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 flattened_checks;
86 if (phase->find_unswitching_candidate(this, flattened_checks) == NULL && flattened_checks.size() == 0) {
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& flattened_checks) 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
122 Node* array;
123 if (unswitch_iff == NULL || unswitch_iff->is_flattened_array_check(&_igvn, array)) {
124 // collect all flattened array checks
125 for (uint i = 0; i < loop->_body.size(); i++) {
126 Node* n = loop->_body.at(i);
127 if (n->is_If() && n->as_If()->is_flattened_array_check(&_igvn, array) &&
128 loop->is_invariant(n->in(1)) &&
129 !loop->is_loop_exit(n)) {
130 flattened_checks.push(n);
131 }
132 }
133 if (flattened_checks.size() > 1) {
134 unswitch_iff = NULL;
135 } else {
136 flattened_checks.clear();
137 }
138 }
139
140 return unswitch_iff;
141 }
142
143 //------------------------------do_unswitching-----------------------------
144 // Clone loop with an invariant test (that does not exit) and
145 // insert a clone of the test that selects which version to
146 // execute.
147 void PhaseIdealLoop::do_unswitching(IdealLoopTree *loop, Node_List &old_new) {
148
149 // Find first invariant test that doesn't exit the loop
150 LoopNode *head = loop->_head->as_Loop();
151
152 Node_List flattened_checks;
153 IfNode* unswitch_iff = find_unswitching_candidate((const IdealLoopTree *)loop, flattened_checks);
154 assert(unswitch_iff != NULL || flattened_checks.size() > 0, "should be at least one");
155 if (unswitch_iff == NULL) {
156 unswitch_iff = flattened_checks.at(0)->as_If();
157 }
158
159 #ifndef PRODUCT
160 if (TraceLoopOpts) {
161 tty->print("Unswitch %d ", head->unswitch_count()+1);
162 loop->dump_head();
163 }
164 #endif
165
166 // Need to revert back to normal loop
167 if (head->is_CountedLoop() && !head->as_CountedLoop()->is_normal_loop()) {
168 head->as_CountedLoop()->set_normal_loop();
169 }
170
171 ProjNode* proj_true = create_slow_version_of_loop(loop, old_new, unswitch_iff->Opcode(), CloneIncludesStripMined);
172
173 #ifdef ASSERT
174 Node* uniqc = proj_true->unique_ctrl_out();
175 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
176 Node* predicate = find_predicate(entry);
177 if (predicate != NULL) {
178 entry = skip_loop_predicates(entry);
179 }
180 if (predicate != NULL && UseLoopPredicate) {
181 // We may have two predicates, find first.
182 Node* n = find_predicate(entry);
183 if (n != NULL) {
184 predicate = n;
185 entry = skip_loop_predicates(entry);
186 }
187 }
188 if (predicate != NULL && UseProfiledLoopPredicate) {
189 entry = find_predicate(entry);
190 if (entry != NULL) predicate = entry;
191 }
192 if (predicate != NULL) predicate = predicate->in(0);
193 assert(proj_true->is_IfTrue() &&
194 (predicate == NULL && uniqc == head && !head->is_strip_mined() ||
195 predicate == NULL && uniqc == head->in(LoopNode::EntryControl) && head->is_strip_mined() ||
196 predicate != NULL && uniqc == predicate), "by construction");
197 #endif
198 // Increment unswitch count
199 LoopNode* head_clone = old_new[head->_idx]->as_Loop();
200 int nct = head->unswitch_count() + 1;
201 head->set_unswitch_count(nct);
202 head_clone->set_unswitch_count(nct);
203 if (flattened_checks.size() > 0) {
204 head->mark_flattened_arrays();
205 }
206
207 // Add test to new "if" outside of loop
208 IfNode* invar_iff = proj_true->in(0)->as_If();
209 Node* invar_iff_c = invar_iff->in(0);
210 invar_iff->_prob = unswitch_iff->_prob;
211 if (flattened_checks.size() > 0) {
212 // Flattened array checks are used in
213 // Parse::array_store()/Parse::array_load() to switch between a
214 // legacy object array access and a flattened value array
215 // access. We want the performance impact on legacy accesses to be
216 // as small as possible so we make 2 copies of the loops: a fast
217 // one where all accesses are known to be legacy, a slow one where
218 // some accesses are to flattened arrays. Flattened array checks
219 // can be removed from the first one but not from the second one
220 // as it can have a mix of flattened/legacy accesses.
