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/connode.hpp"
28 #include "opto/convertnode.hpp"
29 #include "opto/loopnode.hpp"
30 #include "opto/opaquenode.hpp"
31 #include "opto/rootnode.hpp"
32
33 //================= Loop Unswitching =====================
34 //
35 // orig: transformed:
36 // if (invariant-test) then
37 // predicate predicate
38 // loop loop
39 // stmt1 stmt1
40 // if (invariant-test) then stmt2
41 // stmt2 stmt4
42 // else endloop
43 // stmt3 else
44 // endif predicate [clone]
45 // stmt4 loop [clone]
46 // endloop stmt1 [clone]
47 // stmt3
48 // stmt4 [clone]
49 // endloop
50 // endif
51 //
52 // Note: the "else" clause may be empty
53
54 //------------------------------policy_unswitching-----------------------------
55 // Return TRUE or FALSE if the loop should be unswitched
56 // (ie. clone loop with an invariant test that does not exit the loop)
policy_unswitching(PhaseIdealLoop * phase) const57 bool IdealLoopTree::policy_unswitching( PhaseIdealLoop *phase ) const {
58 if (!LoopUnswitching) {
59 return false;
60 }
61 if (!_head->is_Loop()) {
62 return false;
63 }
64
65 // If nodes are depleted, some transform has miscalculated its needs.
66 assert(!phase->exceeding_node_budget(), "sanity");
67
68 // check for vectorized loops, any unswitching was already applied
69 if (_head->is_CountedLoop() && _head->as_CountedLoop()->is_unroll_only()) {
70 return false;
71 }
72
73 LoopNode* head = _head->as_Loop();
74 if (head->unswitch_count() + 1 > head->unswitch_max()) {
75 return false;
76 }
77 if (phase->find_unswitching_candidate(this) == NULL) {
78 return false;
79 }
80
81 // Too speculative if running low on nodes.
82 return phase->may_require_nodes(est_loop_clone_sz(2));
83 }
84
85 //------------------------------find_unswitching_candidate-----------------------------
86 // Find candidate "if" for unswitching
find_unswitching_candidate(const IdealLoopTree * loop) const87 IfNode* PhaseIdealLoop::find_unswitching_candidate(const IdealLoopTree *loop) const {
88
89 // Find first invariant test that doesn't exit the loop
90 LoopNode *head = loop->_head->as_Loop();
91 IfNode* unswitch_iff = NULL;
92 Node* n = head->in(LoopNode::LoopBackControl);
93 while (n != head) {
94 Node* n_dom = idom(n);
95 if (n->is_Region()) {
96 if (n_dom->is_If()) {
97 IfNode* iff = n_dom->as_If();
98 if (iff->in(1)->is_Bool()) {
99 BoolNode* bol = iff->in(1)->as_Bool();
100 if (bol->in(1)->is_Cmp()) {
101 // If condition is invariant and not a loop exit,
102 // then found reason to unswitch.
103 if (loop->is_invariant(bol) && !loop->is_loop_exit(iff)) {
104 unswitch_iff = iff;
105 }
106 }
107 }
108 }
109 }
110 n = n_dom;
111 }
112 return unswitch_iff;
113 }
114
115 //------------------------------do_unswitching-----------------------------
116 // Clone loop with an invariant test (that does not exit) and
117 // insert a clone of the test that selects which version to
118 // execute.
do_unswitching(IdealLoopTree * loop,Node_List & old_new)119 void PhaseIdealLoop::do_unswitching(IdealLoopTree *loop, Node_List &old_new) {
120
121 LoopNode *head = loop->_head->as_Loop();
122 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
123 if (find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check) != NULL
124 || (UseProfiledLoopPredicate && find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate) != NULL)
125 || (UseLoopPredicate && find_predicate_insertion_point(entry, Deoptimization::Reason_predicate) != NULL)) {
126 assert(entry->is_IfProj(), "sanity - must be ifProj since there is at least one predicate");
127 if (entry->outcnt() > 1) {
128 // Bailout if there are loop predicates from which there are additional control dependencies (i.e. from
129 // loop entry 'entry') to previously partially peeled statements since this case is not handled and can lead
130 // to wrong execution. Remove this bailout, once this is fixed.
