1 /*
2 * Copyright (c) 1998, 2018, 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 "ci/ciMethodData.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/divnode.hpp"
38 #include "opto/idealGraphPrinter.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/mulnode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/superword.hpp"
43 #include "utilities/macros.hpp"
44 #if INCLUDE_SHENANDOAHGC
45 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
46 #endif
47
48 //=============================================================================
49 //------------------------------is_loop_iv-------------------------------------
50 // Determine if a node is Counted loop induction variable.
51 // The method is declared in node.hpp.
52 const Node* Node::is_loop_iv() const {
53 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
54 this->as_Phi()->region()->is_CountedLoop() &&
55 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
56 return this;
57 } else {
58 return NULL;
59 }
60 }
61
62 //=============================================================================
63 //------------------------------dump_spec--------------------------------------
64 // Dump special per-node info
65 #ifndef PRODUCT
66 void LoopNode::dump_spec(outputStream *st) const {
67 if (is_inner_loop()) st->print( "inner " );
68 if (is_partial_peel_loop()) st->print( "partial_peel " );
69 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
70 }
71 #endif
72
73 //------------------------------is_valid_counted_loop-------------------------
74 bool LoopNode::is_valid_counted_loop() const {
75 if (is_CountedLoop()) {
76 CountedLoopNode* l = as_CountedLoop();
77 CountedLoopEndNode* le = l->loopexit_or_null();
78 if (le != NULL &&
79 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
80 Node* phi = l->phi();
81 Node* exit = le->proj_out_or_null(0 /* false */);
82 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
83 phi != NULL && phi->is_Phi() &&
84 phi->in(LoopNode::LoopBackControl) == l->incr() &&
85 le->loopnode() == l && le->stride_is_con()) {
86 return true;
87 }
88 }
89 }
90 return false;
91 }
92
93 //------------------------------get_early_ctrl---------------------------------
94 // Compute earliest legal control
95 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
96 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
97 uint i;
98 Node *early;
99 if (n->in(0) && !n->is_expensive()) {
100 early = n->in(0);
101 if (!early->is_CFG()) // Might be a non-CFG multi-def
102 early = get_ctrl(early); // So treat input as a straight data input
103 i = 1;
104 } else {
105 early = get_ctrl(n->in(1));
106 i = 2;
107 }
108 uint e_d = dom_depth(early);
109 assert( early, "" );
110 for (; i < n->req(); i++) {
111 Node *cin = get_ctrl(n->in(i));
112 assert( cin, "" );
113 // Keep deepest dominator depth
114 uint c_d = dom_depth(cin);
115 if (c_d > e_d) { // Deeper guy?
116 early = cin; // Keep deepest found so far
117 e_d = c_d;
118 } else if (c_d == e_d && // Same depth?
119 early != cin) { // If not equal, must use slower algorithm
120 // If same depth but not equal, one _must_ dominate the other
121 // and we want the deeper (i.e., dominated) guy.
122 Node *n1 = early;
123 Node *n2 = cin;
124 while (1) {
125 n1 = idom(n1); // Walk up until break cycle
126 n2 = idom(n2);
127 if (n1 == cin || // Walked early up to cin
128 dom_depth(n2) < c_d)
129 break; // early is deeper; keep him
130 if (n2 == early || // Walked cin up to early
131 dom_depth(n1) < c_d) {
132 early = cin; // cin is deeper; keep him
133 break;
134 }
135 }
136 e_d = dom_depth(early); // Reset depth register cache
137 }
138 }
139
140 // Return earliest legal location
141 assert(early == find_non_split_ctrl(early), "unexpected early control");
142
143 if (n->is_expensive() && !_verify_only && !_verify_me) {
144 assert(n->in(0), "should have control input");
145 early = get_early_ctrl_for_expensive(n, early);
146 }
147
148 return early;
149 }
150
151 //------------------------------get_early_ctrl_for_expensive---------------------------------
152 // Move node up the dominator tree as high as legal while still beneficial
153 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
154 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
155 assert(OptimizeExpensiveOps, "optimization off?");
156
157 Node* ctl = n->in(0);
158 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
159 uint min_dom_depth = dom_depth(earliest);
160 #ifdef ASSERT
161 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
162 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
163 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
164 }
165 #endif
166 if (dom_depth(ctl) < min_dom_depth) {
167 return earliest;
168 }
169
170 while (1) {
171 Node *next = ctl;
172 // Moving the node out of a loop on the projection of a If
173 // confuses loop predication. So once we hit a Loop in a If branch
174 // that doesn't branch to an UNC, we stop. The code that process
175 // expensive nodes will notice the loop and skip over it to try to
176 // move the node further up.
177 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
178 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
179 break;
180 }
181 next = idom(ctl->in(1)->in(0));
182 } else if (ctl->is_Proj()) {
183 // We only move it up along a projection if the projection is
184 // the single control projection for its parent: same code path,
185 // if it's a If with UNC or fallthrough of a call.
186 Node* parent_ctl = ctl->in(0);
187 if (parent_ctl == NULL) {
188 break;
189 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
190 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
191 } else if (parent_ctl->is_If()) {
192 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
193 break;
194 }
195 assert(idom(ctl) == parent_ctl, "strange");
196 next = idom(parent_ctl);
197 } else if (ctl->is_CatchProj()) {
198 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
199 break;
200 }
201 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
202 next = parent_ctl->in(0)->in(0)->in(0);
203 } else {
204 // Check if parent control has a single projection (this
205 // control is the only possible successor of the parent
206 // control). If so, we can try to move the node above the
207 // parent control.
208 int nb_ctl_proj = 0;
209 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
210 Node *p = parent_ctl->fast_out(i);
211 if (p->is_Proj() && p->is_CFG()) {
212 nb_ctl_proj++;
213 if (nb_ctl_proj > 1) {
214 break;
215 }
216 }
217 }
218
219 if (nb_ctl_proj > 1) {
220 break;
221 }
222 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
223 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
224 assert(idom(ctl) == parent_ctl, "strange");
225 next = idom(parent_ctl);
226 }
227 } else {
228 next = idom(ctl);
229 }
230 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
231 break;
232 }
233 ctl = next;
234 }
235
236 if (ctl != n->in(0)) {
237 _igvn.replace_input_of(n, 0, ctl);
238 _igvn.hash_insert(n);
239 }
240
241 return ctl;
242 }
243
244
245 //------------------------------set_early_ctrl---------------------------------
246 // Set earliest legal control
247 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
248 Node *early = get_early_ctrl(n);
249
250 // Record earliest legal location
251 set_ctrl(n, early);
252 }
253
254 //------------------------------set_subtree_ctrl-------------------------------
255 // set missing _ctrl entries on new nodes
256 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
257 // Already set? Get out.
258 if( _nodes[n->_idx] ) return;
259 // Recursively set _nodes array to indicate where the Node goes
260 uint i;
261 for( i = 0; i < n->req(); ++i ) {
262 Node *m = n->in(i);
263 if( m && m != C->root() )
264 set_subtree_ctrl( m );
265 }
266
267 // Fixup self
268 set_early_ctrl( n );
269 }
270
271 // Create a skeleton strip mined outer loop: a Loop head before the
272 // inner strip mined loop, a safepoint and an exit condition guarded
273 // by an opaque node after the inner strip mined loop with a backedge
274 // to the loop head. The inner strip mined loop is left as it is. Only
275 // once loop optimizations are over, do we adjust the inner loop exit
276 // condition to limit its number of iterations, set the outer loop
277 // exit condition and add Phis to the outer loop head. Some loop
278 // optimizations that operate on the inner strip mined loop need to be
279 // aware of the outer strip mined loop: loop unswitching needs to
280 // clone the outer loop as well as the inner, unrolling needs to only
281 // clone the inner loop etc. No optimizations need to change the outer
282 // strip mined loop as it is only a skeleton.
283 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
284 IdealLoopTree* loop, float cl_prob, float le_fcnt,
285 Node*& entry_control, Node*& iffalse) {
286 Node* outer_test = _igvn.intcon(0);
287 set_ctrl(outer_test, C->root());
288 Node *orig = iffalse;
289 iffalse = iffalse->clone();
290 _igvn.register_new_node_with_optimizer(iffalse);
291 set_idom(iffalse, idom(orig), dom_depth(orig));
292
293 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
294 Node *outer_ift = new IfTrueNode (outer_le);
295 Node* outer_iff = orig;
296 _igvn.replace_input_of(outer_iff, 0, outer_le);
297
298 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
299 entry_control = outer_l;
300
301 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
302 IdealLoopTree* parent = loop->_parent;
303 IdealLoopTree* sibling = parent->_child;
304 if (sibling == loop) {
305 parent->_child = outer_ilt;
306 } else {
307 while (sibling->_next != loop) {
308 sibling = sibling->_next;
309 }
310 sibling->_next = outer_ilt;
311 }
312 outer_ilt->_next = loop->_next;
313 outer_ilt->_parent = parent;
314 outer_ilt->_child = loop;
315 outer_ilt->_nest = loop->_nest;
316 loop->_parent = outer_ilt;
317 loop->_next = NULL;
318 loop->_nest++;
319
320 set_loop(iffalse, outer_ilt);
321 register_control(outer_le, outer_ilt, iffalse);
322 register_control(outer_ift, outer_ilt, outer_le);
323 set_idom(outer_iff, outer_le, dom_depth(outer_le));
324 _igvn.register_new_node_with_optimizer(outer_l);
325 set_loop(outer_l, outer_ilt);
326 set_idom(outer_l, init_control, dom_depth(init_control)+1);
327
328 return outer_ilt;
329 }
330
331 void PhaseIdealLoop::insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol) {
332 Node* new_predicate_proj = create_new_if_for_predicate(limit_check_proj, NULL,
333 Deoptimization::Reason_loop_limit_check,
334 Op_If);
335 Node* iff = new_predicate_proj->in(0);
336 assert(iff->Opcode() == Op_If, "bad graph shape");
337 Node* conv = iff->in(1);
338 assert(conv->Opcode() == Op_Conv2B, "bad graph shape");
339 Node* opaq = conv->in(1);
340 assert(opaq->Opcode() == Op_Opaque1, "bad graph shape");
341 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
342 bol = _igvn.register_new_node_with_optimizer(bol);
343 set_subtree_ctrl(bol);
344 _igvn.replace_input_of(iff, 1, bol);
345
346 #ifndef PRODUCT
347 // report that the loop predication has been actually performed
348 // for this loop
349 if (TraceLoopLimitCheck) {
350 tty->print_cr("Counted Loop Limit Check generated:");
351 debug_only( bol->dump(2); )
352 }
353 #endif
354 }
355
356 //------------------------------is_counted_loop--------------------------------
357 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) {
358 PhaseGVN *gvn = &_igvn;
359
360 // Counted loop head must be a good RegionNode with only 3 not NULL
361 // control input edges: Self, Entry, LoopBack.
362 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
363 return false;
364 }
365 Node *init_control = x->in(LoopNode::EntryControl);
366 Node *back_control = x->in(LoopNode::LoopBackControl);
367 if (init_control == NULL || back_control == NULL) // Partially dead
368 return false;
369 // Must also check for TOP when looking for a dead loop
370 if (init_control->is_top() || back_control->is_top())
371 return false;
372
373 // Allow funny placement of Safepoint
374 if (back_control->Opcode() == Op_SafePoint) {
375 if (LoopStripMiningIter != 0) {
376 // Leaving the safepoint on the backedge and creating a
377 // CountedLoop will confuse optimizations. We can't move the
378 // safepoint around because its jvm state wouldn't match a new
379 // location. Give up on that loop.
380 return false;
381 }
382 back_control = back_control->in(TypeFunc::Control);
383 }
384
385 // Controlling test for loop
386 Node *iftrue = back_control;
387 uint iftrue_op = iftrue->Opcode();
388 if (iftrue_op != Op_IfTrue &&
389 iftrue_op != Op_IfFalse)
390 // I have a weird back-control. Probably the loop-exit test is in
391 // the middle of the loop and I am looking at some trailing control-flow
392 // merge point. To fix this I would have to partially peel the loop.
393 return false; // Obscure back-control
394
395 // Get boolean guarding loop-back test
396 Node *iff = iftrue->in(0);
397 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
398 return false;
399 BoolNode *test = iff->in(1)->as_Bool();
400 BoolTest::mask bt = test->_test._test;
401 float cl_prob = iff->as_If()->_prob;
402 if (iftrue_op == Op_IfFalse) {
403 bt = BoolTest(bt).negate();
404 cl_prob = 1.0 - cl_prob;
405 }
406 // Get backedge compare
407 Node *cmp = test->in(1);
408 int cmp_op = cmp->Opcode();
409 if (cmp_op != Op_CmpI)
410 return false; // Avoid pointer & float compares
411
412 // Find the trip-counter increment & limit. Limit must be loop invariant.
413 Node *incr = cmp->in(1);
414 Node *limit = cmp->in(2);
415
416 // ---------
417 // need 'loop()' test to tell if limit is loop invariant
418 // ---------
419
420 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
421 Node *tmp = incr; // Then reverse order into the CmpI
422 incr = limit;
423 limit = tmp;
424 bt = BoolTest(bt).commute(); // And commute the exit test
425 }
426 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
427 return false;
428 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
429 return false;
430
431 Node* phi_incr = NULL;
432 // Trip-counter increment must be commutative & associative.
433 if (incr->Opcode() == Op_CastII) {
434 incr = incr->in(1);
435 }
436 if (incr->is_Phi()) {
437 if (incr->as_Phi()->region() != x || incr->req() != 3)
438 return false; // Not simple trip counter expression
439 phi_incr = incr;
440 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
441 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
442 return false;
443 }
444
445 Node* trunc1 = NULL;
446 Node* trunc2 = NULL;
447 const TypeInt* iv_trunc_t = NULL;
448 Node* orig_incr = incr;
449 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
450 return false; // Funny increment opcode
451 }
452 assert(incr->Opcode() == Op_AddI, "wrong increment code");
453
454 const TypeInt* limit_t = gvn->type(limit)->is_int();
455 if (trunc1 != NULL) {
456 // When there is a truncation, we must be sure that after the truncation
457 // the trip counter will end up higher than the limit, otherwise we are looking
458 // at an endless loop. Can happen with range checks.