221 BoolNode* bol = unswitch_iff->in(1)->clone()->as_Bool();
222 register_new_node(bol, invar_iff->in(0));
223 Node* cmp = bol->in(1)->clone();
224 register_new_node(cmp, invar_iff->in(0));
225 bol->set_req(1, cmp);
226 Node* in1 = NULL;
227 for (uint i = 0; i < flattened_checks.size(); i++) {
228 Node* v = flattened_checks.at(i)->in(1)->in(1)->in(1);
229 if (in1 == NULL) {
230 in1 = v;
231 } else {
232 if (cmp->Opcode() == Op_CmpL) {
233 in1 = new OrLNode(in1, v);
234 } else {
235 in1 = new OrINode(in1, v);
236 }
237 register_new_node(in1, invar_iff->in(0));
238 }
239 }
240 cmp->set_req(1, in1);
241 invar_iff->set_req(1, bol);
242 } else {
243 BoolNode* bol = unswitch_iff->in(1)->as_Bool();
244 invar_iff->set_req(1, bol);
245 }
246
247 ProjNode* proj_false = invar_iff->proj_out(0)->as_Proj();
248
249 // Hoist invariant casts out of each loop to the appropriate
250 // control projection.
251
252 Node_List worklist;
253
254 if (flattened_checks.size() > 0) {
255 for (uint i = 0; i < flattened_checks.size(); i++) {
256 IfNode* iff = flattened_checks.at(i)->as_If();
257 ProjNode* proj= iff->proj_out(0)->as_Proj();
258 // Copy to a worklist for easier manipulation
259 for (DUIterator_Fast jmax, j = proj->fast_outs(jmax); j < jmax; j++) {
260 Node* use = proj->fast_out(j);
261 if (use->Opcode() == Op_CheckCastPP && loop->is_invariant(use->in(1))) {
262 worklist.push(use);
263 }
264 }
265 ProjNode* invar_proj = invar_iff->proj_out(proj->_con)->as_Proj();
266 while (worklist.size() > 0) {
267 Node* use = worklist.pop();
268 Node* nuse = use->clone();
269 nuse->set_req(0, invar_proj);
270 _igvn.replace_input_of(use, 1, nuse);
271 register_new_node(nuse, invar_proj);
272 // Same for the clone
273 Node* use_clone = old_new[use->_idx];
274 _igvn.replace_input_of(use_clone, 1, nuse);
275 }
276 }
277 } else {
278 for (DUIterator_Fast imax, i = unswitch_iff->fast_outs(imax); i < imax; i++) {
279 ProjNode* proj= unswitch_iff->fast_out(i)->as_Proj();
280 // Copy to a worklist for easier manipulation
281 for (DUIterator_Fast jmax, j = proj->fast_outs(jmax); j < jmax; j++) {
282 Node* use = proj->fast_out(j);
283 if (use->Opcode() == Op_CheckCastPP && loop->is_invariant(use->in(1))) {
284 worklist.push(use);
285 }
286 }
287 ProjNode* invar_proj = invar_iff->proj_out(proj->_con)->as_Proj();
288 while (worklist.size() > 0) {
289 Node* use = worklist.pop();
290 Node* nuse = use->clone();
291 nuse->set_req(0, invar_proj);
292 _igvn.replace_input_of(use, 1, nuse);
293 register_new_node(nuse, invar_proj);
294 // Same for the clone
295 Node* use_clone = old_new[use->_idx];
296 _igvn.replace_input_of(use_clone, 1, nuse);
297 }
298 }
299 }
300
301 IfNode* unswitch_iff_clone = old_new[unswitch_iff->_idx]->as_If();
302 if (flattened_checks.size() > 0) {
303 for (uint i = 0; i < flattened_checks.size(); i++) {
304 IfNode* iff = flattened_checks.at(i)->as_If();
305 _igvn.rehash_node_delayed(iff);
306 short_circuit_if(old_new[iff->_idx]->as_If(), proj_false);
307 }
308 } else {
309 // Hardwire the control paths in the loops into if(true) and if(false)
310 _igvn.rehash_node_delayed(unswitch_iff);
311 short_circuit_if(unswitch_iff, proj_true);
312
313 _igvn.rehash_node_delayed(unswitch_iff_clone);
314 short_circuit_if(unswitch_iff_clone, proj_false);
315 }
316
317 // Reoptimize loops
318 loop->record_for_igvn();
319 for(int i = loop->_body.size() - 1; i >= 0 ; i--) {
320 Node *n = loop->_body[i];
321 Node *n_clone = old_new[n->_idx];
322 _igvn._worklist.push(n_clone);
323 }
324
325 #ifndef PRODUCT
326 if (TraceLoopUnswitching) {
327 tty->print_cr("Loop unswitching orig: %d @ %d new: %d @ %d",
328 head->_idx, unswitch_iff->_idx,
329 old_new[head->_idx]->_idx, unswitch_iff_clone->_idx);
330 }
331 #endif
332
333 C->set_major_progress();
334 }
335
336 //-------------------------create_slow_version_of_loop------------------------
337 // Create a slow version of the loop by cloning the loop
338 // and inserting an if to select fast-slow versions.
