1 /*
2 * Copyright (c) 2006, 2012, 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 #include "opto/shenandoahSupport.hpp"
33
34 //================= Loop Unswitching =====================
35 //
36 // orig: transformed:
37 // if (invariant-test) then
38 // predicate predicate
39 // loop loop
40 // stmt1 stmt1
41 // if (invariant-test) then stmt2
42 // stmt2 stmt4
43 // else endloop
44 // stmt3 else
45 // endif predicate [clone]
46 // stmt4 loop [clone]
47 // endloop stmt1 [clone]
48 // stmt3
49 // stmt4 [clone]
50 // endloop
51 // endif
52 //
53 // Note: the "else" clause may be empty
54
55 //------------------------------policy_unswitching-----------------------------
56 // Return TRUE or FALSE if the loop should be unswitched
57 // (ie. clone loop with an invariant test that does not exit the loop)
58 bool IdealLoopTree::policy_unswitching(PhaseIdealLoop *phase) const {
59 if( !LoopUnswitching ) {
60 return false;
61 }
62 if (!_head->is_Loop()) {
63 return false;
64 }
65
66 // check for vectorized loops, any unswitching was already applied
67 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) {
68 return false;
69 }
70
71 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes();
72 if ((int)(2 * _body.size()) > nodes_left) {
73 return false; // Too speculative if running low on nodes.
74 }
75 LoopNode* head = _head->as_Loop();
76 if (head->unswitch_count() + 1 > head->unswitch_max()) {
77 return false;
78 }
79 return phase->find_unswitching_candidate(this) != NULL;
80 }
81
82 //------------------------------find_unswitching_candidate-----------------------------
83 // Find candidate "if" for unswitching
84 IfNode* PhaseIdealLoop::find_unswitching_candidate(const IdealLoopTree *loop) const {
85
86 // Find first invariant test that doesn't exit the loop
87 LoopNode *head = loop->_head->as_Loop();
88 IfNode* unswitch_iff = NULL;
89 Node* n = head->in(LoopNode::LoopBackControl);
90 int loop_has_sfpts = -1;
91 while (n != head) {
92 Node* n_dom = idom(n);
93 if (n->is_Region()) {
94 if (n_dom->is_If()) {
95 IfNode* iff = n_dom->as_If();
96 if (iff->in(1)->is_Bool()) {
97 BoolNode* bol = iff->in(1)->as_Bool();
98 if (bol->in(1)->is_Cmp()) {
99 // If condition is invariant and not a loop exit,
100 // then found reason to unswitch.
101 if (loop->is_invariant(bol) && !loop->is_loop_exit(iff)) {
102 unswitch_iff = iff;
103 } else if (ShenandoahWriteBarrierNode::is_evacuation_in_progress_test(iff) &&
104 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
105 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
106 if (loop_has_sfpts == -1) {
107 for(uint i = 0; i < loop->_body.size(); i++) {
108 Node *m = loop->_body[i];
109 if (m->is_SafePoint() && !m->is_CallLeaf()) {
110 loop_has_sfpts = 1;
111 break;
112 }
113 }
114 if (loop_has_sfpts == -1) {
115 loop_has_sfpts = 0;
116 }
117 }
118 if (!loop_has_sfpts) {
119 unswitch_iff = iff;
120 }
121 }
122 }
123 }
124 }
125 }
126 n = n_dom;
127 }
128 return unswitch_iff;
129 }
130
131 //------------------------------do_unswitching-----------------------------
132 // Clone loop with an invariant test (that does not exit) and
133 // insert a clone of the test that selects which version to
134 // execute.