459
460 // Example:
461 // int i = 0;
462 // while (true)
463 // sum + = array[i];
464 // i++;
465 // i = i && 0x7fff;
466 // }
467 //
468 // If the array is shorter than 0x8000 this exits through a AIOOB
469 // - Counted loop transformation is ok
470 // If the array is longer then this is an endless loop
471 // - No transformation can be done.
472
473 const TypeInt* incr_t = gvn->type(orig_incr)->is_int();
474 if (limit_t->_hi > incr_t->_hi) {
475 // if the limit can have a higher value than the increment (before the phi)
476 return false;
477 }
478 }
479
480 // Get merge point
481 Node *xphi = incr->in(1);
482 Node *stride = incr->in(2);
483 if (!stride->is_Con()) { // Oops, swap these
484 if (!xphi->is_Con()) // Is the other guy a constant?
485 return false; // Nope, unknown stride, bail out
486 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
487 xphi = stride;
488 stride = tmp;
489 }
490 if (xphi->Opcode() == Op_CastII) {
491 xphi = xphi->in(1);
492 }
493 // Stride must be constant
494 int stride_con = stride->get_int();
495 if (stride_con == 0)
496 return false; // missed some peephole opt
497
498 if (!xphi->is_Phi())
499 return false; // Too much math on the trip counter
500 if (phi_incr != NULL && phi_incr != xphi)
501 return false;
502 PhiNode *phi = xphi->as_Phi();
503
504 // Phi must be of loop header; backedge must wrap to increment
505 if (phi->region() != x)
506 return false;
507 if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) ||
508 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) {
509 return false;
510 }
511 Node *init_trip = phi->in(LoopNode::EntryControl);
512
513 // If iv trunc type is smaller than int, check for possible wrap.
514 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
515 assert(trunc1 != NULL, "must have found some truncation");
516
517 // Get a better type for the phi (filtered thru if's)
518 const TypeInt* phi_ft = filtered_type(phi);
519
520 // Can iv take on a value that will wrap?
521 //
522 // Ensure iv's limit is not within "stride" of the wrap value.
523 //
524 // Example for "short" type
525 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
526 // If the stride is +10, then the last value of the induction
527 // variable before the increment (phi_ft->_hi) must be
528 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
529 // ensure no truncation occurs after the increment.
530
531 if (stride_con > 0) {
532 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
533 iv_trunc_t->_lo > phi_ft->_lo) {
534 return false; // truncation may occur
535 }
536 } else if (stride_con < 0) {
537 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
538 iv_trunc_t->_hi < phi_ft->_hi) {
539 return false; // truncation may occur
540 }
541 }
542 // No possibility of wrap so truncation can be discarded
543 // Promote iv type to Int
544 } else {
545 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
546 }
547
548 // If the condition is inverted and we will be rolling
549 // through MININT to MAXINT, then bail out.
550 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
551 // Odd stride
552 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
553 // Count down loop rolls through MAXINT
554 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
555 // Count up loop rolls through MININT
556 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
557 return false; // Bail out
558 }
559
560 const TypeInt* init_t = gvn->type(init_trip)->is_int();
561
562 if (stride_con > 0) {
563 jlong init_p = (jlong)init_t->_lo + stride_con;
564 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
565 return false; // cyclic loop or this loop trips only once
566 } else {
567 jlong init_p = (jlong)init_t->_hi + stride_con;
568 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
569 return false; // cyclic loop or this loop trips only once
570 }
571
572 if (phi_incr != NULL && bt != BoolTest::ne) {
573 // check if there is a possiblity of IV overflowing after the first increment
574 if (stride_con > 0) {
575 if (init_t->_hi > max_jint - stride_con) {
576 return false;
577 }
578 } else {
579 if (init_t->_lo < min_jint - stride_con) {
580 return false;
581 }
582 }
583 }
584
585 // =================================================
586 // ---- SUCCESS! Found A Trip-Counted Loop! -----
587 //
588 assert(x->Opcode() == Op_Loop, "regular loops only");
589 C->print_method(PHASE_BEFORE_CLOOPS, 3);
590
591 Node *hook = new Node(6);
592
593 // ===================================================
594 // Generate loop limit check to avoid integer overflow
595 // in cases like next (cyclic loops):
596 //
597 // for (i=0; i <= max_jint; i++) {}
598 // for (i=0; i < max_jint; i+=2) {}
599 //
600 //
601 // Limit check predicate depends on the loop test:
602 //
603 // for(;i != limit; i++) --> limit <= (max_jint)
604 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
605 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
606 //
607
608 // Check if limit is excluded to do more precise int overflow check.
609 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
610 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
611
612 // If compare points directly to the phi we need to adjust
613 // the compare so that it points to the incr. Limit have
614 // to be adjusted to keep trip count the same and the
615 // adjusted limit should be checked for int overflow.
616 if (phi_incr != NULL) {
617 stride_m += stride_con;
618 }
619
620 if (limit->is_Con()) {
621 int limit_con = limit->get_int();
622 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
623 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
624 // Bailout: it could be integer overflow.
625 return false;
626 }
627 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
628 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
629 // Limit's type may satisfy the condition, for example,
630 // when it is an array length.
631 } else {
632 // Generate loop's limit check.
633 // Loop limit check predicate should be near the loop.
634 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
635 if (!limit_check_proj) {
636 // The limit check predicate is not generated if this method trapped here before.
637 #ifdef ASSERT
638 if (TraceLoopLimitCheck) {
639 tty->print("missing loop limit check:");
640 loop->dump_head();
641 x->dump(1);
642 }
643 #endif
644 return false;
645 }
646
647 IfNode* check_iff = limit_check_proj->in(0)->as_If();
648
649 if (!is_dominator(get_ctrl(limit), check_iff->in(0))) {
650 return false;
651 }
652
653 Node* cmp_limit;
654 Node* bol;
655
656 if (stride_con > 0) {
657 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
658 bol = new BoolNode(cmp_limit, BoolTest::le);
659 } else {
660 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
661 bol = new BoolNode(cmp_limit, BoolTest::ge);
662 }
663
664 insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
665 }
666
667 // Now we need to canonicalize loop condition.
668 if (bt == BoolTest::ne) {
669 assert(stride_con == 1 || stride_con == -1, "simple increment only");
670 if (stride_con > 0 && init_t->_hi < limit_t->_lo) {
671 // 'ne' can be replaced with 'lt' only when init < limit.
672 bt = BoolTest::lt;
673 } else if (stride_con < 0 && init_t->_lo > limit_t->_hi) {
674 // 'ne' can be replaced with 'gt' only when init > limit.
675 bt = BoolTest::gt;
676 } else {
677 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
678 if (!limit_check_proj) {
679 // The limit check predicate is not generated if this method trapped here before.
680 #ifdef ASSERT
681 if (TraceLoopLimitCheck) {
682 tty->print("missing loop limit check:");
683 loop->dump_head();
684 x->dump(1);
685 }
686 #endif
687 return false;
688 }
689 IfNode* check_iff = limit_check_proj->in(0)->as_If();
690
691 if (!is_dominator(get_ctrl(limit), check_iff->in(0)) ||
692 !is_dominator(get_ctrl(init_trip), check_iff->in(0))) {
693 return false;
694 }
695
696 Node* cmp_limit;
697 Node* bol;
698
699 if (stride_con > 0) {
700 cmp_limit = new CmpINode(init_trip, limit);
701 bol = new BoolNode(cmp_limit, BoolTest::lt);
702 } else {
703 cmp_limit = new CmpINode(init_trip, limit);
704 bol = new BoolNode(cmp_limit, BoolTest::gt);
705 }
706
707 insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
708
709 if (stride_con > 0) {
710 // 'ne' can be replaced with 'lt' only when init < limit.
711 bt = BoolTest::lt;
712 } else if (stride_con < 0) {
713 // 'ne' can be replaced with 'gt' only when init > limit.
714 bt = BoolTest::gt;
715 }
716 }
717 }
718
719 if (phi_incr != NULL) {
720 // If compare points directly to the phi we need to adjust
721 // the compare so that it points to the incr. Limit have
722 // to be adjusted to keep trip count the same and we
723 // should avoid int overflow.
724 //
725 // i = init; do {} while(i++ < limit);
726 // is converted to
727 // i = init; do {} while(++i < limit+1);
728 //
729 limit = gvn->transform(new AddINode(limit, stride));
730 }
731
732 if (incl_limit) {
733 // The limit check guaranties that 'limit <= (max_jint - stride)' so
734 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
735 //
736 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
737 limit = gvn->transform(new AddINode(limit, one));
738 if (bt == BoolTest::le)
739 bt = BoolTest::lt;
740 else if (bt == BoolTest::ge)
741 bt = BoolTest::gt;
742 else
743 ShouldNotReachHere();
744 }
745 set_subtree_ctrl( limit );
746
747 if (LoopStripMiningIter == 0) {
748 // Check for SafePoint on backedge and remove
749 Node *sfpt = x->in(LoopNode::LoopBackControl);
750 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
751 lazy_replace( sfpt, iftrue );
752 if (loop->_safepts != NULL) {
753 loop->_safepts->yank(sfpt);
754 }
755 loop->_tail = iftrue;
756 }
757 }
758
759 // Build a canonical trip test.
760 // Clone code, as old values may be in use.
761 incr = incr->clone();
762 incr->set_req(1,phi);
763 incr->set_req(2,stride);
764 incr = _igvn.register_new_node_with_optimizer(incr);
765 set_early_ctrl( incr );
766 _igvn.rehash_node_delayed(phi);
767 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
768
769 // If phi type is more restrictive than Int, raise to
770 // Int to prevent (almost) infinite recursion in igvn
771 // which can only handle integer types for constants or minint..maxint.
772 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
773 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
774 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
775 nphi = _igvn.register_new_node_with_optimizer(nphi);
776 set_ctrl(nphi, get_ctrl(phi));
777 _igvn.replace_node(phi, nphi);
778 phi = nphi->as_Phi();
779 }
780 cmp = cmp->clone();
781 cmp->set_req(1,incr);
782 cmp->set_req(2,limit);
783 cmp = _igvn.register_new_node_with_optimizer(cmp);
784 set_ctrl(cmp, iff->in(0));
785
786 test = test->clone()->as_Bool();
787 (*(BoolTest*)&test->_test)._test = bt;
788 test->set_req(1,cmp);
789 _igvn.register_new_node_with_optimizer(test);
790 set_ctrl(test, iff->in(0));
791
792 // Replace the old IfNode with a new LoopEndNode
793 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
794 IfNode *le = lex->as_If();
795 uint dd = dom_depth(iff);
796 set_idom(le, le->in(0), dd); // Update dominance for loop exit
797 set_loop(le, loop);
798
799 // Get the loop-exit control
800 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
801
802 // Need to swap loop-exit and loop-back control?
803 if (iftrue_op == Op_IfFalse) {
804 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
805 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
806
807 loop->_tail = back_control = ift2;
808 set_loop(ift2, loop);
809 set_loop(iff2, get_loop(iffalse));
810
811 // Lazy update of 'get_ctrl' mechanism.
812 lazy_replace(iffalse, iff2);
813 lazy_replace(iftrue, ift2);
814
815 // Swap names
816 iffalse = iff2;
817 iftrue = ift2;
818 } else {
819 _igvn.rehash_node_delayed(iffalse);
820 _igvn.rehash_node_delayed(iftrue);
821 iffalse->set_req_X( 0, le, &_igvn );
822 iftrue ->set_req_X( 0, le, &_igvn );
823 }
824
825 set_idom(iftrue, le, dd+1);
826 set_idom(iffalse, le, dd+1);
827 assert(iff->outcnt() == 0, "should be dead now");
828 lazy_replace( iff, le ); // fix 'get_ctrl'
829
830 Node *sfpt2 = le->in(0);
831
832 Node* entry_control = init_control;
833 bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL &&
834 sfpt2->Opcode() == Op_SafePoint && !loop->_has_call;
835 IdealLoopTree* outer_ilt = NULL;
836 if (strip_mine_loop) {
837 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
838 cl_prob, le->_fcnt, entry_control,
839 iffalse);
840 }
841
842 // Now setup a new CountedLoopNode to replace the existing LoopNode
843 CountedLoopNode *l = new CountedLoopNode(entry_control, back_control);
844 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
845 // The following assert is approximately true, and defines the intention
846 // of can_be_counted_loop. It fails, however, because phase->type
847 // is not yet initialized for this loop and its parts.
848 //assert(l->can_be_counted_loop(this), "sanity");
849 _igvn.register_new_node_with_optimizer(l);
850 set_loop(l, loop);
851 loop->_head = l;
852 // Fix all data nodes placed at the old loop head.
853 // Uses the lazy-update mechanism of 'get_ctrl'.
854 lazy_replace( x, l );
855 set_idom(l, entry_control, dom_depth(entry_control) + 1);
856
857 if (LoopStripMiningIter == 0 || strip_mine_loop) {
858 // Check for immediately preceding SafePoint and remove
859 if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) {
860 if (strip_mine_loop) {
861 Node* outer_le = outer_ilt->_tail->in(0);
862 Node* sfpt = sfpt2->clone();
863 sfpt->set_req(0, iffalse);
864 outer_le->set_req(0, sfpt);
865 register_control(sfpt, outer_ilt, iffalse);
866 set_idom(outer_le, sfpt, dom_depth(sfpt));
867 }
868 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
869 if (loop->_safepts != NULL) {
870 loop->_safepts->yank(sfpt2);
871 }
872 }
873 }
874
875 // Free up intermediate goo
876 _igvn.remove_dead_node(hook);
877
878 #ifdef ASSERT
879 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
880 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
881 #endif
882 #ifndef PRODUCT
883 if (TraceLoopOpts) {
884 tty->print("Counted ");
885 loop->dump_head();
886 }
887 #endif
888
889 C->print_method(PHASE_AFTER_CLOOPS, 3);
890
891 // Capture bounds of the loop in the induction variable Phi before
892 // subsequent transformation (iteration splitting) obscures the
893 // bounds
894 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
895
896 if (strip_mine_loop) {
897 l->mark_strip_mined();
898 l->verify_strip_mined(1);
899 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
900 loop = outer_ilt;
901 }
902
903 return true;
904 }
905
906 //----------------------exact_limit-------------------------------------------
907 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
908 assert(loop->_head->is_CountedLoop(), "");
909 CountedLoopNode *cl = loop->_head->as_CountedLoop();
910 assert(cl->is_valid_counted_loop(), "");
911
912 if (ABS(cl->stride_con()) == 1 ||
913 cl->limit()->Opcode() == Op_LoopLimit) {
914 // Old code has exact limit (it could be incorrect in case of int overflow).
