >= 1) { |
1151 | stack.push(in, 1); |
1152 | } |
1153 | in = in->in(0); |
1154 | } else { |
1155 | in = in->in(0); |
1156 | } |
1157 | } |
1158 | } else { |
1159 | if (c->is_IfProj()) { |
1160 | if_proj_list.push(c); |
1161 | } |
1162 | stack.pop(); |
1163 | } |
1164 | |
1165 | } while (stack.size() > 0); |
1166 | } |
1167 | |
1168 | |
1169 | bool PhaseIdealLoop::loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj, |
1170 | CountedLoopNode *cl, ConNode* zero, Invariance& invar, |
1171 | Deoptimization::DeoptReason reason) { |
1172 | // Following are changed to nonnull when a predicate can be hoisted |
1173 | ProjNode* new_predicate_proj = NULL; |
1174 | IfNode* iff = proj->in(0)->as_If(); |
1175 | Node* test = iff->in(1); |
1176 | if (!test->is_Bool()){ //Conv2B, ... |
1177 | return false; |
1178 | } |
1179 | BoolNode* bol = test->as_Bool(); |
1180 | if (invar.is_invariant(bol)) { |
1181 | // Invariant test |
1182 | new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL, |
1183 | reason, |
1184 | iff->Opcode()); |
1185 | Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0); |
1186 | BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool(); |
1187 | |
1188 | // Negate test if necessary |
1189 | bool negated = false; |
1190 | if (proj->_con != predicate_proj->_con) { |
1191 | new_predicate_bol = new BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate()); |
1192 | register_new_node(new_predicate_bol, ctrl); |
1193 | negated = true; |
1194 | } |
1195 | IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If(); |
1196 | _igvn.hash_delete(new_predicate_iff); |
1197 | new_predicate_iff->set_req(1, new_predicate_bol); |
1198 | #ifndef PRODUCT |
1199 | if (TraceLoopPredicate) { |
1200 | tty->print("Predicate invariant if%s: %d " , negated ? " negated" : "" , new_predicate_iff->_idx); |
1201 | loop->dump_head(); |
1202 | } else if (TraceLoopOpts) { |
1203 | tty->print("Predicate IC " ); |
1204 | loop->dump_head(); |
1205 | } |
1206 | #endif |
1207 | } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) { |
1208 | // Range check for counted loops |
1209 | const Node* cmp = bol->in(1)->as_Cmp(); |
1210 | Node* idx = cmp->in(1); |
1211 | assert(!invar.is_invariant(idx), "index is variant" ); |
1212 | Node* rng = cmp->in(2); |
1213 | assert(rng->Opcode() == Op_LoadRange || iff->is_RangeCheck() || _igvn.type(rng)->is_int()->_lo >= 0, "must be" ); |
1214 | assert(invar.is_invariant(rng), "range must be invariant" ); |
1215 | int scale = 1; |
1216 | Node* offset = zero; |
1217 | bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset); |
1218 | assert(ok, "must be index expression" ); |
1219 | |
1220 | Node* init = cl->init_trip(); |
1221 | // Limit is not exact. |
1222 | // Calculate exact limit here. |
1223 | // Note, counted loop's test is '<' or '>'. |
1224 | Node* limit = exact_limit(loop); |
1225 | int stride = cl->stride()->get_int(); |
1226 | |
1227 | // Build if's for the upper and lower bound tests. The |
1228 | // lower_bound test will dominate the upper bound test and all |
1229 | // cloned or created nodes will use the lower bound test as |
1230 | // their declared control. |
1231 | |
1232 | // Perform cloning to keep Invariance state correct since the |
1233 | // late schedule will place invariant things in the loop. |
1234 | Node *ctrl = predicate_proj->in(0)->as_If()->in(0); |
1235 | rng = invar.clone(rng, ctrl); |
1236 | if (offset && offset != zero) { |
1237 | assert(invar.is_invariant(offset), "offset must be loop invariant" ); |
1238 | offset = invar.clone(offset, ctrl); |
1239 | } |
1240 | // If predicate expressions may overflow in the integer range, longs are used. |
1241 | bool overflow = false; |
1242 | |
1243 | // Test the lower bound |
1244 | BoolNode* lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false, overflow); |
1245 | // Negate test if necessary |
1246 | bool negated = false; |
1247 | if (proj->_con != predicate_proj->_con) { |