135 void PhaseIdealLoop::do_unswitching (IdealLoopTree *loop, Node_List &old_new) {
136
137 // Find first invariant test that doesn't exit the loop
138 LoopNode *head = loop->_head->as_Loop();
139
140 IfNode* unswitch_iff = find_unswitching_candidate((const IdealLoopTree *)loop);
141 if (ShenandoahWriteBarrierNode::is_evacuation_in_progress_test(unswitch_iff)) {
142 ShenandoahWriteBarrierNode::move_evacuation_test_out_of_loop(unswitch_iff, this);
143 }
144 assert(unswitch_iff != NULL, "should be at least one");
145
146 #ifndef PRODUCT
147 if (TraceLoopOpts) {
148 tty->print("Unswitch %d ", head->unswitch_count()+1);
149 loop->dump_head();
150 }
151 #endif
152
153 // Need to revert back to normal loop
154 if (head->is_CountedLoop() && !head->as_CountedLoop()->is_normal_loop()) {
155 head->as_CountedLoop()->set_normal_loop();
156 }
157
158 ProjNode* proj_true = create_slow_version_of_loop(loop, old_new, unswitch_iff->Opcode(), CloneIncludesStripMined);
159
160 #ifdef ASSERT
161 Node* uniqc = proj_true->unique_ctrl_out();
162 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
163 Node* predicate = find_predicate(entry);
164 if (predicate != NULL && UseLoopPredicate) {
165 // We may have two predicates, find first.
166 entry = find_predicate(entry->in(0)->in(0));
167 if (entry != NULL) predicate = entry;
168 }
169 if (predicate != NULL) predicate = predicate->in(0);
170 assert(proj_true->is_IfTrue() &&
171 (predicate == NULL && uniqc == head && !head->is_strip_mined() ||
172 predicate == NULL && uniqc == head->in(LoopNode::EntryControl) && head->is_strip_mined() ||
173 predicate != NULL && uniqc == predicate), "by construction");
174 #endif
175 // Increment unswitch count
176 LoopNode* head_clone = old_new[head->_idx]->as_Loop();
177 int nct = head->unswitch_count() + 1;
178 head->set_unswitch_count(nct);
179 head_clone->set_unswitch_count(nct);
180
181 // Add test to new "if" outside of loop
182 IfNode* invar_iff = proj_true->in(0)->as_If();
183 Node* invar_iff_c = invar_iff->in(0);
184 BoolNode* bol = unswitch_iff->in(1)->as_Bool();
185 invar_iff->set_req(1, bol);
186 invar_iff->_prob = unswitch_iff->_prob;
187
188 ProjNode* proj_false = invar_iff->proj_out(0)->as_Proj();
189
190 // Hoist invariant casts out of each loop to the appropriate
191 // control projection.
192
193 Node_List worklist;
194
195 for (DUIterator_Fast imax, i = unswitch_iff->fast_outs(imax); i < imax; i++) {
196 ProjNode* proj= unswitch_iff->fast_out(i)->as_Proj();
197 // Copy to a worklist for easier manipulation
198 for (DUIterator_Fast jmax, j = proj->fast_outs(jmax); j < jmax; j++) {
199 Node* use = proj->fast_out(j);
200 if (use->Opcode() == Op_CheckCastPP && loop->is_invariant(use->in(1))) {
201 worklist.push(use);
202 }
203 }
204 ProjNode* invar_proj = invar_iff->proj_out(proj->_con)->as_Proj();
205 while (worklist.size() > 0) {
206 Node* use = worklist.pop();
207 Node* nuse = use->clone();
208 nuse->set_req(0, invar_proj);
209 _igvn.replace_input_of(use, 1, nuse);
210 register_new_node(nuse, invar_proj);
211 // Same for the clone
212 Node* use_clone = old_new[use->_idx];
213 _igvn.replace_input_of(use_clone, 1, nuse);
214 }
215 }
216
217 // Hardwire the control paths in the loops into if(true) and if(false)
218 _igvn.rehash_node_delayed(unswitch_iff);
219 short_circuit_if(unswitch_iff, proj_true);
220
221 IfNode* unswitch_iff_clone = old_new[unswitch_iff->_idx]->as_If();
222 _igvn.rehash_node_delayed(unswitch_iff_clone);
223 short_circuit_if(unswitch_iff_clone, proj_false);
224
225 // Reoptimize loops
226 loop->record_for_igvn();
227 for(int i = loop->_body.size() - 1; i >= 0 ; i--) {
228 Node *n = loop->_body[i];
229 Node *n_clone = old_new[n->_idx];
230 _igvn._worklist.push(n_clone);
231 }
232
233 #ifndef PRODUCT
234 if (TraceLoopUnswitching) {
235 tty->print_cr("Loop unswitching orig: %d @ %d new: %d @ %d",
236 head->_idx, unswitch_iff->_idx,
237 old_new[head->_idx]->_idx, unswitch_iff_clone->_idx);
238 }
239 #endif
240
241 C->set_major_progress();
242 }
243
244 //-------------------------create_slow_version_of_loop------------------------
245 // Create a slow version of the loop by cloning the loop
246 // and inserting an if to select fast-slow versions.