339 // Return control projection of the entry to the fast version.
340 ProjNode* PhaseIdealLoop::create_slow_version_of_loop(IdealLoopTree *loop,
341 Node_List &old_new,
342 int opcode,
343 CloneLoopMode mode) {
344 LoopNode* head = loop->_head->as_Loop();
345 bool counted_loop = head->is_CountedLoop();
346 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
347 _igvn.rehash_node_delayed(entry);
348 IdealLoopTree* outer_loop = loop->_parent;
349
350 head->verify_strip_mined(1);
351
352 Node *cont = _igvn.intcon(1);
353 set_ctrl(cont, C->root());
354 Node* opq = new Opaque1Node(C, cont);
355 register_node(opq, outer_loop, entry, dom_depth(entry));
356 Node *bol = new Conv2BNode(opq);
357 register_node(bol, outer_loop, entry, dom_depth(entry));
358 IfNode* iff = (opcode == Op_RangeCheck) ? new RangeCheckNode(entry, bol, PROB_MAX, COUNT_UNKNOWN) :
359 new IfNode(entry, bol, PROB_MAX, COUNT_UNKNOWN);
360 register_node(iff, outer_loop, entry, dom_depth(entry));
361 ProjNode* iffast = new IfTrueNode(iff);
362 register_node(iffast, outer_loop, iff, dom_depth(iff));
363 ProjNode* ifslow = new IfFalseNode(iff);
364 register_node(ifslow, outer_loop, iff, dom_depth(iff));
365
366 // Clone the loop body. The clone becomes the fast loop. The
367 // original pre-header will (illegally) have 3 control users
368 // (old & new loops & new if).
369 clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), mode, iff);
370 assert(old_new[head->_idx]->is_Loop(), "" );
371
372 // Fast (true) control
373 Node* iffast_pred = clone_loop_predicates(entry, iffast, !counted_loop);
374
375 // Slow (false) control
376 Node* ifslow_pred = clone_loop_predicates(entry, ifslow, !counted_loop);
377
378 Node* l = head->skip_strip_mined();
379 _igvn.replace_input_of(l, LoopNode::EntryControl, iffast_pred);
380 set_idom(l, iffast_pred, dom_depth(l));
381 LoopNode* slow_l = old_new[head->_idx]->as_Loop()->skip_strip_mined();
382 _igvn.replace_input_of(slow_l, LoopNode::EntryControl, ifslow_pred);
383 set_idom(slow_l, ifslow_pred, dom_depth(l));
384
385 recompute_dom_depth();
386
387 return iffast;
388 }
389
390 LoopNode* PhaseIdealLoop::create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk) {
391 Node_List old_new;
392 LoopNode* head = loop->_head->as_Loop();
393 bool counted_loop = head->is_CountedLoop();
394 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
395 _igvn.rehash_node_delayed(entry);
396 IdealLoopTree* outer_loop = head->is_strip_mined() ? loop->_parent->_parent : loop->_parent;
397
398 ConINode* const_1 = _igvn.intcon(1);
399 set_ctrl(const_1, C->root());
400 IfNode* iff = new IfNode(entry, const_1, PROB_MAX, COUNT_UNKNOWN);
401 register_node(iff, outer_loop, entry, dom_depth(entry));
402 ProjNode* iffast = new IfTrueNode(iff);
403 register_node(iffast, outer_loop, iff, dom_depth(iff));
404 ProjNode* ifslow = new IfFalseNode(iff);
405 register_node(ifslow, outer_loop, iff, dom_depth(iff));
406
407 // Clone the loop body. The clone becomes the fast loop. The
408 // original pre-header will (illegally) have 3 control users
409 // (old & new loops & new if).