915 // Loop limit is exact with stride == 1. And loop may already have exact limit.
916 return cl->limit();
917 }
918 Node *limit = NULL;
919 #ifdef ASSERT
920 BoolTest::mask bt = cl->loopexit()->test_trip();
921 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
922 #endif
923 if (cl->has_exact_trip_count()) {
924 // Simple case: loop has constant boundaries.
925 // Use jlongs to avoid integer overflow.
926 int stride_con = cl->stride_con();
927 jlong init_con = cl->init_trip()->get_int();
928 jlong limit_con = cl->limit()->get_int();
929 julong trip_cnt = cl->trip_count();
930 jlong final_con = init_con + trip_cnt*stride_con;
931 int final_int = (int)final_con;
932 // The final value should be in integer range since the loop
933 // is counted and the limit was checked for overflow.
934 assert(final_con == (jlong)final_int, "final value should be integer");
935 limit = _igvn.intcon(final_int);
936 } else {
937 // Create new LoopLimit node to get exact limit (final iv value).
938 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
939 register_new_node(limit, cl->in(LoopNode::EntryControl));
940 }
941 assert(limit != NULL, "sanity");
942 return limit;
943 }
944
945 //------------------------------Ideal------------------------------------------
946 // Return a node which is more "ideal" than the current node.
947 // Attempt to convert into a counted-loop.
948 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
949 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
950 phase->C->set_major_progress();
951 }
952 return RegionNode::Ideal(phase, can_reshape);
953 }
954
955 void LoopNode::verify_strip_mined(int expect_skeleton) const {
956 #ifdef ASSERT
957 const OuterStripMinedLoopNode* outer = NULL;
958 const CountedLoopNode* inner = NULL;
959 if (is_strip_mined()) {
960 assert(is_CountedLoop(), "no Loop should be marked strip mined");
961 inner = as_CountedLoop();
962 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
963 } else if (is_OuterStripMinedLoop()) {
964 outer = this->as_OuterStripMinedLoop();
965 inner = outer->unique_ctrl_out()->as_CountedLoop();
966 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
967 }
968 if (inner != NULL || outer != NULL) {
969 assert(inner != NULL && outer != NULL, "missing loop in strip mined nest");
970 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
971 Node* outer_le = outer_tail->in(0);
972 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
973 Node* sfpt = outer_le->in(0);
974 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
975 Node* inner_out = sfpt->in(0);
976 if (inner_out->outcnt() != 1) {
977 ResourceMark rm;
978 Unique_Node_List wq;
979
980 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
981 Node* u = inner_out->fast_out(i);
982 if (u == sfpt) {
983 continue;
984 }
985 wq.clear();
986 wq.push(u);
987 bool found_sfpt = false;
988 for (uint next = 0; next < wq.size() && !found_sfpt; next++) {
989 Node *n = wq.at(next);
990 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) {
991 Node* u = n->fast_out(i);
992 if (u == sfpt) {
993 found_sfpt = true;
994 }
995 if (!u->is_CFG()) {
996 wq.push(u);
997 }
998 }
999 }
1000 assert(found_sfpt, "no node in loop that's not input to safepoint");
1001 }
1002 }
1003 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
1004 assert(cle == inner->loopexit_or_null(), "mismatch");
1005 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
1006 if (has_skeleton) {
1007 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
1008 assert(outer->outcnt() == 2, "only phis");
1009 } else {
1010 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
1011 uint phis = 0;
1012 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
1013 Node* u = inner->fast_out(i);
1014 if (u->is_Phi()) {
1015 phis++;
1016 }
1017 }
1018 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
1019 Node* u = outer->fast_out(i);
1020 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
1021 }
1022 uint stores = 0;
1023 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
1024 Node* u = inner_out->fast_out(i);
1025 if (u->is_Store()) {
1026 stores++;
1027 }
1028 }
1029 assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
1030 }
1031 assert(sfpt->outcnt() == 1, "no data node");
1032 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
1033 }
1034 #endif
1035 }
1036
1037 //=============================================================================
1038 //------------------------------Ideal------------------------------------------
1039 // Return a node which is more "ideal" than the current node.
1040 // Attempt to convert into a counted-loop.
1041 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1042 return RegionNode::Ideal(phase, can_reshape);
1043 }
1044
1045 //------------------------------dump_spec--------------------------------------
1046 // Dump special per-node info
1047 #ifndef PRODUCT
1048 void CountedLoopNode::dump_spec(outputStream *st) const {
1049 LoopNode::dump_spec(st);
1050 if (stride_is_con()) {
1051 st->print("stride: %d ",stride_con());
1052 }
1053 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
1054 if (is_main_loop()) st->print("main of N%d", _idx);
1055 if (is_post_loop()) st->print("post of N%d", _main_idx);
1056 if (is_strip_mined()) st->print(" strip mined");
1057 }
1058 #endif
1059
1060 //=============================================================================
1061 int CountedLoopEndNode::stride_con() const {
1062 return stride()->bottom_type()->is_int()->get_con();
1063 }
1064
1065 //=============================================================================
1066 //------------------------------Value-----------------------------------------
1067 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
1068 const Type* init_t = phase->type(in(Init));
1069 const Type* limit_t = phase->type(in(Limit));
1070 const Type* stride_t = phase->type(in(Stride));
1071 // Either input is TOP ==> the result is TOP
1072 if (init_t == Type::TOP) return Type::TOP;
1073 if (limit_t == Type::TOP) return Type::TOP;
1074 if (stride_t == Type::TOP) return Type::TOP;
1075
1076 int stride_con = stride_t->is_int()->get_con();
1077 if (stride_con == 1)
1078 return NULL; // Identity
1079
1080 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
1081 // Use jlongs to avoid integer overflow.
1082 jlong init_con = init_t->is_int()->get_con();
1083 jlong limit_con = limit_t->is_int()->get_con();
1084 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
1085 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
1086 jlong final_con = init_con + stride_con*trip_count;
1087 int final_int = (int)final_con;
1088 // The final value should be in integer range since the loop
1089 // is counted and the limit was checked for overflow.
1090 assert(final_con == (jlong)final_int, "final value should be integer");
1091 return TypeInt::make(final_int);
1092 }
1093
1094 return bottom_type(); // TypeInt::INT
1095 }
1096
1097 //------------------------------Ideal------------------------------------------
1098 // Return a node which is more "ideal" than the current node.
1099 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1100 if (phase->type(in(Init)) == Type::TOP ||
1101 phase->type(in(Limit)) == Type::TOP ||
1102 phase->type(in(Stride)) == Type::TOP)
1103 return NULL; // Dead
1104
1105 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1106 if (stride_con == 1)
1107 return NULL; // Identity
1108
1109 if (in(Init)->is_Con() && in(Limit)->is_Con())
1110 return NULL; // Value
1111
1112 // Delay following optimizations until all loop optimizations
1113 // done to keep Ideal graph simple.
1114 if (!can_reshape || phase->C->major_progress())
1115 return NULL;
1116
1117 const TypeInt* init_t = phase->type(in(Init) )->is_int();
1118 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
1119 int stride_p;
1120 jlong lim, ini;
1121 julong max;
1122 if (stride_con > 0) {
1123 stride_p = stride_con;
1124 lim = limit_t->_hi;
1125 ini = init_t->_lo;
1126 max = (julong)max_jint;
1127 } else {
1128 stride_p = -stride_con;
1129 lim = init_t->_hi;
1130 ini = limit_t->_lo;
1131 max = (julong)min_jint;
1132 }
1133 julong range = lim - ini + stride_p;
1134 if (range <= max) {
1135 // Convert to integer expression if it is not overflow.
1136 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
1137 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
1138 Node *bias = phase->transform(new AddINode(range, stride_m));
1139 Node *trip = phase->transform(new DivINode(0, bias, in(Stride)));
1140 Node *span = phase->transform(new MulINode(trip, in(Stride)));
1141 return new AddINode(span, in(Init)); // exact limit
1142 }
1143
1144 if (is_power_of_2(stride_p) || // divisor is 2^n
1145 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
1146 // Convert to long expression to avoid integer overflow
1147 // and let igvn optimizer convert this division.
1148 //
1149 Node* init = phase->transform( new ConvI2LNode(in(Init)));
1150 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
1151 Node* stride = phase->longcon(stride_con);
1152 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
1153
1154 Node *range = phase->transform(new SubLNode(limit, init));
1155 Node *bias = phase->transform(new AddLNode(range, stride_m));
1156 Node *span;
1157 if (stride_con > 0 && is_power_of_2(stride_p)) {
1158 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
1159 // and avoid generating rounding for division. Zero trip guard should
1160 // guarantee that init < limit but sometimes the guard is missing and
1161 // we can get situation when init > limit. Note, for the empty loop
1162 // optimization zero trip guard is generated explicitly which leaves
1163 // only RCE predicate where exact limit is used and the predicate
1164 // will simply fail forcing recompilation.
1165 Node* neg_stride = phase->longcon(-stride_con);
1166 span = phase->transform(new AndLNode(bias, neg_stride));
1167 } else {
1168 Node *trip = phase->transform(new DivLNode(0, bias, stride));
1169 span = phase->transform(new MulLNode(trip, stride));
1170 }
1171 // Convert back to int
1172 Node *span_int = phase->transform(new ConvL2INode(span));
1173 return new AddINode(span_int, in(Init)); // exact limit
1174 }
1175
1176 return NULL; // No progress
1177 }
1178
1179 //------------------------------Identity---------------------------------------
1180 // If stride == 1 return limit node.
1181 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1182 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1183 if (stride_con == 1 || stride_con == -1)
1184 return in(Limit);
1185 return this;
1186 }
1187
1188 //=============================================================================
1189 //----------------------match_incr_with_optional_truncation--------------------
1190 // Match increment with optional truncation:
1191 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1192 // Return NULL for failure. Success returns the increment node.
1193 Node* CountedLoopNode::match_incr_with_optional_truncation(
1194 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1195 // Quick cutouts:
1196 if (expr == NULL || expr->req() != 3) return NULL;
1197
1198 Node *t1 = NULL;
1199 Node *t2 = NULL;
1200 const TypeInt* trunc_t = TypeInt::INT;
1201 Node* n1 = expr;
1202 int n1op = n1->Opcode();
1203
1204 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1205 if (n1op == Op_AndI &&
1206 n1->in(2)->is_Con() &&
1207 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1208 // %%% This check should match any mask of 2**K-1.
1209 t1 = n1;
1210 n1 = t1->in(1);
1211 n1op = n1->Opcode();
1212 trunc_t = TypeInt::CHAR;
1213 } else if (n1op == Op_RShiftI &&
1214 n1->in(1) != NULL &&
1215 n1->in(1)->Opcode() == Op_LShiftI &&
1216 n1->in(2) == n1->in(1)->in(2) &&
1217 n1->in(2)->is_Con()) {
1218 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1219 // %%% This check should match any shift in [1..31].
1220 if (shift == 16 || shift == 8) {
1221 t1 = n1;
1222 t2 = t1->in(1);
1223 n1 = t2->in(1);
1224 n1op = n1->Opcode();
1225 if (shift == 16) {
1226 trunc_t = TypeInt::SHORT;
1227 } else if (shift == 8) {
1228 trunc_t = TypeInt::BYTE;
1229 }
1230 }
1231 }
1232
1233 // If (maybe after stripping) it is an AddI, we won:
1234 if (n1op == Op_AddI) {
1235 *trunc1 = t1;
1236 *trunc2 = t2;
1237 *trunc_type = trunc_t;
1238 return n1;
1239 }
1240
1241 // failed
1242 return NULL;
1243 }
1244
1245 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
1246 if (is_strip_mined()) {
1247 verify_strip_mined(expect_skeleton);
1248 return in(EntryControl)->as_Loop();
1249 }
1250 return this;
1251 }
1252
1253 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
1254 assert(is_strip_mined(), "not a strip mined loop");
1255 Node* c = in(EntryControl);
1256 if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) {
1257 return NULL;
1258 }
1259 return c->as_OuterStripMinedLoop();
1260 }
1261
1262 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
1263 Node* c = in(LoopBackControl);
1264 if (c == NULL || c->is_top()) {
1265 return NULL;
1266 }
1267 return c->as_IfTrue();
1268 }
1269
1270 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
1271 LoopNode* l = outer_loop();
1272 if (l == NULL) {
1273 return NULL;
1274 }
1275 return l->outer_loop_tail();
1276 }
1277
1278 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
1279 IfTrueNode* proj = outer_loop_tail();
1280 if (proj == NULL) {
1281 return NULL;
1282 }
1283 Node* c = proj->in(0);
1284 if (c == NULL || c->is_top() || c->outcnt() != 2) {
1285 return NULL;
1286 }
1287 return c->as_OuterStripMinedLoopEnd();
1288 }
1289
1290 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
1291 LoopNode* l = outer_loop();
1292 if (l == NULL) {
1293 return NULL;
1294 }
1295 return l->outer_loop_end();
1296 }
1297
1298 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
1299 IfNode* le = outer_loop_end();
1300 if (le == NULL) {
1301 return NULL;
1302 }
1303 Node* c = le->proj_out_or_null(false);
1304 if (c == NULL) {
1305 return NULL;
1306 }
1307 return c->as_IfFalse();
1308 }
1309
1310 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
1311 LoopNode* l = outer_loop();
1312 if (l == NULL) {
1313 return NULL;
1314 }
1315 return l->outer_loop_exit();
1316 }
1317
1318 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
1319 IfNode* le = outer_loop_end();
1320 if (le == NULL) {
1321 return NULL;
1322 }
1323 Node* c = le->in(0);
1324 if (c == NULL || c->is_top()) {
1325 return NULL;
1326 }
1327 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
1328 return c->as_SafePoint();
1329 }
1330
1331 SafePointNode* CountedLoopNode::outer_safepoint() const {
1332 LoopNode* l = outer_loop();
1333 if (l == NULL) {
1334 return NULL;
1335 }
1336 return l->outer_safepoint();
1337 }
1338
1339 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) {
1340 while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() &&
1341 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 &&
1342 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) {
1343 ctrl = ctrl->in(0)->in(0);
1344 }
1345
1346 return ctrl;
1347 }
1348
1349 Node* CountedLoopNode::skip_predicates() {
1350 if (is_main_loop()) {
1351 Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
1352
1353 return skip_predicates_from_entry(ctrl);
1354 }
1355 return in(LoopNode::EntryControl);
1356 }
1357
1358 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
1359 // Look for the outer & inner strip mined loop, reduce number of
1360 // iterations of the inner loop, set exit condition of outer loop,
1361 // construct required phi nodes for outer loop.