247 // Return control projection of the entry to the fast version.
248 ProjNode* PhaseIdealLoop::create_slow_version_of_loop(IdealLoopTree *loop,
249 Node_List &old_new,
250 int opcode,
251 CloneLoopMode mode) {
252 LoopNode* head = loop->_head->as_Loop();
253 bool counted_loop = head->is_CountedLoop();
254 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
255 _igvn.rehash_node_delayed(entry);
256 IdealLoopTree* outer_loop = loop->_parent;
257
258 head->verify_strip_mined(1);
259
260 Node *cont = _igvn.intcon(1);
261 set_ctrl(cont, C->root());
262 Node* opq = new Opaque1Node(C, cont);
263 register_node(opq, outer_loop, entry, dom_depth(entry));
264 Node *bol = new Conv2BNode(opq);
265 register_node(bol, outer_loop, entry, dom_depth(entry));
266 IfNode* iff = (opcode == Op_RangeCheck) ? new RangeCheckNode(entry, bol, PROB_MAX, COUNT_UNKNOWN) :
267 new IfNode(entry, bol, PROB_MAX, COUNT_UNKNOWN);
268 register_node(iff, outer_loop, entry, dom_depth(entry));
269 ProjNode* iffast = new IfTrueNode(iff);
270 register_node(iffast, outer_loop, iff, dom_depth(iff));
271 ProjNode* ifslow = new IfFalseNode(iff);
272 register_node(ifslow, outer_loop, iff, dom_depth(iff));
273
274 // Clone the loop body. The clone becomes the fast loop. The
275 // original pre-header will (illegally) have 3 control users
276 // (old & new loops & new if).
277 clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), mode, iff);
278 assert(old_new[head->_idx]->is_Loop(), "" );
279
280 // Fast (true) control
281 Node* iffast_pred = clone_loop_predicates(entry, iffast, !counted_loop);
282
283 // Slow (false) control
284 Node* ifslow_pred = clone_loop_predicates(entry, ifslow, !counted_loop);
285
286 Node* l = head->skip_strip_mined();
287 _igvn.replace_input_of(l, LoopNode::EntryControl, iffast_pred);
288 set_idom(l, iffast_pred, dom_depth(l));
289 LoopNode* slow_l = old_new[head->_idx]->as_Loop()->skip_strip_mined();
290 _igvn.replace_input_of(slow_l, LoopNode::EntryControl, ifslow_pred);
291 set_idom(slow_l, ifslow_pred, dom_depth(l));
292
293 recompute_dom_depth();
294
295 return iffast;
296 }
297
298 LoopNode* PhaseIdealLoop::create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk) {
299 Node_List old_new;
300 LoopNode* head = loop->_head->as_Loop();
301 bool counted_loop = head->is_CountedLoop();
302 Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
303 _igvn.rehash_node_delayed(entry);
304 IdealLoopTree* outer_loop = head->is_strip_mined() ? loop->_parent->_parent : loop->_parent;
305
306 ConINode* const_1 = _igvn.intcon(1);
307 set_ctrl(const_1, C->root());
308 IfNode* iff = new IfNode(entry, const_1, PROB_MAX, COUNT_UNKNOWN);
309 register_node(iff, outer_loop, entry, dom_depth(entry));
310 ProjNode* iffast = new IfTrueNode(iff);
311 register_node(iffast, outer_loop, iff, dom_depth(iff));
312 ProjNode* ifslow = new IfFalseNode(iff);
313 register_node(ifslow, outer_loop, iff, dom_depth(iff));
314
315 // Clone the loop body. The clone becomes the fast loop. The
316 // original pre-header will (illegally) have 3 control users
317 // (old & new loops & new if).
