1 /*
2 * Copyright (c) 2006, 2019, 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 "memory/allocation.inline.hpp"
27 #include "opto/mulnode.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/connode.hpp"
30 #include "opto/convertnode.hpp"
31 #include "opto/loopnode.hpp"
32 #include "opto/opaquenode.hpp"
33 #include "opto/rootnode.hpp"
34
35 //================= Loop Unswitching =====================
36 //
37 // orig: transformed:
38 // if (invariant-test) then
39 // predicate predicate
40 // loop loop
41 // stmt1 stmt1
42 // if (invariant-test) then stmt2
43 // stmt2 stmt4
44 // else endloop
45 // stmt3 else
46 // endif predicate [clone]
47 // stmt4 loop [clone]
48 // endloop stmt1 [clone]
49 // stmt3
50 // stmt4 [clone]
51 // endloop
52 // endif
53 //
54 // Note: the "else" clause may be empty
55
56
57 //------------------------------policy_unswitching-----------------------------
58 // Return TRUE or FALSE if the loop should be unswitched
59 // (ie. clone loop with an invariant test that does not exit the loop)
60 bool IdealLoopTree::policy_unswitching( PhaseIdealLoop *phase ) const {
61 if (!LoopUnswitching) {
62 return false;
63 }
64 if (!_head->is_Loop()) {
65 return false;
66 }
67
68 // If nodes are depleted, some transform has miscalculated its needs.
69 assert(!phase->exceeding_node_budget(), "sanity");
70
71 // check for vectorized loops, any unswitching was already applied
72 if (_head->is_CountedLoop() && _head->as_CountedLoop()->is_unroll_only()) {
73 return false;
74 }
75
76 LoopNode* head = _head->as_Loop();
77 if (head->unswitch_count() + 1 > head->unswitch_max()) {
78 return false;
79 }
80
81 if (head->is_flattened_arrays()) {
82 return false;
83 }
84
85 Node_List unswitch_iffs;
86 if (phase->find_unswitching_candidate(this, unswitch_iffs) == NULL) {
87 return false;
88 }
89
90 // Too speculative if running low on nodes.
91 return phase->may_require_nodes(est_loop_clone_sz(2));
92 }
93
94 //------------------------------find_unswitching_candidate-----------------------------
95 // Find candidate "if" for unswitching
96 IfNode* PhaseIdealLoop::find_unswitching_candidate(const IdealLoopTree *loop, Node_List& unswitch_iffs) const {
97
98 // Find first invariant test that doesn't exit the loop
99 LoopNode *head = loop->_head->as_Loop();
100 IfNode* unswitch_iff = NULL;
101 Node* n = head->in(LoopNode::LoopBackControl);
102 while (n != head) {
103 Node* n_dom = idom(n);
104 if (n->is_Region()) {
105 if (n_dom->is_If()) {
106 IfNode* iff = n_dom->as_If();
107 if (iff->in(1)->is_Bool()) {
108 BoolNode* bol = iff->in(1)->as_Bool();
109 if (bol->in(1)->is_Cmp()) {
110 // If condition is invariant and not a loop exit,
111 // then found reason to unswitch.
112 if (loop->is_invariant(bol) && !loop->is_loop_exit(iff)) {
113 unswitch_iff = iff;
114 }
115 }
116 }
117 }
118 }
119 n = n_dom;
120 }
121 if (unswitch_iff != NULL) {
122 unswitch_iffs.push(unswitch_iff);
123 }
124
125 // Collect all non-flattened array checks for unswitching to create a fast loop
126 // without checks (only non-flattened array accesses) and a slow loop with checks.
127 if (unswitch_iff == NULL || unswitch_iff->is_flat_array_check(&_igvn)) {
128 for (uint i = 0; i < loop->_body.size(); i++) {
129 IfNode* n = loop->_body.at(i)->isa_If();
130 if (n != NULL && n != unswitch_iff && n->is_flat_array_check(&_igvn) &&
131 loop->is_invariant(n->in(1)) && !loop->is_loop_exit(n)) {
132 unswitch_iffs.push(n);
133 if (unswitch_iff == NULL) {
134 unswitch_iff = n;
135 }
136 }
137 }
138 }
139 return unswitch_iff;
140 }
141
142 //------------------------------do_unswitching-----------------------------
143 // Clone loop with an invariant test (that does not exit) and
144 // insert a clone of the test that selects which version to
145 // execute.