1248 | lower_bound_bol = new BoolNode(lower_bound_bol->in(1), lower_bound_bol->_test.negate()); |
1249 | register_new_node(lower_bound_bol, ctrl); |
1250 | negated = true; |
1251 | } |
1252 | ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, reason, overflow ? Op_If : iff->Opcode()); |
1253 | IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If(); |
1254 | _igvn.hash_delete(lower_bound_iff); |
1255 | lower_bound_iff->set_req(1, lower_bound_bol); |
1256 | if (TraceLoopPredicate) tty->print_cr("lower bound check if: %s %d " , negated ? " negated" : "" , lower_bound_iff->_idx); |
1257 | |
1258 | // Test the upper bound |
1259 | BoolNode* upper_bound_bol = rc_predicate(loop, lower_bound_proj, scale, offset, init, limit, stride, rng, true, overflow); |
1260 | negated = false; |
1261 | if (proj->_con != predicate_proj->_con) { |
1262 | upper_bound_bol = new BoolNode(upper_bound_bol->in(1), upper_bound_bol->_test.negate()); |
1263 | register_new_node(upper_bound_bol, ctrl); |
1264 | negated = true; |
1265 | } |
1266 | ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, reason, overflow ? Op_If : iff->Opcode()); |
1267 | assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate" ); |
1268 | IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If(); |
1269 | _igvn.hash_delete(upper_bound_iff); |
1270 | upper_bound_iff->set_req(1, upper_bound_bol); |
1271 | if (TraceLoopPredicate) tty->print_cr("upper bound check if: %s %d " , negated ? " negated" : "" , lower_bound_iff->_idx); |
1272 | |
1273 | // Fall through into rest of the clean up code which will move |
1274 | // any dependent nodes onto the upper bound test. |
1275 | new_predicate_proj = upper_bound_proj; |
1276 | |
1277 | if (iff->is_RangeCheck()) { |
1278 | new_predicate_proj = insert_skeleton_predicate(iff, loop, proj, predicate_proj, upper_bound_proj, scale, offset, init, limit, stride, rng, overflow, reason); |
1279 | } |
1280 | |
1281 | #ifndef PRODUCT |
1282 | if (TraceLoopOpts && !TraceLoopPredicate) { |
1283 | tty->print("Predicate RC " ); |
1284 | loop->dump_head(); |
1285 | } |
1286 | #endif |
1287 | } else { |
1288 | // Loop variant check (for example, range check in non-counted loop) |
1289 | // with uncommon trap. |
1290 | return false; |
1291 | } |
1292 | assert(new_predicate_proj != NULL, "sanity" ); |
1293 | // Success - attach condition (new_predicate_bol) to predicate if |
1294 | invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate |
1295 | |
1296 | // Eliminate the old If in the loop body |
1297 | dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con ); |
1298 | |
1299 | C->set_major_progress(); |
1300 | return true; |
1301 | } |
1302 | |
1303 | |
1304 | // After pre/main/post loops are created, we'll put a copy of some |
1305 | // range checks between the pre and main loop to validate the value |
1306 | // of the main loop induction variable. Make a copy of the predicates |
1307 | // here with an opaque node as a place holder for the value (will be |
1308 | // updated by PhaseIdealLoop::clone_skeleton_predicate()). |
1309 | ProjNode* PhaseIdealLoop::insert_skeleton_predicate(IfNode* iff, IdealLoopTree *loop, |
1310 | ProjNode* proj, ProjNode *predicate_proj, |
1311 | ProjNode* upper_bound_proj, |
1312 | int scale, Node* offset, |
1313 | Node* init, Node* limit, jint stride, |
1314 | Node* rng, bool &overflow, |
1315 | Deoptimization::DeoptReason reason) { |
1316 | assert(proj->_con && predicate_proj->_con, "not a range check?" ); |
1317 | Node* opaque_init = new Opaque1Node(C, init); |
1318 | register_new_node(opaque_init, upper_bound_proj); |
1319 | BoolNode* bol = rc_predicate(loop, upper_bound_proj, scale, offset, opaque_init, limit, stride, rng, (stride > 0) != (scale > 0), overflow); |
1320 | Node* opaque_bol = new Opaque4Node(C, bol, _igvn.intcon(1)); // This will go away once loop opts are over |
1321 | register_new_node(opaque_bol, upper_bound_proj); |
1322 | ProjNode* new_proj = create_new_if_for_predicate(predicate_proj, NULL, reason, overflow ? Op_If : iff->Opcode()); |