318 clone_loop(loop, old_new, dom_depth(head), CloneIncludesStripMined, iff);
319 assert(old_new[head->_idx]->is_Loop(), "" );
320
321 LoopNode* slow_head = old_new[head->_idx]->as_Loop();
322
323 #ifndef PRODUCT
324 if (TraceLoopOpts) {
325 tty->print_cr("PhaseIdealLoop::create_reserve_version_of_loop:");
326 tty->print("\t iff = %d, ", iff->_idx); iff->dump();
327 tty->print("\t iffast = %d, ", iffast->_idx); iffast->dump();
328 tty->print("\t ifslow = %d, ", ifslow->_idx); ifslow->dump();
329 tty->print("\t before replace_input_of: head = %d, ", head->_idx); head->dump();
330 tty->print("\t before replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
331 }
332 #endif
333
334 // Fast (true) control
335 _igvn.replace_input_of(head->skip_strip_mined(), LoopNode::EntryControl, iffast);
336 // Slow (false) control
337 _igvn.replace_input_of(slow_head->skip_strip_mined(), LoopNode::EntryControl, ifslow);
338
339 recompute_dom_depth();
340
341 lk->set_iff(iff);
342
343 #ifndef PRODUCT
344 if (TraceLoopOpts ) {
345 tty->print("\t after replace_input_of: head = %d, ", head->_idx); head->dump();
346 tty->print("\t after replace_input_of: slow_head = %d, ", slow_head->_idx); slow_head->dump();
347 }
348 #endif
349
350 return slow_head->as_Loop();
351 }
352
353 CountedLoopReserveKit::CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active = true) :
354 _phase(phase),
355 _lpt(loop),
356 _lp(NULL),
357 _iff(NULL),
358 _lp_reserved(NULL),
359 _has_reserved(false),
360 _use_new(false),
361 _active(active)
362 {
363 create_reserve();
364 };
365
366 CountedLoopReserveKit::~CountedLoopReserveKit() {
367 if (!_active) {
368 return;
369 }
370
371 if (_has_reserved && !_use_new) {
372 // intcon(0)->iff-node reverts CF to the reserved copy
373 ConINode* const_0 = _phase->_igvn.intcon(0);
374 _phase->set_ctrl(const_0, _phase->C->root());
375 _iff->set_req(1, const_0);
376
377 #ifndef PRODUCT
378 if (TraceLoopOpts) {
379 tty->print_cr("CountedLoopReserveKit::~CountedLoopReserveKit()");
380 tty->print("\t discard loop %d and revert to the reserved loop clone %d: ", _lp->_idx, _lp_reserved->_idx);
381 _lp_reserved->dump();
382 }
383 #endif
384 }
385 }
386
387 bool CountedLoopReserveKit::create_reserve() {
388 if (!_active) {
389 return false;
390 }
391
392 if(!_lpt->_head->is_CountedLoop()) {
393 if (TraceLoopOpts) {
394 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not counted loop", _lpt->_head->_idx);
395 }
396 return false;
397 }
398 CountedLoopNode *cl = _lpt->_head->as_CountedLoop();
399 if (!cl->is_valid_counted_loop()) {
400 if (TraceLoopOpts) {
401 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not valid counted loop", cl->_idx);
402 }
403 return false; // skip malformed counted loop
404 }
405 if (!cl->is_main_loop()) {
406 bool loop_not_canonical = true;
407 if (cl->is_post_loop() && (cl->slp_max_unroll() > 0)) {
408 loop_not_canonical = false;
409 }
410 // only reject some loop forms
411 if (loop_not_canonical) {
412 if (TraceLoopOpts) {
413 tty->print_cr("CountedLoopReserveKit::create_reserve: %d not canonical loop", cl->_idx);
414 }
415 return false; // skip normal, pre, and post (conditionally) loops
416 }
417 }
418
419 _lp = _lpt->_head->as_Loop();
420 _lp_reserved = _phase->create_reserve_version_of_loop(_lpt, this);
421
422 if (!_lp_reserved->is_CountedLoop()) {
423 return false;
424 }
425
426 Node* ifslow_pred = _lp_reserved->skip_strip_mined()->in(LoopNode::EntryControl);
427
428 if (!ifslow_pred->is_IfFalse()) {
429 return false;
430 }
431
432 Node* iff = ifslow_pred->in(0);
433 if (!iff->is_If() || iff != _iff) {
434 return false;
435 }
436
437 if (iff->in(1)->Opcode() != Op_ConI) {
438 return false;
439 }
440
441 return _has_reserved = true;
442 }