146 void PhaseIdealLoop::do_unswitching(IdealLoopTree *loop, Node_List &old_new) {
147
148 LoopNode *head = loop->_head->as_Loop();
149 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
150 if (find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check) != NULL
151 || (UseProfiledLoopPredicate && find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate) != NULL)
152 || (UseLoopPredicate && find_predicate_insertion_point(entry, Deoptimization::Reason_predicate) != NULL)) {
153 assert(entry->is_IfProj(), "sanity - must be ifProj since there is at least one predicate");
154 if (entry->outcnt() > 1) {
155 // Bailout if there are loop predicates from which there are additional control dependencies (i.e. from
156 // loop entry 'entry') to previously partially peeled statements since this case is not handled and can lead
157 // to wrong execution. Remove this bailout, once this is fixed.
158 return;
159 }
160 }
161 // Find first invariant test that doesn't exit the loop
162 Node_List unswitch_iffs;
163 IfNode* unswitch_iff = find_unswitching_candidate((const IdealLoopTree *)loop, unswitch_iffs);
164 assert(unswitch_iff != NULL && unswitch_iff == unswitch_iffs.at(0), "should be at least one");
165 bool flat_array_checks = unswitch_iffs.size() > 1;
166
167 #ifndef PRODUCT
168 if (TraceLoopOpts) {
169 tty->print("Unswitch %d ", head->unswitch_count()+1);
170 loop->dump_head();
171 for (uint i = 0; i < unswitch_iffs.size(); i++) {
172 unswitch_iffs.at(i)->dump(3);
173 tty->cr();
174 }
175 }
176 #endif
177
178 // Need to revert back to normal loop
179 if (head->is_CountedLoop() && !head->as_CountedLoop()->is_normal_loop()) {
180 head->as_CountedLoop()->set_normal_loop();
181 }
182
183 IfNode* invar_iff = create_slow_version_of_loop(loop, old_new, unswitch_iffs, CloneIncludesStripMined);
184 ProjNode* proj_true = invar_iff->proj_out(1);
185 ProjNode* proj_false = invar_iff->proj_out(0);
186
187 #ifdef ASSERT
188 assert(proj_true->is_IfTrue(), "must be true projection");
189 entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
190 Node* predicate = find_predicate(entry);
191 if (predicate == NULL) {
192 // No empty predicate
193 Node* uniqc = proj_true->unique_ctrl_out();
194 assert((uniqc == head && !head->is_strip_mined()) || (uniqc == head->in(LoopNode::EntryControl)
195 && head->is_strip_mined()), "must hold by construction if no predicates");
196 } else {
197 // There is at least one empty predicate. When calling 'skip_loop_predicates' on each found empty predicate,
198 // we should end up at 'proj_true'.
199 Node* proj_before_first_empty_predicate = skip_loop_predicates(entry);
200 if (UseProfiledLoopPredicate) {
201 predicate = find_predicate(proj_before_first_empty_predicate);
202 if (predicate != NULL) {
203 proj_before_first_empty_predicate = skip_loop_predicates(predicate);
204 }
205 }
206 if (UseLoopPredicate) {
207 predicate = find_predicate(proj_before_first_empty_predicate);
208 if (predicate != NULL) {
209 proj_before_first_empty_predicate = skip_loop_predicates(predicate);
210 }
211 }
212 assert(proj_true == proj_before_first_empty_predicate, "must hold by construction if at least one predicate");
213 }
214 #endif
215 // Increment unswitch count
216 LoopNode* head_clone = old_new[head->_idx]->as_Loop();
217 int nct = head->unswitch_count() + 1;
218 head->set_unswitch_count(nct);
219 head_clone->set_unswitch_count(nct);
220
221 // Hoist invariant casts out of each loop to the appropriate control projection.