1323 | _igvn.replace_input_of(new_proj->in(0), 1, opaque_bol); |
1324 | assert(opaque_init->outcnt() > 0, "should be used" ); |
1325 | return new_proj; |
1326 | } |
1327 | |
1328 | //------------------------------ loop_predication_impl-------------------------- |
1329 | // Insert loop predicates for null checks and range checks |
1330 | bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) { |
1331 | if (!UseLoopPredicate) return false; |
1332 | |
1333 | if (!loop->_head->is_Loop()) { |
1334 | // Could be a simple region when irreducible loops are present. |
1335 | return false; |
1336 | } |
1337 | LoopNode* head = loop->_head->as_Loop(); |
1338 | |
1339 | if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) { |
1340 | // do nothing for infinite loops |
1341 | return false; |
1342 | } |
1343 | |
1344 | if (head->is_OuterStripMinedLoop()) { |
1345 | return false; |
1346 | } |
1347 | |
1348 | CountedLoopNode *cl = NULL; |
1349 | if (head->is_valid_counted_loop()) { |
1350 | cl = head->as_CountedLoop(); |
1351 | // do nothing for iteration-splitted loops |
1352 | if (!cl->is_normal_loop()) return false; |
1353 | // Avoid RCE if Counted loop's test is '!='. |
1354 | BoolTest::mask bt = cl->loopexit()->test_trip(); |
1355 | if (bt != BoolTest::lt && bt != BoolTest::gt) |
1356 | cl = NULL; |
1357 | } |
1358 | |
1359 | Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl); |
1360 | ProjNode *loop_limit_proj = NULL; |
1361 | ProjNode *predicate_proj = NULL; |
1362 | ProjNode *profile_predicate_proj = NULL; |
1363 | // Loop limit check predicate should be near the loop. |
1364 | loop_limit_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); |
1365 | if (loop_limit_proj != NULL) { |
1366 | entry = skip_loop_predicates(loop_limit_proj); |
1367 | } |
1368 | bool has_profile_predicates = false; |
1369 | profile_predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate); |
1370 | if (profile_predicate_proj != NULL) { |
1371 | Node* n = skip_loop_predicates(entry); |
1372 | // Check if predicates were already added to the profile predicate |
1373 | // block |
1374 | if (n != entry->in(0)->in(0) || n->outcnt() != 1) { |
1375 | has_profile_predicates = true; |
1376 | } |
1377 | entry = n; |
1378 | } |
1379 | predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); |
1380 | |
1381 | float loop_trip_cnt = -1; |
1382 | bool follow_branches = loop_predication_should_follow_branches(loop, profile_predicate_proj, loop_trip_cnt); |
1383 | assert(!follow_branches || loop_trip_cnt >= 0, "negative trip count?" ); |
1384 | |
1385 | if (predicate_proj == NULL && !follow_branches) { |
1386 | #ifndef PRODUCT |
1387 | if (TraceLoopPredicate) { |
1388 | tty->print("missing predicate:" ); |
1389 | loop->dump_head(); |
1390 | head->dump(1); |
1391 | } |
1392 | #endif |
1393 | return false; |
1394 | } |
1395 | ConNode* zero = _igvn.intcon(0); |
1396 | set_ctrl(zero, C->root()); |
1397 | |
1398 | ResourceArea *area = Thread::current()->resource_area(); |
1399 | Invariance invar(area, loop); |
1400 | |
1401 | // Create list of if-projs such that a newer proj dominates all older |
1402 | // projs in the list, and they all dominate loop->tail() |
1403 | Node_List if_proj_list(area); |
1404 | Node_List regions(area); |
1405 | Node *current_proj = loop->tail(); //start from tail |
1406 | |
1407 | |
1408 | Node_List controls(area); |
1409 | while (current_proj != head) { |
1410 | if (loop == get_loop(current_proj) && // still in the loop ? |
1411 | current_proj->is_Proj() && // is a projection ? |
1412 | (current_proj->in(0)->Opcode() == Op_If || |
1413 | current_proj->in(0)->Opcode() == Op_RangeCheck)) { // is a if projection ? |
1414 | if_proj_list.push(current_proj); |
1415 | } |
1416 | if (follow_branches && |
1417 | current_proj->Opcode() == Op_Region && |
1418 | loop == get_loop(current_proj)) { |
1419 | regions.push(current_proj); |
1420 | } |
1421 | current_proj = idom(current_proj); |
1422 | } |
1423 | |
