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