222 Node_List worklist;
223 for (uint i = 0; i < unswitch_iffs.size(); i++) {
224 IfNode* iff = unswitch_iffs.at(i)->as_If();
225 for (DUIterator_Fast imax, i = iff->fast_outs(imax); i < imax; i++) {
226 ProjNode* proj = iff->fast_out(i)->as_Proj();
227 // Copy to a worklist for easier manipulation
228 for (DUIterator_Fast jmax, j = proj->fast_outs(jmax); j < jmax; j++) {
229 Node* use = proj->fast_out(j);
230 if (use->Opcode() == Op_CheckCastPP && loop->is_invariant(use->in(1))) {
231 worklist.push(use);
232 }
233 }
234 ProjNode* invar_proj = invar_iff->proj_out(proj->_con)->as_Proj();
235 while (worklist.size() > 0) {
236 Node* use = worklist.pop();
237 Node* nuse = use->clone();
238 nuse->set_req(0, invar_proj);
239 _igvn.replace_input_of(use, 1, nuse);
240 register_new_node(nuse, invar_proj);
241 // Same for the clone if we are removing checks from the slow loop
242 if (!flat_array_checks) {
243 Node* use_clone = old_new[use->_idx];
244 _igvn.replace_input_of(use_clone, 1, nuse);
245 }
246 }
247 }
248 }
249
250 // Hardwire the control paths in the loops into if(true) and if(false)
251 for (uint i = 0; i < unswitch_iffs.size(); i++) {
252 IfNode* iff = unswitch_iffs.at(i)->as_If();
253 _igvn.rehash_node_delayed(iff);
254 dominated_by(proj_true, iff);
255 }
256 IfNode* unswitch_iff_clone = old_new[unswitch_iff->_idx]->as_If();
257 if (!flat_array_checks) {
258 ProjNode* proj_false = invar_iff->proj_out(0)->as_Proj();
259 _igvn.rehash_node_delayed(unswitch_iff_clone);
260 dominated_by(proj_false, unswitch_iff_clone);
261 } else {
262 // Leave the flattened array checks in the slow loop and
263 // prevent it from being unswitched again based on these checks.
264 head_clone->mark_flattened_arrays();
265 }
266
267 // Reoptimize loops
268 loop->record_for_igvn();
269 for(int i = loop->_body.size() - 1; i >= 0 ; i--) {
270 Node *n = loop->_body[i];
271 Node *n_clone = old_new[n->_idx];
272 _igvn._worklist.push(n_clone);
273 }
274
275 #ifndef PRODUCT
276 if (TraceLoopUnswitching) {
277 for (uint i = 0; i < unswitch_iffs.size(); i++) {
278 tty->print_cr("Loop unswitching orig: %d @ %d new: %d @ %d",
279 head->_idx, unswitch_iffs.at(i)->_idx,
280 old_new[head->_idx]->_idx, old_new[unswitch_iffs.at(i)->_idx]->_idx);
281 }
282 }
283 #endif
284
285 C->set_major_progress();
286 }
287
288 //-------------------------create_slow_version_of_loop------------------------
289 // Create a slow version of the loop by cloning the loop
290 // and inserting an if to select fast-slow versions.
291 // Return the inserted if.
292 IfNode* PhaseIdealLoop::create_slow_version_of_loop(IdealLoopTree *loop,
293 Node_List &old_new,
294 Node_List &unswitch_iffs,
295 CloneLoopMode mode) {
296 LoopNode* head = loop->_head->as_Loop();
297 bool counted_loop = head->is_CountedLoop();
298 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
299 _igvn.rehash_node_delayed(entry);
300 IdealLoopTree* outer_loop = loop->_parent;
301
302 head->verify_strip_mined(1);
303
304 // Add test to new "if" outside of loop
305 IfNode* unswitch_iff = unswitch_iffs.at(0)->as_If();
306 BoolNode* bol = unswitch_iff->in(1)->as_Bool();
307 if (unswitch_iffs.size() > 1) {
308 // Flattened array checks are used on array access to switch between
309 // a legacy object array access and a flattened inline type array
310 // access. We want the performance impact on legacy accesses to be
311 // as small as possible so we make two copies of the loop: a fast
312 // one where all accesses are known to be legacy, a slow one where
313 // some accesses are to flattened arrays. Flattened array checks
314 // can be removed from the fast loop (true proj) but not from the
315 // slow loop (false proj) as it can have a mix of flattened/legacy accesses.