131 return;
132 }
133 }
134 // Find first invariant test that doesn't exit the loop
135 IfNode* unswitch_iff = find_unswitching_candidate((const IdealLoopTree *)loop);
136 assert(unswitch_iff != NULL, "should be at least one");
137
138 #ifndef PRODUCT
139 if (TraceLoopOpts) {
140 tty->print("Unswitch %d ", head->unswitch_count()+1);
141 loop->dump_head();
142 }
143 #endif
144
145 // Need to revert back to normal loop
146 if (head->is_CountedLoop() && !head->as_CountedLoop()->is_normal_loop()) {
147 head->as_CountedLoop()->set_normal_loop();
148 }
149
150 ProjNode* proj_true = create_slow_version_of_loop(loop, old_new, unswitch_iff->Opcode(), CloneIncludesStripMined);
151
152 #ifdef ASSERT
153 assert(proj_true->is_IfTrue(), "must be true projection");
154 entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
155 Node* predicate = find_predicate(entry);
156 if (predicate == NULL) {
157 // No empty predicate
158 Node* uniqc = proj_true->unique_ctrl_out();
159 assert((uniqc == head && !head->is_strip_mined()) || (uniqc == head->in(LoopNode::EntryControl)
160 && head->is_strip_mined()), "must hold by construction if no predicates");
161 } else {
162 // There is at least one empty predicate. When calling 'skip_loop_predicates' on each found empty predicate,
163 // we should end up at 'proj_true'.
164 Node* proj_before_first_empty_predicate = skip_loop_predicates(entry);
165 if (UseProfiledLoopPredicate) {
166 predicate = find_predicate(proj_before_first_empty_predicate);
167 if (predicate != NULL) {
168 proj_before_first_empty_predicate = skip_loop_predicates(predicate);
169 }
170 }
171 if (UseLoopPredicate) {
172 predicate = find_predicate(proj_before_first_empty_predicate);
173 if (predicate != NULL) {
174 proj_before_first_empty_predicate = skip_loop_predicates(predicate);
175 }
176 }
177 assert(proj_true == proj_before_first_empty_predicate, "must hold by construction if at least one predicate");
178 }
179 #endif
180 // Increment unswitch count
181 LoopNode* head_clone = old_new[head->_idx]->as_Loop();
182 int nct = head->unswitch_count() + 1;
183 head->set_unswitch_count(nct);
184 head_clone->set_unswitch_count(nct);
185
186 // Add test to new "if" outside of loop
187 IfNode* invar_iff = proj_true->in(0)->as_If();
188 Node* invar_iff_c = invar_iff->in(0);
189 BoolNode* bol = unswitch_iff->in(1)->as_Bool();
190 invar_iff->set_req(1, bol);
191 invar_iff->_prob = unswitch_iff->_prob;
192
193 ProjNode* proj_false = invar_iff->proj_out(0)->as_Proj();
194
195 // Hoist invariant casts out of each loop to the appropriate
196 // control projection.
197
198 Node_List worklist;
199
200 for (DUIterator_Fast imax, i = unswitch_iff->fast_outs(imax); i < imax; i++) {
201 ProjNode* proj= unswitch_iff->fast_out(i)->as_Proj();
202 // Copy to a worklist for easier manipulation
203 for (DUIterator_Fast jmax, j = proj->fast_outs(jmax); j < jmax; j++) {
204 Node* use = proj->fast_out(j);
205 if (use->Opcode() == Op_CheckCastPP && loop->is_invariant(use->in(1))) {
206 worklist.push(use);
207 }
208 }
209 ProjNode* invar_proj = invar_iff->proj_out(proj->_con)->as_Proj();
210 while (worklist.size() > 0) {
211 Node* use = worklist.pop();
212 Node* nuse = use->clone();
213 nuse->set_req(0, invar_proj);
214 _igvn.replace_input_of(use, 1, nuse);
215 register_new_node(nuse, invar_proj);
216 // Same for the clone
217 Node* use_clone = old_new[use->_idx];
218 _igvn.replace_input_of(use_clone, 1, nuse);
219 }
220 }
221
222 // Hardwire the control paths in the loops into if(true) and if(false)
223 _igvn.rehash_node_delayed(unswitch_iff);
224 dominated_by(proj_true, unswitch_iff, false, false);
225
226 IfNode* unswitch_iff_clone = old_new[unswitch_iff->_idx]->as_If();
227 _igvn.rehash_node_delayed(unswitch_iff_clone);
228 dominated_by(proj_false, unswitch_iff_clone, false, false);
229
230 // Reoptimize loops
231 loop->record_for_igvn();
232 for(int i = loop->_body.size() - 1; i >= 0 ; i--) {
233 Node *n = loop->_body[i];
234 Node *n_clone = old_new[n->_idx];
235 _igvn._worklist.push(n_clone);
236 }
237
238 #ifndef PRODUCT
239 if (TraceLoopUnswitching) {
240 tty->print_cr("Loop unswitching orig: %d @ %d new: %d @ %d",
241 head->_idx, unswitch_iff->_idx,
242 old_new[head->_idx]->_idx, unswitch_iff_clone->_idx);
243 }
244 #endif
245
246 C->set_major_progress();
247 }
248
249 //-------------------------create_slow_version_of_loop------------------------
250 // Create a slow version of the loop by cloning the loop
251 // and inserting an if to select fast-slow versions.