1362 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
1363 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
1364 Node* inner_iv_phi = inner_cl->phi();
1365 if (inner_iv_phi == NULL) {
1366 IfNode* outer_le = outer_loop_end();
1367 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1368 igvn->replace_node(outer_le, iff);
1369 inner_cl->clear_strip_mined();
1370 return;
1371 }
1372 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
1373
1374 int stride = inner_cl->stride_con();
1375 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride);
1376 int scaled_iters = (int)scaled_iters_long;
1377 int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride);
1378 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
1379 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
1380 assert(iter_estimate > 0, "broken");
1381 if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) {
1382 // Remove outer loop and safepoint (too few iterations)
1383 Node* outer_sfpt = outer_safepoint();
1384 Node* outer_out = outer_loop_exit();
1385 igvn->replace_node(outer_out, outer_sfpt->in(0));
1386 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
1387 inner_cl->clear_strip_mined();
1388 return;
1389 }
1390 if (iter_estimate <= scaled_iters_long) {
1391 // We would only go through one iteration of
1392 // the outer loop: drop the outer loop but
1393 // keep the safepoint so we don't run for
1394 // too long without a safepoint
1395 IfNode* outer_le = outer_loop_end();
1396 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1397 igvn->replace_node(outer_le, iff);
1398 inner_cl->clear_strip_mined();
1399 return;
1400 }
1401
1402 Node* cle_tail = inner_cle->proj_out(true);
1403 ResourceMark rm;
1404 Node_List old_new;
1405 if (cle_tail->outcnt() > 1) {
1406 // Look for nodes on backedge of inner loop and clone them
1407 Unique_Node_List backedge_nodes;
1408 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
1409 Node* u = cle_tail->fast_out(i);
1410 if (u != inner_cl) {
1411 assert(!u->is_CFG(), "control flow on the backedge?");
1412 backedge_nodes.push(u);
1413 }
1414 }
1415 uint last = igvn->C->unique();
1416 for (uint next = 0; next < backedge_nodes.size(); next++) {
1417 Node* n = backedge_nodes.at(next);
1418 old_new.map(n->_idx, n->clone());
1419 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1420 Node* u = n->fast_out(i);
1421 assert(!u->is_CFG(), "broken");
1422 if (u->_idx >= last) {
1423 continue;
1424 }
1425 if (!u->is_Phi()) {
1426 backedge_nodes.push(u);
1427 } else {
1428 assert(u->in(0) == inner_cl, "strange phi on the backedge");
1429 }
1430 }
1431 }
1432 // Put the clones on the outer loop backedge
1433 Node* le_tail = outer_loop_tail();
1434 for (uint next = 0; next < backedge_nodes.size(); next++) {
1435 Node *n = old_new[backedge_nodes.at(next)->_idx];
1436 for (uint i = 1; i < n->req(); i++) {
1437 if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) {
1438 n->set_req(i, old_new[n->in(i)->_idx]);
1439 }
1440 }
1441 if (n->in(0) != NULL && n->in(0) == cle_tail) {
1442 n->set_req(0, le_tail);
1443 }
1444 igvn->register_new_node_with_optimizer(n);
1445 }
1446 }
1447
1448 Node* iv_phi = NULL;
1449 // Make a clone of each phi in the inner loop
1450 // for the outer loop
1451 for (uint i = 0; i < inner_cl->outcnt(); i++) {
1452 Node* u = inner_cl->raw_out(i);
1453 if (u->is_Phi()) {
1454 assert(u->in(0) == inner_cl, "inconsistent");
1455 Node* phi = u->clone();
1456 phi->set_req(0, this);
1457 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
1458 if (be != NULL) {
1459 phi->set_req(LoopNode::LoopBackControl, be);
1460 }
1461 phi = igvn->transform(phi);
1462 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
1463 if (u == inner_iv_phi) {
1464 iv_phi = phi;
1465 }
1466 }
1467 }
1468 Node* cle_out = inner_cle->proj_out(false);
1469 if (cle_out->outcnt() > 1) {
1470 // Look for chains of stores that were sunk
1471 // out of the inner loop and are in the outer loop
1472 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
1473 Node* u = cle_out->fast_out(i);
1474 if (u->is_Store()) {
1475 Node* first = u;
1476 for(;;) {
1477 Node* next = first->in(MemNode::Memory);
1478 if (!next->is_Store() || next->in(0) != cle_out) {
1479 break;
1480 }
1481 first = next;
1482 }
1483 Node* last = u;
1484 for(;;) {
1485 Node* next = NULL;
1486 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
1487 Node* uu = last->fast_out(j);
1488 if (uu->is_Store() && uu->in(0) == cle_out) {
1489 assert(next == NULL, "only one in the outer loop");
1490 next = uu;
1491 }
1492 }
1493 if (next == NULL) {
1494 break;
1495 }
1496 last = next;
1497 }
1498 Node* phi = NULL;
1499 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1500 Node* uu = fast_out(j);
1501 if (uu->is_Phi()) {
1502 Node* be = uu->in(LoopNode::LoopBackControl);
1503 if (be->is_Store() && old_new[be->_idx] != NULL) {
1504 assert(false, "store on the backedge + sunk stores: unsupported");
1505 // drop outer loop
1506 IfNode* outer_le = outer_loop_end();
1507 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1508 igvn->replace_node(outer_le, iff);
1509 inner_cl->clear_strip_mined();
1510 return;
1511 }
1512 if (be == last || be == first->in(MemNode::Memory)) {
1513 assert(phi == NULL, "only one phi");
1514 phi = uu;
1515 }
1516 }
1517 }
1518 #ifdef ASSERT
1519 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1520 Node* uu = fast_out(j);
1521 if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) {
1522 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
1523 assert(phi == uu, "what's that phi?");
1524 } else if (uu->adr_type() == TypePtr::BOTTOM) {
1525 Node* n = uu->in(LoopNode::LoopBackControl);
1526 uint limit = igvn->C->live_nodes();
1527 uint i = 0;
1528 while (n != uu) {
1529 i++;
1530 assert(i < limit, "infinite loop");
1531 if (n->is_Proj()) {
1532 n = n->in(0);
1533 } else if (n->is_SafePoint() || n->is_MemBar()) {
1534 n = n->in(TypeFunc::Memory);
1535 } else if (n->is_Phi()) {
1536 n = n->in(1);
1537 } else if (n->is_MergeMem()) {
1538 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
1539 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
1540 n = n->in(MemNode::Memory);
1541 } else {
1542 n->dump();
1543 ShouldNotReachHere();
1544 }
1545 }
1546 }
1547 }
1548 }
1549 #endif
1550 if (phi == NULL) {
1551 // If the an entire chains was sunk, the
1552 // inner loop has no phi for that memory
1553 // slice, create one for the outer loop
1554 phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY,
1555 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
1556 phi->set_req(LoopNode::LoopBackControl, last);
1557 phi = igvn->transform(phi);
1558 igvn->replace_input_of(first, MemNode::Memory, phi);
1559 } else {
1560 // Or fix the outer loop fix to include
1561 // that chain of stores.
1562 Node* be = phi->in(LoopNode::LoopBackControl);
1563 assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported");
1564 if (be == first->in(MemNode::Memory)) {
1565 if (be == phi->in(LoopNode::LoopBackControl)) {
1566 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
1567 } else {
1568 igvn->replace_input_of(be, MemNode::Memory, last);
1569 }
1570 } else {
1571 #ifdef ASSERT
1572 if (be == phi->in(LoopNode::LoopBackControl)) {
1573 assert(phi->in(LoopNode::LoopBackControl) == last, "");
1574 } else {
1575 assert(be->in(MemNode::Memory) == last, "");
1576 }
1577 #endif
1578 }
1579 }
1580 }
1581 }
1582 }
1583
1584 if (iv_phi != NULL) {
1585 // Now adjust the inner loop's exit condition
1586 Node* limit = inner_cl->limit();
1587 Node* sub = NULL;
1588 if (stride > 0) {
1589 sub = igvn->transform(new SubINode(limit, iv_phi));
1590 } else {
1591 sub = igvn->transform(new SubINode(iv_phi, limit));
1592 }
1593 Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters)));
1594 Node* new_limit = NULL;
1595 if (stride > 0) {
1596 new_limit = igvn->transform(new AddINode(min, iv_phi));
1597 } else {
1598 new_limit = igvn->transform(new SubINode(iv_phi, min));
1599 }
1600 Node* inner_cmp = inner_cle->cmp_node();
1601 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
1602 Node* outer_bol = inner_bol;
1603 // cmp node for inner loop may be shared
1604 inner_cmp = inner_cmp->clone();
1605 inner_cmp->set_req(2, new_limit);
1606 inner_bol = inner_bol->clone();
1607 inner_bol->set_req(1, igvn->transform(inner_cmp));
1608 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
1609 // Set the outer loop's exit condition too
1610 igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
1611 } else {
1612 assert(false, "should be able to adjust outer loop");
1613 IfNode* outer_le = outer_loop_end();
1614 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1615 igvn->replace_node(outer_le, iff);
1616 inner_cl->clear_strip_mined();
1617 }
1618 }
1619
1620 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
1621 if (!in(0)) return Type::TOP;
1622 if (phase->type(in(0)) == Type::TOP)
1623 return Type::TOP;
1624
1625 return TypeTuple::IFBOTH;
1626 }
1627
1628 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1629 if (remove_dead_region(phase, can_reshape)) return this;
1630
1631 return NULL;
1632 }
1633
1634 //------------------------------filtered_type--------------------------------
1635 // Return a type based on condition control flow
1636 // A successful return will be a type that is restricted due
1637 // to a series of dominating if-tests, such as:
1638 // if (i < 10) {
1639 // if (i > 0) {
1640 // here: "i" type is [1..10)
1641 // }
1642 // }
1643 // or a control flow merge
1644 // if (i < 10) {
1645 // do {
1646 // phi( , ) -- at top of loop type is [min_int..10)
1647 // i = ?
1648 // } while ( i < 10)
1649 //
1650 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1651 assert(n && n->bottom_type()->is_int(), "must be int");
1652 const TypeInt* filtered_t = NULL;
1653 if (!n->is_Phi()) {
1654 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1655 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1656
1657 } else {
1658 Node* phi = n->as_Phi();
1659 Node* region = phi->in(0);
1660 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1661 if (region && region != C->top()) {
1662 for (uint i = 1; i < phi->req(); i++) {
1663 Node* val = phi->in(i);
1664 Node* use_c = region->in(i);
1665 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1666 if (val_t != NULL) {
1667 if (filtered_t == NULL) {
1668 filtered_t = val_t;
1669 } else {
1670 filtered_t = filtered_t->meet(val_t)->is_int();
1671 }
1672 }
1673 }
1674 }
1675 }
1676 const TypeInt* n_t = _igvn.type(n)->is_int();
1677 if (filtered_t != NULL) {
1678 n_t = n_t->join(filtered_t)->is_int();
1679 }
1680 return n_t;
1681 }
1682
1683
1684 //------------------------------filtered_type_from_dominators--------------------------------
1685 // Return a possibly more restrictive type for val based on condition control flow of dominators
1686 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1687 if (val->is_Con()) {
1688 return val->bottom_type()->is_int();
1689 }
1690 uint if_limit = 10; // Max number of dominating if's visited
1691 const TypeInt* rtn_t = NULL;
1692
1693 if (use_ctrl && use_ctrl != C->top()) {
1694 Node* val_ctrl = get_ctrl(val);
1695 uint val_dom_depth = dom_depth(val_ctrl);
1696 Node* pred = use_ctrl;
1697 uint if_cnt = 0;
1698 while (if_cnt < if_limit) {
1699 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1700 if_cnt++;
1701 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1702 if (if_t != NULL) {
1703 if (rtn_t == NULL) {
1704 rtn_t = if_t;
1705 } else {
1706 rtn_t = rtn_t->join(if_t)->is_int();
1707 }
1708 }
1709 }
1710 pred = idom(pred);
1711 if (pred == NULL || pred == C->top()) {
1712 break;
1713 }
1714 // Stop if going beyond definition block of val
1715 if (dom_depth(pred) < val_dom_depth) {
1716 break;
1717 }
1718 }
1719 }
1720 return rtn_t;
1721 }
1722
1723
1724 //------------------------------dump_spec--------------------------------------
1725 // Dump special per-node info
1726 #ifndef PRODUCT
1727 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1728 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1729 BoolTest bt( test_trip()); // Added this for g++.