316 assert(bol->_test._test == BoolTest::ne, "IfTrue proj must point to flat array");
317 bol = bol->clone()->as_Bool();
318 register_new_node(bol, entry);
319 FlatArrayCheckNode* cmp = bol->in(1)->clone()->as_FlatArrayCheck();
320 register_new_node(cmp, entry);
321 bol->set_req(1, cmp);
322 // Combine all checks into a single one that fails if one array is flattened
323 assert(cmp->req() == 3, "unexpected number of inputs for FlatArrayCheck");
324 cmp->add_req_batch(C->top(), unswitch_iffs.size() - 1);
325 for (uint i = 0; i < unswitch_iffs.size(); i++) {
326 Node* array = unswitch_iffs.at(i)->in(1)->in(1)->in(FlatArrayCheckNode::Array);
327 cmp->set_req(FlatArrayCheckNode::Array + i, array);
328 }
329 }
330
331 IfNode* iff = (unswitch_iff->Opcode() == Op_RangeCheck) ? new RangeCheckNode(entry, bol, unswitch_iff->_prob, unswitch_iff->_fcnt) :
332 new IfNode(entry, bol, unswitch_iff->_prob, unswitch_iff->_fcnt);
333 register_node(iff, outer_loop, entry, dom_depth(entry));
334 ProjNode* iffast = new IfTrueNode(iff);
335 register_node(iffast, outer_loop, iff, dom_depth(iff));
336 ProjNode* ifslow = new IfFalseNode(iff);
337 register_node(ifslow, outer_loop, iff, dom_depth(iff));
338
339 // Clone the loop body. The clone becomes the slow loop. The
340 // original pre-header will (illegally) have 3 control users
341 // (old & new loops & new if).
342 clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), mode, iff);
343 assert(old_new[head->_idx]->is_Loop(), "" );
344
345 // Fast (true) and Slow (false) control
346 ProjNode* iffast_pred = iffast;
347 ProjNode* ifslow_pred = ifslow;
348 clone_predicates_to_unswitched_loop(loop, old_new, iffast_pred, ifslow_pred);
349
350 Node* l = head->skip_strip_mined();
351 _igvn.replace_input_of(l, LoopNode::EntryControl, iffast_pred);
352 set_idom(l, iffast_pred, dom_depth(l));
353 LoopNode* slow_l = old_new[head->_idx]->as_Loop()->skip_strip_mined();
354 _igvn.replace_input_of(slow_l, LoopNode::EntryControl, ifslow_pred);
355 set_idom(slow_l, ifslow_pred, dom_depth(l));
356
357 recompute_dom_depth();
358
359 return iff;
360 }
361
362 LoopNode* PhaseIdealLoop::create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk) {
363 Node_List old_new;
364 LoopNode* head = loop->_head->as_Loop();
365 bool counted_loop = head->is_CountedLoop();
366 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
367 _igvn.rehash_node_delayed(entry);
368 IdealLoopTree* outer_loop = head->is_strip_mined() ? loop->_parent->_parent : loop->_parent;
369
370 ConINode* const_1 = _igvn.intcon(1);
371 set_ctrl(const_1, C->root());
372 IfNode* iff = new IfNode(entry, const_1, PROB_MAX, COUNT_UNKNOWN);
373 register_node(iff, outer_loop, entry, dom_depth(entry));
374 ProjNode* iffast = new IfTrueNode(iff);
375 register_node(iffast, outer_loop, iff, dom_depth(iff));
376 ProjNode* ifslow = new IfFalseNode(iff);
377 register_node(ifslow, outer_loop, iff, dom_depth(iff));
378
379 // Clone the loop body. The clone becomes the slow loop. The
380 // original pre-header will (illegally) have 3 control users
381 // (old & new loops & new if).
