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