1730
1731 st->print("[");
1732 bt.dump_on(st);
1733 st->print("]");
1734 }
1735 st->print(" ");
1736 IfNode::dump_spec(st);
1737 }
1738 #endif
1739
1740 //=============================================================================
1741 //------------------------------is_member--------------------------------------
1742 // Is 'l' a member of 'this'?
1743 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1744 while( l->_nest > _nest ) l = l->_parent;
1745 return l == this;
1746 }
1747
1748 //------------------------------set_nest---------------------------------------
1749 // Set loop tree nesting depth. Accumulate _has_call bits.
1750 int IdealLoopTree::set_nest( uint depth ) {
1751 _nest = depth;
1752 int bits = _has_call;
1753 if( _child ) bits |= _child->set_nest(depth+1);
1754 if( bits ) _has_call = 1;
1755 if( _next ) bits |= _next ->set_nest(depth );
1756 return bits;
1757 }
1758
1759 //------------------------------split_fall_in----------------------------------
1760 // Split out multiple fall-in edges from the loop header. Move them to a
1761 // private RegionNode before the loop. This becomes the loop landing pad.
1762 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1763 PhaseIterGVN &igvn = phase->_igvn;
1764 uint i;
1765
1766 // Make a new RegionNode to be the landing pad.
1767 Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1768 phase->set_loop(landing_pad,_parent);
1769 // Gather all the fall-in control paths into the landing pad
1770 uint icnt = fall_in_cnt;
1771 uint oreq = _head->req();
1772 for( i = oreq-1; i>0; i-- )
1773 if( !phase->is_member( this, _head->in(i) ) )
1774 landing_pad->set_req(icnt--,_head->in(i));
1775
1776 // Peel off PhiNode edges as well
1777 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1778 Node *oj = _head->fast_out(j);
1779 if( oj->is_Phi() ) {
1780 PhiNode* old_phi = oj->as_Phi();
1781 assert( old_phi->region() == _head, "" );
1782 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1783 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1784 uint icnt = fall_in_cnt;
1785 for( i = oreq-1; i>0; i-- ) {
1786 if( !phase->is_member( this, _head->in(i) ) ) {
1787 p->init_req(icnt--, old_phi->in(i));
1788 // Go ahead and clean out old edges from old phi
1789 old_phi->del_req(i);
1790 }
1791 }
1792 // Search for CSE's here, because ZKM.jar does a lot of
1793 // loop hackery and we need to be a little incremental
1794 // with the CSE to avoid O(N^2) node blow-up.
1795 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1796 if( p2 ) { // Found CSE
1797 p->destruct(); // Recover useless new node
1798 p = p2; // Use old node
1799 } else {
1800 igvn.register_new_node_with_optimizer(p, old_phi);
1801 }
1802 // Make old Phi refer to new Phi.
1803 old_phi->add_req(p);
1804 // Check for the special case of making the old phi useless and
1805 // disappear it. In JavaGrande I have a case where this useless
1806 // Phi is the loop limit and prevents recognizing a CountedLoop
1807 // which in turn prevents removing an empty loop.
1808 Node *id_old_phi = igvn.apply_identity(old_phi);
1809 if( id_old_phi != old_phi ) { // Found a simple identity?
1810 // Note that I cannot call 'replace_node' here, because
1811 // that will yank the edge from old_phi to the Region and
1812 // I'm mid-iteration over the Region's uses.
1813 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1814 Node* use = old_phi->last_out(i);
1815 igvn.rehash_node_delayed(use);
1816 uint uses_found = 0;
1817 for (uint j = 0; j < use->len(); j++) {
1818 if (use->in(j) == old_phi) {
1819 if (j < use->req()) use->set_req (j, id_old_phi);
1820 else use->set_prec(j, id_old_phi);
1821 uses_found++;
1822 }
1823 }
1824 i -= uses_found; // we deleted 1 or more copies of this edge
1825 }
1826 }
1827 igvn._worklist.push(old_phi);
1828 }
1829 }
1830 // Finally clean out the fall-in edges from the RegionNode
1831 for( i = oreq-1; i>0; i-- ) {
1832 if( !phase->is_member( this, _head->in(i) ) ) {
1833 _head->del_req(i);
1834 }
1835 }
1836 igvn.rehash_node_delayed(_head);
1837 // Transform landing pad
1838 igvn.register_new_node_with_optimizer(landing_pad, _head);
1839 // Insert landing pad into the header
1840 _head->add_req(landing_pad);
1841 }
1842
1843 //------------------------------split_outer_loop-------------------------------
1844 // Split out the outermost loop from this shared header.
1845 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1846 PhaseIterGVN &igvn = phase->_igvn;
1847
1848 // Find index of outermost loop; it should also be my tail.
1849 uint outer_idx = 1;
1850 while( _head->in(outer_idx) != _tail ) outer_idx++;
1851
1852 // Make a LoopNode for the outermost loop.
1853 Node *ctl = _head->in(LoopNode::EntryControl);
1854 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1855 outer = igvn.register_new_node_with_optimizer(outer, _head);
1856 phase->set_created_loop_node();
1857
1858 // Outermost loop falls into '_head' loop
1859 _head->set_req(LoopNode::EntryControl, outer);
1860 _head->del_req(outer_idx);
1861 // Split all the Phis up between '_head' loop and 'outer' loop.
1862 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1863 Node *out = _head->fast_out(j);
1864 if( out->is_Phi() ) {
1865 PhiNode *old_phi = out->as_Phi();
1866 assert( old_phi->region() == _head, "" );
1867 Node *phi = PhiNode::make_blank(outer, old_phi);
1868 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1869 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1870 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1871 // Make old Phi point to new Phi on the fall-in path
1872 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1873 old_phi->del_req(outer_idx);
1874 }
1875 }
1876
1877 // Use the new loop head instead of the old shared one
1878 _head = outer;
1879 phase->set_loop(_head, this);
1880 }
1881
1882 //------------------------------fix_parent-------------------------------------
1883 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1884 loop->_parent = parent;
1885 if( loop->_child ) fix_parent( loop->_child, loop );
1886 if( loop->_next ) fix_parent( loop->_next , parent );
1887 }
1888
1889 //------------------------------estimate_path_freq-----------------------------
1890 static float estimate_path_freq( Node *n ) {
1891 // Try to extract some path frequency info
1892 IfNode *iff;
1893 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1894 uint nop = n->Opcode();
1895 if( nop == Op_SafePoint ) { // Skip any safepoint
1896 n = n->in(0);
1897 continue;
1898 }
1899 if( nop == Op_CatchProj ) { // Get count from a prior call
1900 // Assume call does not always throw exceptions: means the call-site
1901 // count is also the frequency of the fall-through path.
1902 assert( n->is_CatchProj(), "" );
1903 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1904 return 0.0f; // Assume call exception path is rare
1905 Node *call = n->in(0)->in(0)->in(0);
1906 assert( call->is_Call(), "expect a call here" );
1907 const JVMState *jvms = ((CallNode*)call)->jvms();
1908 ciMethodData* methodData = jvms->method()->method_data();
1909 if (!methodData->is_mature()) return 0.0f; // No call-site data
1910 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1911 if ((data == NULL) || !data->is_CounterData()) {
1912 // no call profile available, try call's control input
1913 n = n->in(0);
1914 continue;
1915 }
1916 return data->as_CounterData()->count()/FreqCountInvocations;
1917 }
1918 // See if there's a gating IF test
1919 Node *n_c = n->in(0);
1920 if( !n_c->is_If() ) break; // No estimate available
1921 iff = n_c->as_If();
1922 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1923 // Compute how much count comes on this path
1924 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1925 // Have no count info. Skip dull uncommon-trap like branches.
1926 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1927 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1928 break;
1929 // Skip through never-taken branch; look for a real loop exit.
1930 n = iff->in(0);
1931 }
1932 return 0.0f; // No estimate available
1933 }
1934
1935 //------------------------------merge_many_backedges---------------------------
1936 // Merge all the backedges from the shared header into a private Region.
1937 // Feed that region as the one backedge to this loop.
1938 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1939 uint i;
1940
1941 // Scan for the top 2 hottest backedges
1942 float hotcnt = 0.0f;
1943 float warmcnt = 0.0f;
1944 uint hot_idx = 0;
1945 // Loop starts at 2 because slot 1 is the fall-in path
1946 for( i = 2; i < _head->req(); i++ ) {
1947 float cnt = estimate_path_freq(_head->in(i));
1948 if( cnt > hotcnt ) { // Grab hottest path
1949 warmcnt = hotcnt;
1950 hotcnt = cnt;
1951 hot_idx = i;
1952 } else if( cnt > warmcnt ) { // And 2nd hottest path
1953 warmcnt = cnt;
1954 }
1955 }
1956
1957 // See if the hottest backedge is worthy of being an inner loop
1958 // by being much hotter than the next hottest backedge.
1959 if( hotcnt <= 0.0001 ||
1960 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1961
1962 // Peel out the backedges into a private merge point; peel
1963 // them all except optionally hot_idx.
1964 PhaseIterGVN &igvn = phase->_igvn;
1965
1966 Node *hot_tail = NULL;
1967 // Make a Region for the merge point
1968 Node *r = new RegionNode(1);
1969 for( i = 2; i < _head->req(); i++ ) {
1970 if( i != hot_idx )
1971 r->add_req( _head->in(i) );
1972 else hot_tail = _head->in(i);
1973 }
1974 igvn.register_new_node_with_optimizer(r, _head);
1975 // Plug region into end of loop _head, followed by hot_tail
1976 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1977 igvn.replace_input_of(_head, 2, r);
1978 if( hot_idx ) _head->add_req(hot_tail);
1979
1980 // Split all the Phis up between '_head' loop and the Region 'r'
1981 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1982 Node *out = _head->fast_out(j);
1983 if( out->is_Phi() ) {
1984 PhiNode* n = out->as_Phi();
1985 igvn.hash_delete(n); // Delete from hash before hacking edges
1986 Node *hot_phi = NULL;
1987 Node *phi = new PhiNode(r, n->type(), n->adr_type());
1988 // Check all inputs for the ones to peel out
1989 uint j = 1;
1990 for( uint i = 2; i < n->req(); i++ ) {
1991 if( i != hot_idx )
1992 phi->set_req( j++, n->in(i) );
1993 else hot_phi = n->in(i);
1994 }
1995 // Register the phi but do not transform until whole place transforms
1996 igvn.register_new_node_with_optimizer(phi, n);
1997 // Add the merge phi to the old Phi
1998 while( n->req() > 3 ) n->del_req( n->req()-1 );
1999 igvn.replace_input_of(n, 2, phi);
2000 if( hot_idx ) n->add_req(hot_phi);
2001 }
2002 }
2003
2004
2005 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
2006 // of self loop tree. Turn self into a loop headed by _head and with
2007 // tail being the new merge point.
2008 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
2009 phase->set_loop(_tail,ilt); // Adjust tail
2010 _tail = r; // Self's tail is new merge point
2011 phase->set_loop(r,this);
2012 ilt->_child = _child; // New guy has my children
2013 _child = ilt; // Self has new guy as only child
2014 ilt->_parent = this; // new guy has self for parent
2015 ilt->_nest = _nest; // Same nesting depth (for now)
2016
2017 // Starting with 'ilt', look for child loop trees using the same shared
2018 // header. Flatten these out; they will no longer be loops in the end.
2019 IdealLoopTree **pilt = &_child;
2020 while( ilt ) {
2021 if( ilt->_head == _head ) {
2022 uint i;
2023 for( i = 2; i < _head->req(); i++ )
2024 if( _head->in(i) == ilt->_tail )
2025 break; // Still a loop
2026 if( i == _head->req() ) { // No longer a loop
2027 // Flatten ilt. Hang ilt's "_next" list from the end of
2028 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
2029 IdealLoopTree **cp = &ilt->_child;
2030 while( *cp ) cp = &(*cp)->_next; // Find end of child list
2031 *cp = ilt->_next; // Hang next list at end of child list
2032 *pilt = ilt->_child; // Move child up to replace ilt
2033 ilt->_head = NULL; // Flag as a loop UNIONED into parent
2034 ilt = ilt->_child; // Repeat using new ilt
2035 continue; // do not advance over ilt->_child
2036 }
2037 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
2038 phase->set_loop(_head,ilt);
2039 }
2040 pilt = &ilt->_child; // Advance to next
2041 ilt = *pilt;
2042 }
2043
2044 if( _child ) fix_parent( _child, this );
2045 }
2046
2047 //------------------------------beautify_loops---------------------------------
2048 // Split shared headers and insert loop landing pads.
2049 // Insert a LoopNode to replace the RegionNode.
2050 // Return TRUE if loop tree is structurally changed.
2051 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
2052 bool result = false;
2053 // Cache parts in locals for easy
2054 PhaseIterGVN &igvn = phase->_igvn;
2055
2056 igvn.hash_delete(_head); // Yank from hash before hacking edges
2057
2058 // Check for multiple fall-in paths. Peel off a landing pad if need be.
2059 int fall_in_cnt = 0;
2060 for( uint i = 1; i < _head->req(); i++ )
2061 if( !phase->is_member( this, _head->in(i) ) )
2062 fall_in_cnt++;
2063 assert( fall_in_cnt, "at least 1 fall-in path" );
2064 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
2065 split_fall_in( phase, fall_in_cnt );
2066
2067 // Swap inputs to the _head and all Phis to move the fall-in edge to
2068 // the left.
2069 fall_in_cnt = 1;
2070 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
2071 fall_in_cnt++;
2072 if( fall_in_cnt > 1 ) {
2073 // Since I am just swapping inputs I do not need to update def-use info
2074 Node *tmp = _head->in(1);
2075 igvn.rehash_node_delayed(_head);
2076 _head->set_req( 1, _head->in(fall_in_cnt) );
2077 _head->set_req( fall_in_cnt, tmp );
2078 // Swap also all Phis
2079 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
2080 Node* phi = _head->fast_out(i);
2081 if( phi->is_Phi() ) {
2082 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
2083 tmp = phi->in(1);
2084 phi->set_req( 1, phi->in(fall_in_cnt) );
2085 phi->set_req( fall_in_cnt, tmp );
2086 }
2087 }
2088 }
2089 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
2090 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
2091
2092 // If I am a shared header (multiple backedges), peel off the many
2093 // backedges into a private merge point and use the merge point as
2094 // the one true backedge.