382 clone_loop(loop, old_new, dom_depth(head), CloneIncludesStripMined, iff);
383 assert(old_new[head->_idx]->is_Loop(), "" );
384
385 LoopNode* slow_head = old_new[head->_idx]->as_Loop();
386
387 #ifndef PRODUCT
388 if (TraceLoopOpts) {
389 tty->print_cr("PhaseIdealLoop::create_reserve_version_of_loop:");
390 tty->print("\t iff = %d, ", iff->_idx); iff->dump();
391 tty->print("\t iffast = %d, ", iffast->_idx); iffast->dump();
392 tty->print("\t ifslow = %d, ", ifslow->_idx); ifslow->dump();
393 tty->print("\t before replace_input_of: head = %d, ", head->_idx); head->dump();
394 tty->print("\t before replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
395 }
396 #endif
397
398 // Fast (true) control
399 _igvn.replace_input_of(head->skip_strip_mined(), LoopNode::EntryControl, iffast);
400 // Slow (false) control
401 _igvn.replace_input_of(slow_head->skip_strip_mined(), LoopNode::EntryControl, ifslow);
402
403 recompute_dom_depth();
404
405 lk->set_iff(iff);
406
407 #ifndef PRODUCT
408 if (TraceLoopOpts ) {
409 tty->print("\t after replace_input_of: head = %d, ", head->_idx); head->dump();
410 tty->print("\t after replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
411 }
412 #endif
413
414 return slow_head->as_Loop();
415 }
416
417 CountedLoopReserveKit::CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active = true) :
418 _phase(phase),
419 _lpt(loop),
420 _lp(NULL),
421 _iff(NULL),
422 _lp_reserved(NULL),
423 _has_reserved(false),
424 _use_new(false),
425 _active(active)
426 {
427 create_reserve();
428 };
429
430 CountedLoopReserveKit::~CountedLoopReserveKit() {
431 if (!_active) {
432 return;
433 }
434
435 if (_has_reserved && !_use_new) {
436 // intcon(0)->iff-node reverts CF to the reserved copy
437 ConINode* const_0 = _phase->_igvn.intcon(0);
438 _phase->set_ctrl(const_0, _phase->C->root());
439 _iff->set_req(1, const_0);
440
441 #ifndef PRODUCT
442 if (TraceLoopOpts) {
443 tty->print_cr("CountedLoopReserveKit::~CountedLoopReserveKit()");
444 tty->print("\t discard loop %d and revert to the reserved loop clone %d: ", _lp->_idx, _lp_reserved->_idx);
445 _lp_reserved->dump();
446 }
447 #endif
448 }
449 }
450
451 bool CountedLoopReserveKit::create_reserve() {
452 if (!_active) {
453 return false;
454 }
455
456 if(!_lpt->_head->is_CountedLoop()) {
457 if (TraceLoopOpts) {
458 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not counted loop", _lpt->_head->_idx);
459 }
460 return false;
461 }
462 CountedLoopNode *cl = _lpt->_head->as_CountedLoop();
463 if (!cl->is_valid_counted_loop(T_INT)) {
464 if (TraceLoopOpts) {
465 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not valid counted loop", cl->_idx);
466 }
467 return false; // skip malformed counted loop
468 }
469 if (!cl->is_main_loop()) {
470 bool loop_not_canonical = true;
471 if (cl->is_post_loop() && (cl->slp_max_unroll() > 0)) {
472 loop_not_canonical = false;
473 }
474 // only reject some loop forms
475 if (loop_not_canonical) {
476 if (TraceLoopOpts) {
477 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not canonical loop", cl->_idx);
478 }
479 return false; // skip normal, pre, and post (conditionally) loops
480 }
481 }
482
483 _lp = _lpt->_head->as_Loop();
484 _lp_reserved = _phase->create_reserve_version_of_loop(_lpt, this);
485
486 if (!_lp_reserved->is_CountedLoop()) {
487 return false;
488 }
489
490 Node* ifslow_pred = _lp_reserved->skip_strip_mined()->in(LoopNode::EntryControl);
491
492 if (!ifslow_pred->is_IfFalse()) {
493 return false;
494 }
495
496 Node* iff = ifslow_pred->in(0);
497 if (!iff->is_If() || iff != _iff) {
498 return false;
499 }
500
501 if (iff->in(1)->Opcode() != Op_ConI) {
502 return false;
503 }
504
505 _has_reserved = true;
506 return true;
507 }