1424 | bool hoisted = false; // true if at least one proj is promoted |
1425 | |
1426 | if (!has_profile_predicates) { |
1427 | while (if_proj_list.size() > 0) { |
1428 | Node* n = if_proj_list.pop(); |
1429 | |
1430 | ProjNode* proj = n->as_Proj(); |
1431 | IfNode* iff = proj->in(0)->as_If(); |
1432 | |
1433 | CallStaticJavaNode* call = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); |
1434 | if (call == NULL) { |
1435 | if (loop->is_loop_exit(iff)) { |
1436 | // stop processing the remaining projs in the list because the execution of them |
1437 | // depends on the condition of "iff" (iff->in(1)). |
1438 | break; |
1439 | } else { |
1440 | // Both arms are inside the loop. There are two cases: |
1441 | // (1) there is one backward branch. In this case, any remaining proj |
1442 | // in the if_proj list post-dominates "iff". So, the condition of "iff" |
1443 | // does not determine the execution the remining projs directly, and we |
1444 | // can safely continue. |
1445 | // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj" |
1446 | // does not dominate loop->tail(), so it can not be in the if_proj list. |
1447 | continue; |
1448 | } |
1449 | } |
1450 | Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(call->uncommon_trap_request()); |
1451 | if (reason == Deoptimization::Reason_predicate) { |
1452 | break; |
1453 | } |
1454 | |
1455 | if (predicate_proj != NULL) { |
1456 | hoisted = loop_predication_impl_helper(loop, proj, predicate_proj, cl, zero, invar, Deoptimization::Reason_predicate) | hoisted; |
1457 | } |
1458 | } // end while |
1459 | } |
1460 | |
1461 | Node_List if_proj_list_freq(area); |
1462 | if (follow_branches) { |
1463 | PathFrequency pf(loop->_head, this); |
1464 | |
1465 | // Some projections were skipped by regular predicates because of |
1466 | // an early loop exit. Try them with profile data. |
1467 | while (if_proj_list.size() > 0) { |
1468 | Node* proj = if_proj_list.pop(); |
1469 | float f = pf.to(proj); |
1470 | if (proj->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && |
1471 | f * loop_trip_cnt >= 1) { |
1472 | hoisted = loop_predication_impl_helper(loop, proj->as_Proj(), profile_predicate_proj, cl, zero, invar, Deoptimization::Reason_profile_predicate) | hoisted; |
1473 | } |
1474 | } |
1475 | |
1476 | // And look into all branches |
1477 | Node_Stack stack(0); |
1478 | VectorSet seen(Thread::current()->resource_area()); |
1479 | while (regions.size() > 0) { |
1480 | Node* c = regions.pop(); |
1481 | loop_predication_follow_branches(c, loop, loop_trip_cnt, pf, stack, seen, if_proj_list_freq); |
1482 | } |
1483 | |
1484 | for (uint i = 0; i < if_proj_list_freq.size(); i++) { |
1485 | ProjNode* proj = if_proj_list_freq.at(i)->as_Proj(); |
1486 | hoisted = loop_predication_impl_helper(loop, proj, profile_predicate_proj, cl, zero, invar, Deoptimization::Reason_profile_predicate) | hoisted; |
1487 | } |
1488 | } |
1489 | |
1490 | #ifndef PRODUCT |
1491 | // report that the loop predication has been actually performed |
1492 | // for this loop |
1493 | if (TraceLoopPredicate && hoisted) { |
1494 | tty->print("Loop Predication Performed:" ); |
1495 | loop->dump_head(); |
1496 | } |
1497 | #endif |
1498 | |
1499 | head->verify_strip_mined(1); |
1500 | |
1501 | return hoisted; |
1502 | } |
1503 | |
1504 | //------------------------------loop_predication-------------------------------- |
1505 | // driver routine for loop predication optimization |
1506 | bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) { |
1507 | bool hoisted = false; |
1508 | // Recursively promote predicates |
1509 | if (_child) { |
1510 | hoisted = _child->loop_predication( phase); |
1511 | } |
1512 | |
1513 | // self |
1514 | if (!_irreducible && !tail()->is_top()) { |
1515 | hoisted |= phase->loop_predication_impl(this); |
1516 | } |
1517 | |
1518 | if (_next) { //sibling |
1519 | hoisted |= _next->loop_predication( phase); |
1520 | } |
1521 | |
1522 | return hoisted; |
1523 | } |
1524 | |