2095 if( _head->req() > 3 ) {
2096 // Merge the many backedges into a single backedge but leave
2097 // the hottest backedge as separate edge for the following peel.
2098 merge_many_backedges( phase );
2099 result = true;
2100 }
2101
2102 // If I have one hot backedge, peel off myself loop.
2103 // I better be the outermost loop.
2104 if (_head->req() > 3 && !_irreducible) {
2105 split_outer_loop( phase );
2106 result = true;
2107
2108 } else if (!_head->is_Loop() && !_irreducible) {
2109 // Make a new LoopNode to replace the old loop head
2110 Node *l = new LoopNode( _head->in(1), _head->in(2) );
2111 l = igvn.register_new_node_with_optimizer(l, _head);
2112 phase->set_created_loop_node();
2113 // Go ahead and replace _head
2114 phase->_igvn.replace_node( _head, l );
2115 _head = l;
2116 phase->set_loop(_head, this);
2117 }
2118
2119 // Now recursively beautify nested loops
2120 if( _child ) result |= _child->beautify_loops( phase );
2121 if( _next ) result |= _next ->beautify_loops( phase );
2122 return result;
2123 }
2124
2125 //------------------------------allpaths_check_safepts----------------------------
2126 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
2127 // encountered. Helper for check_safepts.
2128 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
2129 assert(stack.size() == 0, "empty stack");
2130 stack.push(_tail);
2131 visited.Clear();
2132 visited.set(_tail->_idx);
2133 while (stack.size() > 0) {
2134 Node* n = stack.pop();
2135 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2136 // Terminate this path
2137 } else if (n->Opcode() == Op_SafePoint) {
2138 if (_phase->get_loop(n) != this) {
2139 if (_required_safept == NULL) _required_safept = new Node_List();
2140 _required_safept->push(n); // save the one closest to the tail
2141 }
2142 // Terminate this path
2143 } else {
2144 uint start = n->is_Region() ? 1 : 0;
2145 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
2146 for (uint i = start; i < end; i++) {
2147 Node* in = n->in(i);
2148 assert(in->is_CFG(), "must be");
2149 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
2150 stack.push(in);
2151 }
2152 }
2153 }
2154 }
2155 }
2156
2157 //------------------------------check_safepts----------------------------
2158 // Given dominators, try to find loops with calls that must always be
2159 // executed (call dominates loop tail). These loops do not need non-call
2160 // safepoints (ncsfpt).
2161 //
2162 // A complication is that a safepoint in a inner loop may be needed
2163 // by an outer loop. In the following, the inner loop sees it has a
2164 // call (block 3) on every path from the head (block 2) to the
2165 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
2166 // in block 2, _but_ this leaves the outer loop without a safepoint.
2167 //
2168 // entry 0
2169 // |
2170 // v
2171 // outer 1,2 +->1
2172 // | |
2173 // | v
2174 // | 2<---+ ncsfpt in 2
2175 // |_/|\ |
2176 // | v |
2177 // inner 2,3 / 3 | call in 3
2178 // / | |
2179 // v +--+
2180 // exit 4
2181 //
2182 //
2183 // This method creates a list (_required_safept) of ncsfpt nodes that must
2184 // be protected is created for each loop. When a ncsfpt maybe deleted, it
2185 // is first looked for in the lists for the outer loops of the current loop.
2186 //
2187 // The insights into the problem:
2188 // A) counted loops are okay
2189 // B) innermost loops are okay (only an inner loop can delete
2190 // a ncsfpt needed by an outer loop)
2191 // C) a loop is immune from an inner loop deleting a safepoint
2192 // if the loop has a call on the idom-path
2193 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
2194 // idom-path that is not in a nested loop
2195 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
2196 // loop needs to be prevented from deletion by an inner loop
2197 //
2198 // There are two analyses:
2199 // 1) The first, and cheaper one, scans the loop body from
2200 // tail to head following the idom (immediate dominator)
2201 // chain, looking for the cases (C,D,E) above.
2202 // Since inner loops are scanned before outer loops, there is summary
2203 // information about inner loops. Inner loops can be skipped over
2204 // when the tail of an inner loop is encountered.
2205 //
2206 // 2) The second, invoked if the first fails to find a call or ncsfpt on
2207 // the idom path (which is rare), scans all predecessor control paths
2208 // from the tail to the head, terminating a path when a call or sfpt
2209 // is encountered, to find the ncsfpt's that are closest to the tail.
2210 //
2211 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
2212 // Bottom up traversal
2213 IdealLoopTree* ch = _child;
2214 if (_child) _child->check_safepts(visited, stack);
2215 if (_next) _next ->check_safepts(visited, stack);
2216
2217 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
2218 bool has_call = false; // call on dom-path
2219 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
2220 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
2221 // Scan the dom-path nodes from tail to head
2222 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
2223 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2224 has_call = true;
2225 _has_sfpt = 1; // Then no need for a safept!
2226 break;
2227 } else if (n->Opcode() == Op_SafePoint) {
2228 if (_phase->get_loop(n) == this) {
2229 has_local_ncsfpt = true;
2230 break;
2231 }
2232 if (nonlocal_ncsfpt == NULL) {
2233 nonlocal_ncsfpt = n; // save the one closest to the tail
2234 }
2235 } else {
2236 IdealLoopTree* nlpt = _phase->get_loop(n);
2237 if (this != nlpt) {
2238 // If at an inner loop tail, see if the inner loop has already
2239 // recorded seeing a call on the dom-path (and stop.) If not,
2240 // jump to the head of the inner loop.
2241 assert(is_member(nlpt), "nested loop");
2242 Node* tail = nlpt->_tail;
2243 if (tail->in(0)->is_If()) tail = tail->in(0);
2244 if (n == tail) {
2245 // If inner loop has call on dom-path, so does outer loop
2246 if (nlpt->_has_sfpt) {
2247 has_call = true;
2248 _has_sfpt = 1;
2249 break;
2250 }
2251 // Skip to head of inner loop
2252 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
2253 n = nlpt->_head;
2254 }
2255 }
2256 }
2257 }
2258 // Record safept's that this loop needs preserved when an
2259 // inner loop attempts to delete it's safepoints.
2260 if (_child != NULL && !has_call && !has_local_ncsfpt) {
2261 if (nonlocal_ncsfpt != NULL) {
2262 if (_required_safept == NULL) _required_safept = new Node_List();
2263 _required_safept->push(nonlocal_ncsfpt);
2264 } else {
2265 // Failed to find a suitable safept on the dom-path. Now use
2266 // an all paths walk from tail to head, looking for safepoints to preserve.
2267 allpaths_check_safepts(visited, stack);
2268 }
2269 }
2270 }
2271 }
2272
2273 //---------------------------is_deleteable_safept----------------------------
2274 // Is safept not required by an outer loop?
2275 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
2276 assert(sfpt->Opcode() == Op_SafePoint, "");
2277 IdealLoopTree* lp = get_loop(sfpt)->_parent;
2278 while (lp != NULL) {
2279 Node_List* sfpts = lp->_required_safept;
2280 if (sfpts != NULL) {
2281 for (uint i = 0; i < sfpts->size(); i++) {
2282 if (sfpt == sfpts->at(i))
2283 return false;
2284 }
2285 }
2286 lp = lp->_parent;
2287 }
2288 return true;
2289 }
2290
2291 //---------------------------replace_parallel_iv-------------------------------
2292 // Replace parallel induction variable (parallel to trip counter)
2293 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
2294 assert(loop->_head->is_CountedLoop(), "");
2295 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2296 if (!cl->is_valid_counted_loop())
2297 return; // skip malformed counted loop
2298 Node *incr = cl->incr();
2299 if (incr == NULL)
2300 return; // Dead loop?
2301 Node *init = cl->init_trip();
2302 Node *phi = cl->phi();
2303 int stride_con = cl->stride_con();
2304
2305 // Visit all children, looking for Phis
2306 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
2307 Node *out = cl->out(i);
2308 // Look for other phis (secondary IVs). Skip dead ones
2309 if (!out->is_Phi() || out == phi || !has_node(out))
2310 continue;
2311 PhiNode* phi2 = out->as_Phi();
2312 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
2313 // Look for induction variables of the form: X += constant
2314 if (phi2->region() != loop->_head ||
2315 incr2->req() != 3 ||
2316 incr2->in(1) != phi2 ||
2317 incr2 == incr ||
2318 incr2->Opcode() != Op_AddI ||
2319 !incr2->in(2)->is_Con())
2320 continue;
2321
2322 // Check for parallel induction variable (parallel to trip counter)
2323 // via an affine function. In particular, count-down loops with
2324 // count-up array indices are common. We only RCE references off
2325 // the trip-counter, so we need to convert all these to trip-counter
2326 // expressions.
2327 Node *init2 = phi2->in( LoopNode::EntryControl );
2328 int stride_con2 = incr2->in(2)->get_int();
2329
2330 // The ratio of the two strides cannot be represented as an int
2331 // if stride_con2 is min_int and stride_con is -1.
2332 if (stride_con2 == min_jint && stride_con == -1) {
2333 continue;
2334 }
2335
2336 // The general case here gets a little tricky. We want to find the
2337 // GCD of all possible parallel IV's and make a new IV using this
2338 // GCD for the loop. Then all possible IVs are simple multiples of
2339 // the GCD. In practice, this will cover very few extra loops.
2340 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
2341 // where +/-1 is the common case, but other integer multiples are
2342 // also easy to handle.
2343 int ratio_con = stride_con2/stride_con;
2344
2345 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
2346 #ifndef PRODUCT
2347 if (TraceLoopOpts) {
2348 tty->print("Parallel IV: %d ", phi2->_idx);
2349 loop->dump_head();
2350 }
2351 #endif
2352 // Convert to using the trip counter. The parallel induction
2353 // variable differs from the trip counter by a loop-invariant
2354 // amount, the difference between their respective initial values.
2355 // It is scaled by the 'ratio_con'.
2356 Node* ratio = _igvn.intcon(ratio_con);
2357 set_ctrl(ratio, C->root());
2358 Node* ratio_init = new MulINode(init, ratio);
2359 _igvn.register_new_node_with_optimizer(ratio_init, init);
2360 set_early_ctrl(ratio_init);
2361 Node* diff = new SubINode(init2, ratio_init);
2362 _igvn.register_new_node_with_optimizer(diff, init2);
2363 set_early_ctrl(diff);
2364 Node* ratio_idx = new MulINode(phi, ratio);
2365 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
2366 set_ctrl(ratio_idx, cl);
2367 Node* add = new AddINode(ratio_idx, diff);
2368 _igvn.register_new_node_with_optimizer(add);
2369 set_ctrl(add, cl);
2370 _igvn.replace_node( phi2, add );
2371 // Sometimes an induction variable is unused
2372 if (add->outcnt() == 0) {
2373 _igvn.remove_dead_node(add);
2374 }
2375 --i; // deleted this phi; rescan starting with next position
2376 continue;
2377 }
2378 }
2379 }
2380
2381 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
2382 Node* keep = NULL;
2383 if (keep_one) {
2384 // Look for a safepoint on the idom-path.
2385 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
2386 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
2387 keep = i;
2388 break; // Found one
2389 }
2390 }
2391 }
2392
2393 // Don't remove any safepoints if it is requested to keep a single safepoint and
2394 // no safepoint was found on idom-path. It is not safe to remove any safepoint
2395 // in this case since there's no safepoint dominating all paths in the loop body.
2396 bool prune = !keep_one || keep != NULL;
2397
2398 // Delete other safepoints in this loop.
2399 Node_List* sfpts = _safepts;
2400 if (prune && sfpts != NULL) {
2401 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
2402 for (uint i = 0; i < sfpts->size(); i++) {
2403 Node* n = sfpts->at(i);
2404 assert(phase->get_loop(n) == this, "");
2405 if (n != keep && phase->is_deleteable_safept(n)) {
2406 phase->lazy_replace(n, n->in(TypeFunc::Control));
2407 }
2408 }
2409 }
2410 }
2411
2412 //------------------------------counted_loop-----------------------------------
2413 // Convert to counted loops where possible
2414 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
2415
2416 // For grins, set the inner-loop flag here
2417 if (!_child) {
2418 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
2419 }
2420
2421 IdealLoopTree* loop = this;
2422 if (_head->is_CountedLoop() ||
2423 phase->is_counted_loop(_head, loop)) {
2424
2425 if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) {
2426 // Indicate we do not need a safepoint here
2427 _has_sfpt = 1;
2428 }
2429
2430 // Remove safepoints
2431 bool keep_one_sfpt = !(_has_call || _has_sfpt);
2432 remove_safepoints(phase, keep_one_sfpt);
2433
2434 // Look for induction variables
2435 phase->replace_parallel_iv(this);
2436
2437 } else if (_parent != NULL && !_irreducible) {
2438 // Not a counted loop. Keep one safepoint.