252 // Return control projection of the entry to the fast version.
create_slow_version_of_loop(IdealLoopTree * loop,Node_List & old_new,int opcode,CloneLoopMode mode)253 ProjNode* PhaseIdealLoop::create_slow_version_of_loop(IdealLoopTree *loop,
254 Node_List &old_new,
255 int opcode,
256 CloneLoopMode mode) {
257 LoopNode* head = loop->_head->as_Loop();
258 bool counted_loop = head->is_CountedLoop();
259 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
260 _igvn.rehash_node_delayed(entry);
261 IdealLoopTree* outer_loop = loop->_parent;
262
263 head->verify_strip_mined(1);
264
265 Node *cont = _igvn.intcon(1);
266 set_ctrl(cont, C->root());
267 Node* opq = new Opaque1Node(C, cont);
268 register_node(opq, outer_loop, entry, dom_depth(entry));
269 Node *bol = new Conv2BNode(opq);
270 register_node(bol, outer_loop, entry, dom_depth(entry));
271 IfNode* iff = (opcode == Op_RangeCheck) ? new RangeCheckNode(entry, bol, PROB_MAX, COUNT_UNKNOWN) :
272 new IfNode(entry, bol, PROB_MAX, COUNT_UNKNOWN);
273 register_node(iff, outer_loop, entry, dom_depth(entry));
274 ProjNode* iffast = new IfTrueNode(iff);
275 register_node(iffast, outer_loop, iff, dom_depth(iff));
276 ProjNode* ifslow = new IfFalseNode(iff);
277 register_node(ifslow, outer_loop, iff, dom_depth(iff));
278
279 // Clone the loop body. The clone becomes the slow loop. The
280 // original pre-header will (illegally) have 3 control users
281 // (old & new loops & new if).
282 clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), mode, iff);
283 assert(old_new[head->_idx]->is_Loop(), "" );
284
285 // Fast (true) and Slow (false) control
286 ProjNode* iffast_pred = iffast;
287 ProjNode* ifslow_pred = ifslow;
288 clone_predicates_to_unswitched_loop(loop, old_new, iffast_pred, ifslow_pred);
289
290 Node* l = head->skip_strip_mined();
291 _igvn.replace_input_of(l, LoopNode::EntryControl, iffast_pred);
292 set_idom(l, iffast_pred, dom_depth(l));
293 LoopNode* slow_l = old_new[head->_idx]->as_Loop()->skip_strip_mined();
294 _igvn.replace_input_of(slow_l, LoopNode::EntryControl, ifslow_pred);
295 set_idom(slow_l, ifslow_pred, dom_depth(l));
296
297 recompute_dom_depth();
298
299 return iffast;
300 }
301
create_reserve_version_of_loop(IdealLoopTree * loop,CountedLoopReserveKit * lk)302 LoopNode* PhaseIdealLoop::create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk) {
303 Node_List old_new;
304 LoopNode* head = loop->_head->as_Loop();
305 bool counted_loop = head->is_CountedLoop();
306 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
307 _igvn.rehash_node_delayed(entry);
308 IdealLoopTree* outer_loop = head->is_strip_mined() ? loop->_parent->_parent : loop->_parent;
309
310 ConINode* const_1 = _igvn.intcon(1);
311 set_ctrl(const_1, C->root());
312 IfNode* iff = new IfNode(entry, const_1, PROB_MAX, COUNT_UNKNOWN);
313 register_node(iff, outer_loop, entry, dom_depth(entry));
314 ProjNode* iffast = new IfTrueNode(iff);
315 register_node(iffast, outer_loop, iff, dom_depth(iff));
316 ProjNode* ifslow = new IfFalseNode(iff);
317 register_node(ifslow, outer_loop, iff, dom_depth(iff));
318
319 // Clone the loop body. The clone becomes the slow loop. The
320 // original pre-header will (illegally) have 3 control users
321 // (old & new loops & new if).