410 clone_loop(loop, old_new, dom_depth(head), CloneIncludesStripMined, iff);
411 assert(old_new[head->_idx]->is_Loop(), "" );
412
413 LoopNode* slow_head = old_new[head->_idx]->as_Loop();
414
415 #ifndef PRODUCT
416 if (TraceLoopOpts) {
417 tty->print_cr("PhaseIdealLoop::create_reserve_version_of_loop:");
418 tty->print("\t iff = %d, ", iff->_idx); iff->dump();
419 tty->print("\t iffast = %d, ", iffast->_idx); iffast->dump();
420 tty->print("\t ifslow = %d, ", ifslow->_idx); ifslow->dump();
421 tty->print("\t before replace_input_of: head = %d, ", head->_idx); head->dump();
422 tty->print("\t before replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
423 }
424 #endif
425
426 // Fast (true) control
427 _igvn.replace_input_of(head->skip_strip_mined(), LoopNode::EntryControl, iffast);
428 // Slow (false) control
429 _igvn.replace_input_of(slow_head->skip_strip_mined(), LoopNode::EntryControl, ifslow);
430
431 recompute_dom_depth();
432
433 lk->set_iff(iff);
434
435 #ifndef PRODUCT
436 if (TraceLoopOpts ) {
437 tty->print("\t after replace_input_of: head = %d, ", head->_idx); head->dump();
438 tty->print("\t after replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
439 }
440 #endif
441
442 return slow_head->as_Loop();
443 }
444
445 CountedLoopReserveKit::CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active = true) :
446 _phase(phase),
447 _lpt(loop),
448 _lp(NULL),
449 _iff(NULL),
450 _lp_reserved(NULL),
451 _has_reserved(false),
452 _use_new(false),
453 _active(active)
454 {
455 create_reserve();
456 };
457
458 CountedLoopReserveKit::~CountedLoopReserveKit() {
459 if (!_active) {
460 return;
461 }
462
463 if (_has_reserved && !_use_new) {
464 // intcon(0)->iff-node reverts CF to the reserved copy
465 ConINode* const_0 = _phase->_igvn.intcon(0);
466 _phase->set_ctrl(const_0, _phase->C->root());
467 _iff->set_req(1, const_0);
468
469 #ifndef PRODUCT
470 if (TraceLoopOpts) {
471 tty->print_cr("CountedLoopReserveKit::~CountedLoopReserveKit()");
472 tty->print("\t discard loop %d and revert to the reserved loop clone %d: ", _lp->_idx, _lp_reserved->_idx);
473 _lp_reserved->dump();
474 }
475 #endif
476 }
477 }
478
479 bool CountedLoopReserveKit::create_reserve() {
480 if (!_active) {
481 return false;
482 }
483
484 if(!_lpt->_head->is_CountedLoop()) {
485 if (TraceLoopOpts) {
486 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not counted loop", _lpt->_head->_idx);
487 }
488 return false;
489 }
490 CountedLoopNode *cl = _lpt->_head->as_CountedLoop();
491 if (!cl->is_valid_counted_loop()) {
492 if (TraceLoopOpts) {
493 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not valid counted loop", cl->_idx);
494 }
495 return false; // skip malformed counted loop
496 }
497 if (!cl->is_main_loop()) {
498 bool loop_not_canonical = true;
499 if (cl->is_post_loop() && (cl->slp_max_unroll() > 0)) {
500 loop_not_canonical = false;
501 }
502 // only reject some loop forms
503 if (loop_not_canonical) {
504 if (TraceLoopOpts) {
505 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not canonical loop", cl->_idx);
506 }
507 return false; // skip normal, pre, and post (conditionally) loops
508 }
509 }
510
511 _lp = _lpt->_head->as_Loop();
512 _lp_reserved = _phase->create_reserve_version_of_loop(_lpt, this);
513
514 if (!_lp_reserved->is_CountedLoop()) {
515 return false;
516 }
517
518 Node* ifslow_pred = _lp_reserved->skip_strip_mined()->in(LoopNode::EntryControl);
519
520 if (!ifslow_pred->is_IfFalse()) {
521 return false;
522 }
523
524 Node* iff = ifslow_pred->in(0);
525 if (!iff->is_If() || iff != _iff) {
526 return false;
527 }
528
529 if (iff->in(1)->Opcode() != Op_ConI) {
530 return false;
531 }
532
533 return _has_reserved = true;
534 }