2439 bool keep_one_sfpt = true;
2440 remove_safepoints(phase, keep_one_sfpt);
2441 }
2442
2443 // Recursively
2444 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
2445 assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops");
2446 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
2447 if (loop->_next) loop->_next ->counted_loop(phase);
2448 }
2449
2450 #ifndef PRODUCT
2451 //------------------------------dump_head--------------------------------------
2452 // Dump 1 liner for loop header info
2453 void IdealLoopTree::dump_head( ) const {
2454 for (uint i=0; i<_nest; i++)
2455 tty->print(" ");
2456 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
2457 if (_irreducible) tty->print(" IRREDUCIBLE");
2458 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl);
2459 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
2460 if (predicate != NULL ) {
2461 tty->print(" limit_check");
2462 entry = PhaseIdealLoop::skip_loop_predicates(entry);
2463 }
2464 if (UseLoopPredicate) {
2465 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2466 if (entry != NULL) {
2467 tty->print(" predicated");
2468 entry = PhaseIdealLoop::skip_loop_predicates(entry);
2469 }
2470 }
2471 if (UseProfiledLoopPredicate) {
2472 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
2473 if (entry != NULL) {
2474 tty->print(" profile_predicated");
2475 }
2476 }
2477 if (_head->is_CountedLoop()) {
2478 CountedLoopNode *cl = _head->as_CountedLoop();
2479 tty->print(" counted");
2480
2481 Node* init_n = cl->init_trip();
2482 if (init_n != NULL && init_n->is_Con())
2483 tty->print(" [%d,", cl->init_trip()->get_int());
2484 else
2485 tty->print(" [int,");
2486 Node* limit_n = cl->limit();
2487 if (limit_n != NULL && limit_n->is_Con())
2488 tty->print("%d),", cl->limit()->get_int());
2489 else
2490 tty->print("int),");
2491 int stride_con = cl->stride_con();
2492 if (stride_con > 0) tty->print("+");
2493 tty->print("%d", stride_con);
2494
2495 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
2496
2497 if (cl->is_pre_loop ()) tty->print(" pre" );
2498 if (cl->is_main_loop()) tty->print(" main");
2499 if (cl->is_post_loop()) tty->print(" post");
2500 if (cl->is_vectorized_loop()) tty->print(" vector");
2501 if (cl->range_checks_present()) tty->print(" rc ");
2502 if (cl->is_multiversioned()) tty->print(" multi ");
2503 }
2504 if (_has_call) tty->print(" has_call");
2505 if (_has_sfpt) tty->print(" has_sfpt");
2506 if (_rce_candidate) tty->print(" rce");
2507 if (_safepts != NULL && _safepts->size() > 0) {
2508 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
2509 }
2510 if (_required_safept != NULL && _required_safept->size() > 0) {
2511 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
2512 }
2513 if (Verbose) {
2514 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
2515 }
2516 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
2517 tty->print(" strip_mined");
2518 }
2519 tty->cr();
2520 }
2521
2522 //------------------------------dump-------------------------------------------
2523 // Dump loops by loop tree
2524 void IdealLoopTree::dump( ) const {
2525 dump_head();
2526 if (_child) _child->dump();
2527 if (_next) _next ->dump();
2528 }
2529
2530 #endif
2531
2532 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
2533 if (loop == root) {
2534 if (loop->_child != NULL) {
2535 log->begin_head("loop_tree");
2536 log->end_head();
2537 if( loop->_child ) log_loop_tree(root, loop->_child, log);
2538 log->tail("loop_tree");
2539 assert(loop->_next == NULL, "what?");
2540 }
2541 } else {
2542 Node* head = loop->_head;
2543 log->begin_head("loop");
2544 log->print(" idx='%d' ", head->_idx);
2545 if (loop->_irreducible) log->print("irreducible='1' ");
2546 if (head->is_Loop()) {
2547 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
2548 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
2549 }
2550 if (head->is_CountedLoop()) {
2551 CountedLoopNode* cl = head->as_CountedLoop();
2552 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
2553 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
2554 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
2555 }
2556 log->end_head();
2557 if( loop->_child ) log_loop_tree(root, loop->_child, log);
2558 log->tail("loop");
2559 if( loop->_next ) log_loop_tree(root, loop->_next, log);
2560 }
2561 }
2562
2563 //---------------------collect_potentially_useful_predicates-----------------------
2564 // Helper function to collect potentially useful predicates to prevent them from
2565 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
2566 void PhaseIdealLoop::collect_potentially_useful_predicates(
2567 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
2568 if (loop->_child) { // child
2569 collect_potentially_useful_predicates(loop->_child, useful_predicates);
2570 }
2571
2572 // self (only loops that we can apply loop predication may use their predicates)
2573 if (loop->_head->is_Loop() &&
2574 !loop->_irreducible &&
2575 !loop->tail()->is_top()) {
2576 LoopNode* lpn = loop->_head->as_Loop();
2577 Node* entry = lpn->in(LoopNode::EntryControl);
2578 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2579 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
2580 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2581 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2582 entry = skip_loop_predicates(entry);
2583 }
2584 predicate_proj = find_predicate(entry); // Predicate
2585 if (predicate_proj != NULL ) {
2586 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2587 entry = skip_loop_predicates(entry);
2588 }
2589 if (UseProfiledLoopPredicate) {
2590 predicate_proj = find_predicate(entry); // Predicate
2591 if (predicate_proj != NULL ) {
2592 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2593 }
2594 }
2595 }
2596
2597 if (loop->_next) { // sibling
2598 collect_potentially_useful_predicates(loop->_next, useful_predicates);
2599 }
2600 }
2601
2602 //------------------------eliminate_useless_predicates-----------------------------
2603 // Eliminate all inserted predicates if they could not be used by loop predication.
2604 // Note: it will also eliminates loop limits check predicate since it also uses
2605 // Opaque1 node (see Parse::add_predicate()).
2606 void PhaseIdealLoop::eliminate_useless_predicates() {
2607 if (C->predicate_count() == 0)
2608 return; // no predicate left
2609
2610 Unique_Node_List useful_predicates; // to store useful predicates
2611 if (C->has_loops()) {
2612 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2613 }
2614
2615 for (int i = C->predicate_count(); i > 0; i--) {
2616 Node * n = C->predicate_opaque1_node(i-1);
2617 assert(n->Opcode() == Op_Opaque1, "must be");
2618 if (!useful_predicates.member(n)) { // not in the useful list
2619 _igvn.replace_node(n, n->in(1));
2620 }
2621 }
2622 }
2623
2624 //------------------------process_expensive_nodes-----------------------------
2625 // Expensive nodes have their control input set to prevent the GVN
2626 // from commoning them and as a result forcing the resulting node to
2627 // be in a more frequent path. Use CFG information here, to change the
2628 // control inputs so that some expensive nodes can be commoned while
2629 // not executed more frequently.
2630 bool PhaseIdealLoop::process_expensive_nodes() {
2631 assert(OptimizeExpensiveOps, "optimization off?");
2632
2633 // Sort nodes to bring similar nodes together
2634 C->sort_expensive_nodes();
2635
2636 bool progress = false;
2637
2638 for (int i = 0; i < C->expensive_count(); ) {
2639 Node* n = C->expensive_node(i);
2640 int start = i;
2641 // Find nodes similar to n
2642 i++;
2643 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2644 int end = i;
2645 // And compare them two by two
2646 for (int j = start; j < end; j++) {
2647 Node* n1 = C->expensive_node(j);
2648 if (is_node_unreachable(n1)) {
2649 continue;
2650 }
2651 for (int k = j+1; k < end; k++) {
2652 Node* n2 = C->expensive_node(k);
2653 if (is_node_unreachable(n2)) {
2654 continue;
2655 }
2656
2657 assert(n1 != n2, "should be pair of nodes");
2658
2659 Node* c1 = n1->in(0);
2660 Node* c2 = n2->in(0);
2661
2662 Node* parent_c1 = c1;
2663 Node* parent_c2 = c2;
2664
2665 // The call to get_early_ctrl_for_expensive() moves the
2666 // expensive nodes up but stops at loops that are in a if
2667 // branch. See whether we can exit the loop and move above the
2668 // If.
2669 if (c1->is_Loop()) {
2670 parent_c1 = c1->in(1);
2671 }
2672 if (c2->is_Loop()) {
2673 parent_c2 = c2->in(1);
2674 }
2675
2676 if (parent_c1 == parent_c2) {
2677 _igvn._worklist.push(n1);
2678 _igvn._worklist.push(n2);
2679 continue;
2680 }
2681
2682 // Look for identical expensive node up the dominator chain.
2683 if (is_dominator(c1, c2)) {
2684 c2 = c1;
2685 } else if (is_dominator(c2, c1)) {
2686 c1 = c2;
2687 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2688 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2689 // Both branches have the same expensive node so move it up
2690 // before the if.
2691 c1 = c2 = idom(parent_c1->in(0));
2692 }
2693 // Do the actual moves
2694 if (n1->in(0) != c1) {
2695 _igvn.hash_delete(n1);
2696 n1->set_req(0, c1);
2697 _igvn.hash_insert(n1);
2698 _igvn._worklist.push(n1);
2699 progress = true;
2700 }
2701 if (n2->in(0) != c2) {
2702 _igvn.hash_delete(n2);
2703 n2->set_req(0, c2);
2704 _igvn.hash_insert(n2);
2705 _igvn._worklist.push(n2);
2706 progress = true;
2707 }
2708 }
2709 }
2710 }
2711
2712 return progress;
2713 }
2714
2715
2716 //=============================================================================
2717 //----------------------------build_and_optimize-------------------------------
2718 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2719 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2720 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
2721 bool do_split_ifs = (mode == LoopOptsDefault || mode == LoopOptsZgcLastRound);
2722 bool skip_loop_opts = (mode == LoopOptsNone) ;
2723 bool shenandoah_opts = (mode == LoopOptsShenandoahExpand ||
2724 mode == LoopOptsShenandoahPostExpand);
2725
2726 ResourceMark rm;
2727
2728 int old_progress = C->major_progress();
2729 uint orig_worklist_size = _igvn._worklist.size();
2730
2731 // Reset major-progress flag for the driver's heuristics
2732 C->clear_major_progress();
2733
2734 #ifndef PRODUCT
2735 // Capture for later assert
2736 uint unique = C->unique();
2737 _loop_invokes++;
2738 _loop_work += unique;
2739 #endif
2740
2741 // True if the method has at least 1 irreducible loop
2742 _has_irreducible_loops = false;
2743
2744 _created_loop_node = false;
2745
2746 Arena *a = Thread::current()->resource_area();
2747 VectorSet visited(a);
2748 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2749 _nodes.map(C->unique(), NULL);
2750 memset(_nodes.adr(), 0, wordSize * C->unique());
2751
2752 // Pre-build the top-level outermost loop tree entry
2753 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2754 // Do not need a safepoint at the top level
2755 _ltree_root->_has_sfpt = 1;
2756
2757 // Initialize Dominators.
2758 // Checked in clone_loop_predicate() during beautify_loops().
2759 _idom_size = 0;
2760 _idom = NULL;
2761 _dom_depth = NULL;
2762 _dom_stk = NULL;
2763
2764 // Empty pre-order array
2765 allocate_preorders();
2766
2767 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2768 // IdealLoopTree entries. Data nodes are NOT walked.
2769 build_loop_tree();
2770 // Check for bailout, and return
2771 if (C->failing()) {
2772 return;
2773 }
2774
2775 // No loops after all
2776 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2777
2778 // There should always be an outer loop containing the Root and Return nodes.
2779 // If not, we have a degenerate empty program. Bail out in this case.
2780 if (!has_node(C->root())) {
2781 if (!_verify_only) {
2782 C->clear_major_progress();
2783 C->record_method_not_compilable("empty program detected during loop optimization");
2784 }
2785 return;
2786 }
2787
2788 // Nothing to do, so get out
2789 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only && !shenandoah_opts;
2790 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2791 if (stop_early && !do_expensive_nodes) {
2792 _igvn.optimize(); // Cleanup NeverBranches
2793 return;
2794 }
2795
2796 // Set loop nesting depth
2797 _ltree_root->set_nest( 0 );
2798
2799 // Split shared headers and insert loop landing pads.
2800 // Do not bother doing this on the Root loop of course.
2801 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2802 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2803 if( _ltree_root->_child->beautify_loops( this ) ) {
2804 // Re-build loop tree!
2805 _ltree_root->_child = NULL;
2806 _nodes.clear();
2807 reallocate_preorders();
2808 build_loop_tree();
2809 // Check for bailout, and return
2810 if (C->failing()) {
2811 return;
2812 }
2813 // Reset loop nesting depth
2814 _ltree_root->set_nest( 0 );
2815
2816 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2817 }
2818 }
2819
2820 // Build Dominators for elision of NULL checks & loop finding.
2821 // Since nodes do not have a slot for immediate dominator, make
2822 // a persistent side array for that info indexed on node->_idx.
2823 _idom_size = C->unique();
2824 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2825 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2826 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2827 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2828
2829 Dominators();
2830
2831 if (!_verify_only) {
2832 // As a side effect, Dominators removed any unreachable CFG paths
2833 // into RegionNodes. It doesn't do this test against Root, so
2834 // we do it here.
2835 for( uint i = 1; i < C->root()->req(); i++ ) {
2836 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2837 _igvn.delete_input_of(C->root(), i);
2838 i--; // Rerun same iteration on compressed edges
2839 }
2840 }
2841
2842 // Given dominators, try to find inner loops with calls that must
2843 // always be executed (call dominates loop tail). These loops do
2844 // not need a separate safepoint.
2845 Node_List cisstack(a);
2846 _ltree_root->check_safepts(visited, cisstack);
2847 }
2848
2849 // Walk the DATA nodes and place into loops. Find earliest control
2850 // node. For CFG nodes, the _nodes array starts out and remains
2851 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2852 // _nodes array holds the earliest legal controlling CFG node.
2853
2854 // Allocate stack with enough space to avoid frequent realloc
2855 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2856 Node_Stack nstack( a, stack_size );
2857
2858 visited.Clear();
2859 Node_List worklist(a);
2860 // Don't need C->root() on worklist since
2861 // it will be processed among C->top() inputs
2862 worklist.push( C->top() );
2863 visited.set( C->top()->_idx ); // Set C->top() as visited now
2864 build_loop_early( visited, worklist, nstack );
2865
2866 // Given early legal placement, try finding counted loops. This placement
2867 // is good enough to discover most loop invariants.
2868 if (!_verify_me && !_verify_only && !shenandoah_opts) {
2869 _ltree_root->counted_loop(this);
2870 }
2871
2872 // Find latest loop placement. Find ideal loop placement.