322 clone_loop(loop, old_new, dom_depth(head), CloneIncludesStripMined, iff);
323 assert(old_new[head->_idx]->is_Loop(), "" );
324
325 LoopNode* slow_head = old_new[head->_idx]->as_Loop();
326
327 #ifndef PRODUCT
328 if (TraceLoopOpts) {
329 tty->print_cr("PhaseIdealLoop::create_reserve_version_of_loop:");
330 tty->print("\t iff = %d, ", iff->_idx); iff->dump();
331 tty->print("\t iffast = %d, ", iffast->_idx); iffast->dump();
332 tty->print("\t ifslow = %d, ", ifslow->_idx); ifslow->dump();
333 tty->print("\t before replace_input_of: head = %d, ", head->_idx); head->dump();
334 tty->print("\t before replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
335 }
336 #endif
337
338 // Fast (true) control
339 _igvn.replace_input_of(head->skip_strip_mined(), LoopNode::EntryControl, iffast);
340 // Slow (false) control
341 _igvn.replace_input_of(slow_head->skip_strip_mined(), LoopNode::EntryControl, ifslow);
342
343 recompute_dom_depth();
344
345 lk->set_iff(iff);
346
347 #ifndef PRODUCT
348 if (TraceLoopOpts ) {
349 tty->print("\t after replace_input_of: head = %d, ", head->_idx); head->dump();
350 tty->print("\t after replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
351 }
352 #endif
353
354 return slow_head->as_Loop();
355 }
356
CountedLoopReserveKit(PhaseIdealLoop * phase,IdealLoopTree * loop,bool active=true)357 CountedLoopReserveKit::CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active = true) :
358 _phase(phase),
359 _lpt(loop),
360 _lp(NULL),
361 _iff(NULL),
362 _lp_reserved(NULL),
363 _has_reserved(false),
364 _use_new(false),
365 _active(active)
366 {
367 create_reserve();
368 };
369
~CountedLoopReserveKit()370 CountedLoopReserveKit::~CountedLoopReserveKit() {
371 if (!_active) {
372 return;
373 }
374
375 if (_has_reserved && !_use_new) {
376 // intcon(0)->iff-node reverts CF to the reserved copy
377 ConINode* const_0 = _phase->_igvn.intcon(0);
378 _phase->set_ctrl(const_0, _phase->C->root());
379 _iff->set_req(1, const_0);
380
381 #ifndef PRODUCT
382 if (TraceLoopOpts) {
383 tty->print_cr("CountedLoopReserveKit::~CountedLoopReserveKit()");
384 tty->print("\t discard loop %d and revert to the reserved loop clone %d: ", _lp->_idx, _lp_reserved->_idx);
385 _lp_reserved->dump();
386 }
387 #endif
388 }
389 }
390
create_reserve()391 bool CountedLoopReserveKit::create_reserve() {
392 if (!_active) {
393 return false;
394 }
395
396 if(!_lpt->_head->is_CountedLoop()) {
397 if (TraceLoopOpts) {
398 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not counted loop", _lpt->_head->_idx);
399 }
400 return false;
401 }
402 CountedLoopNode *cl = _lpt->_head->as_CountedLoop();
403 if (!cl->is_valid_counted_loop(T_INT)) {
404 if (TraceLoopOpts) {
405 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not valid counted loop", cl->_idx);
406 }
407 return false; // skip malformed counted loop
408 }
409 if (!cl->is_main_loop()) {
410 bool loop_not_canonical = true;
411 if (cl->is_post_loop() && (cl->slp_max_unroll() > 0)) {
412 loop_not_canonical = false;
413 }
414 // only reject some loop forms
415 if (loop_not_canonical) {
416 if (TraceLoopOpts) {
417 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not canonical loop", cl->_idx);
418 }
419 return false; // skip normal, pre, and post (conditionally) loops
420 }
421 }
422
423 _lp = _lpt->_head->as_Loop();
424 _lp_reserved = _phase->create_reserve_version_of_loop(_lpt, this);
425
426 if (!_lp_reserved->is_CountedLoop()) {
427 return false;
428 }
429
430 Node* ifslow_pred = _lp_reserved->skip_strip_mined()->in(LoopNode::EntryControl);
431
432 if (!ifslow_pred->is_IfFalse()) {
433 return false;
434 }
435
436 Node* iff = ifslow_pred->in(0);
437 if (!iff->is_If() || iff != _iff) {
438 return false;
439 }
440
441 if (iff->in(1)->Opcode() != Op_ConI) {
442 return false;
443 }
444
445 _has_reserved = true;
446 return true;
447 }
448