2873 visited.Clear();
2874 init_dom_lca_tags();
2875 // Need C->root() on worklist when processing outs
2876 worklist.push( C->root() );
2877 NOT_PRODUCT( C->verify_graph_edges(); )
2878 worklist.push( C->top() );
2879 build_loop_late( visited, worklist, nstack );
2880
2881 if (_verify_only) {
2882 // restore major progress flag
2883 for (int i = 0; i < old_progress; i++)
2884 C->set_major_progress();
2885 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2886 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2887 return;
2888 }
2889
2890 // clear out the dead code after build_loop_late
2891 while (_deadlist.size()) {
2892 _igvn.remove_globally_dead_node(_deadlist.pop());
2893 }
2894
2895 if (stop_early) {
2896 assert(do_expensive_nodes, "why are we here?");
2897 if (process_expensive_nodes()) {
2898 // If we made some progress when processing expensive nodes then
2899 // the IGVN may modify the graph in a way that will allow us to
2900 // make some more progress: we need to try processing expensive
2901 // nodes again.
2902 C->set_major_progress();
2903 }
2904 _igvn.optimize();
2905 return;
2906 }
2907
2908 // Some parser-inserted loop predicates could never be used by loop
2909 // predication or they were moved away from loop during some optimizations.
2910 // For example, peeling. Eliminate them before next loop optimizations.
2911 eliminate_useless_predicates();
2912
2913 #ifndef PRODUCT
2914 C->verify_graph_edges();
2915 if (_verify_me) { // Nested verify pass?
2916 // Check to see if the verify mode is broken
2917 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2918 return;
2919 }
2920 if(VerifyLoopOptimizations) verify();
2921 if(TraceLoopOpts && C->has_loops()) {
2922 _ltree_root->dump();
2923 }
2924 #endif
2925
2926 if (skip_loop_opts) {
2927 // restore major progress flag
2928 for (int i = 0; i < old_progress; i++) {
2929 C->set_major_progress();
2930 }
2931
2932 // Cleanup any modified bits
2933 _igvn.optimize();
2934
2935 if (C->log() != NULL) {
2936 log_loop_tree(_ltree_root, _ltree_root, C->log());
2937 }
2938 return;
2939 }
2940
2941 #if INCLUDE_SHENANDOAHGC
2942 if (UseShenandoahGC && ((ShenandoahBarrierSetC2*) BarrierSet::barrier_set()->barrier_set_c2())->optimize_loops(this, mode, visited, nstack, worklist)) {
2943 _igvn.optimize();
2944 if (C->log() != NULL) {
2945 log_loop_tree(_ltree_root, _ltree_root, C->log());
2946 }
2947 return;
2948 }
2949 #endif
2950
2951 if (ReassociateInvariants) {
2952 // Reassociate invariants and prep for split_thru_phi
2953 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2954 IdealLoopTree* lpt = iter.current();
2955 bool is_counted = lpt->is_counted();
2956 if (!is_counted || !lpt->is_inner()) continue;
2957
2958 // check for vectorized loops, any reassociation of invariants was already done
2959 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue;
2960
2961 lpt->reassociate_invariants(this);
2962
2963 // Because RCE opportunities can be masked by split_thru_phi,
2964 // look for RCE candidates and inhibit split_thru_phi
2965 // on just their loop-phi's for this pass of loop opts
2966 if (SplitIfBlocks && do_split_ifs) {
2967 if (lpt->policy_range_check(this)) {
2968 lpt->_rce_candidate = 1; // = true
2969 }
2970 }
2971 }
2972 }
2973
2974 // Check for aggressive application of split-if and other transforms
2975 // that require basic-block info (like cloning through Phi's)
2976 if( SplitIfBlocks && do_split_ifs ) {
2977 visited.Clear();
2978 split_if_with_blocks( visited, nstack, mode == LoopOptsZgcLastRound );
2979 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2980 if (mode == LoopOptsZgcLastRound) {
2981 C->set_major_progress();
2982 }
2983 }
2984
2985 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2986 C->set_major_progress();
2987 }
2988
2989 // Perform loop predication before iteration splitting
2990 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2991 _ltree_root->_child->loop_predication(this);
2992 }
2993
2994 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2995 if (do_intrinsify_fill()) {
2996 C->set_major_progress();
2997 }
2998 }
2999
3000 // Perform iteration-splitting on inner loops. Split iterations to avoid
3001 // range checks or one-shot null checks.
3002
3003 // If split-if's didn't hack the graph too bad (no CFG changes)
3004 // then do loop opts.
3005 if (C->has_loops() && !C->major_progress()) {
3006 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
3007 _ltree_root->_child->iteration_split( this, worklist );
3008 // No verify after peeling! GCM has hoisted code out of the loop.
3009 // After peeling, the hoisted code could sink inside the peeled area.
3010 // The peeling code does not try to recompute the best location for
3011 // all the code before the peeled area, so the verify pass will always
3012 // complain about it.
3013 }
3014 // Do verify graph edges in any case
3015 NOT_PRODUCT( C->verify_graph_edges(); );
3016
3017 if (!do_split_ifs) {
3018 // We saw major progress in Split-If to get here. We forced a
3019 // pass with unrolling and not split-if, however more split-if's
3020 // might make progress. If the unrolling didn't make progress
3021 // then the major-progress flag got cleared and we won't try
3022 // another round of Split-If. In particular the ever-common
3023 // instance-of/check-cast pattern requires at least 2 rounds of
3024 // Split-If to clear out.
3025 C->set_major_progress();
3026 }
3027
3028 // Repeat loop optimizations if new loops were seen
3029 if (created_loop_node()) {
3030 C->set_major_progress();
3031 }
3032
3033 // Keep loop predicates and perform optimizations with them
3034 // until no more loop optimizations could be done.
3035 // After that switch predicates off and do more loop optimizations.
3036 if (!C->major_progress() && (C->predicate_count() > 0)) {
3037 C->cleanup_loop_predicates(_igvn);
3038 if (TraceLoopOpts) {
3039 tty->print_cr("PredicatesOff");
3040 }
3041 C->set_major_progress();
3042 }
3043
3044 // Convert scalar to superword operations at the end of all loop opts.
3045 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
3046 // SuperWord transform
3047 SuperWord sw(this);
3048 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3049 IdealLoopTree* lpt = iter.current();
3050 if (lpt->is_counted()) {
3051 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
3052
3053 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
3054 // Check that the rce'd post loop is encountered first, multiversion after all
3055 // major main loop optimization are concluded
3056 if (!C->major_progress()) {
3057 IdealLoopTree *lpt_next = lpt->_next;
3058 if (lpt_next && lpt_next->is_counted()) {
3059 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
3060 has_range_checks(lpt_next);
3061 if (cl->is_post_loop() && cl->range_checks_present()) {
3062 if (!cl->is_multiversioned()) {
3063 if (multi_version_post_loops(lpt, lpt_next) == false) {
3064 // Cause the rce loop to be optimized away if we fail
3065 cl->mark_is_multiversioned();
3066 cl->set_slp_max_unroll(0);
3067 poison_rce_post_loop(lpt);
3068 }
3069 }
3070 }
3071 }
3072 sw.transform_loop(lpt, true);
3073 }
3074 } else if (cl->is_main_loop()) {
3075 sw.transform_loop(lpt, true);
3076 }
3077 }
3078 }
3079 }
3080
3081 // Cleanup any modified bits
3082 _igvn.optimize();
3083
3084 // disable assert until issue with split_flow_path is resolved (6742111)
3085 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
3086 // "shouldn't introduce irreducible loops");
3087
3088 if (C->log() != NULL) {
3089 log_loop_tree(_ltree_root, _ltree_root, C->log());
3090 }
3091 }
3092
3093 #ifndef PRODUCT
3094 //------------------------------print_statistics-------------------------------
3095 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
3096 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
3097 void PhaseIdealLoop::print_statistics() {
3098 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
3099 }
3100
3101 //------------------------------verify-----------------------------------------
3102 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
3103 static int fail; // debug only, so its multi-thread dont care
3104 void PhaseIdealLoop::verify() const {
3105 int old_progress = C->major_progress();
3106 ResourceMark rm;
3107 PhaseIdealLoop loop_verify( _igvn, this );
3108 VectorSet visited(Thread::current()->resource_area());
3109
3110 fail = 0;
3111 verify_compare( C->root(), &loop_verify, visited );
3112 assert( fail == 0, "verify loops failed" );
3113 // Verify loop structure is the same
3114 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
3115 // Reset major-progress. It was cleared by creating a verify version of
3116 // PhaseIdealLoop.
3117 for( int i=0; i<old_progress; i++ )
3118 C->set_major_progress();
3119 }
3120
3121 //------------------------------verify_compare---------------------------------
3122 // Make sure me and the given PhaseIdealLoop agree on key data structures
3123 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
3124 if( !n ) return;
3125 if( visited.test_set( n->_idx ) ) return;
3126 if( !_nodes[n->_idx] ) { // Unreachable
3127 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
3128 return;
3129 }
3130
3131 uint i;
3132 for( i = 0; i < n->req(); i++ )
3133 verify_compare( n->in(i), loop_verify, visited );
3134
3135 // Check the '_nodes' block/loop structure
3136 i = n->_idx;
3137 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
3138 if( _nodes[i] != loop_verify->_nodes[i] &&
3139 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
3140 tty->print("Mismatched control setting for: ");
3141 n->dump();
3142 if( fail++ > 10 ) return;
3143 Node *c = get_ctrl_no_update(n);
3144 tty->print("We have it as: ");
3145 if( c->in(0) ) c->dump();
3146 else tty->print_cr("N%d",c->_idx);
3147 tty->print("Verify thinks: ");
3148 if( loop_verify->has_ctrl(n) )
3149 loop_verify->get_ctrl_no_update(n)->dump();
3150 else
3151 loop_verify->get_loop_idx(n)->dump();
3152 tty->cr();
3153 }
3154 } else { // We have a loop
3155 IdealLoopTree *us = get_loop_idx(n);
3156 if( loop_verify->has_ctrl(n) ) {
3157 tty->print("Mismatched loop setting for: ");
3158 n->dump();
3159 if( fail++ > 10 ) return;
3160 tty->print("We have it as: ");
3161 us->dump();
3162 tty->print("Verify thinks: ");
3163 loop_verify->get_ctrl_no_update(n)->dump();
3164 tty->cr();
3165 } else if (!C->major_progress()) {
3166 // Loop selection can be messed up if we did a major progress
3167 // operation, like split-if. Do not verify in that case.
3168 IdealLoopTree *them = loop_verify->get_loop_idx(n);
3169 if( us->_head != them->_head || us->_tail != them->_tail ) {
3170 tty->print("Unequals loops for: ");
3171 n->dump();
3172 if( fail++ > 10 ) return;
3173 tty->print("We have it as: ");
3174 us->dump();
3175 tty->print("Verify thinks: ");
3176 them->dump();
3177 tty->cr();
3178 }
3179 }
3180 }
3181
3182 // Check for immediate dominators being equal
3183 if( i >= _idom_size ) {
3184 if( !n->is_CFG() ) return;
3185 tty->print("CFG Node with no idom: ");
3186 n->dump();
3187 return;
3188 }
3189 if( !n->is_CFG() ) return;
3190 if( n == C->root() ) return; // No IDOM here
3191
3192 assert(n->_idx == i, "sanity");
3193 Node *id = idom_no_update(n);
3194 if( id != loop_verify->idom_no_update(n) ) {
3195 tty->print("Unequals idoms for: ");
3196 n->dump();
3197 if( fail++ > 10 ) return;
3198 tty->print("We have it as: ");
3199 id->dump();
3200 tty->print("Verify thinks: ");
3201 loop_verify->idom_no_update(n)->dump();
3202 tty->cr();
3203 }
3204
3205 }
3206
3207 //------------------------------verify_tree------------------------------------
3208 // Verify that tree structures match. Because the CFG can change, siblings
3209 // within the loop tree can be reordered. We attempt to deal with that by
3210 // reordering the verify's loop tree if possible.
3211 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
3212 assert( _parent == parent, "Badly formed loop tree" );
3213
3214 // Siblings not in same order? Attempt to re-order.
3215 if( _head != loop->_head ) {
3216 // Find _next pointer to update
3217 IdealLoopTree **pp = &loop->_parent->_child;
3218 while( *pp != loop )
3219 pp = &((*pp)->_next);
3220 // Find proper sibling to be next
3221 IdealLoopTree **nn = &loop->_next;
3222 while( (*nn) && (*nn)->_head != _head )
3223 nn = &((*nn)->_next);
3224
3225 // Check for no match.
3226 if( !(*nn) ) {
3227 // Annoyingly, irreducible loops can pick different headers
3228 // after a major_progress operation, so the rest of the loop
3229 // tree cannot be matched.
3230 if (_irreducible && Compile::current()->major_progress()) return;
3231 assert( 0, "failed to match loop tree" );
3232 }
3233
3234 // Move (*nn) to (*pp)
3235 IdealLoopTree *hit = *nn;
3236 *nn = hit->_next;
3237 hit->_next = loop;
3238 *pp = loop;
3239 loop = hit;
3240 // Now try again to verify
3241 }
3242
3243 assert( _head == loop->_head , "mismatched loop head" );
3244 Node *tail = _tail; // Inline a non-updating version of
3245 while( !tail->in(0) ) // the 'tail()' call.
3246 tail = tail->in(1);
3247 assert( tail == loop->_tail, "mismatched loop tail" );
3248
3249 // Counted loops that are guarded should be able to find their guards
3250 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
3251 CountedLoopNode *cl = _head->as_CountedLoop();
3252 Node *init = cl->init_trip();
3253 Node *ctrl = cl->in(LoopNode::EntryControl);
3254 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
3255 Node *iff = ctrl->in(0);
3256 assert( iff->Opcode() == Op_If, "" );
3257 Node *bol = iff->in(1);
3258 assert( bol->Opcode() == Op_Bool, "" );
3259 Node *cmp = bol->in(1);
3260 assert( cmp->Opcode() == Op_CmpI, "" );
3261 Node *add = cmp->in(1);
3262 Node *opaq;
3263 if( add->Opcode() == Op_Opaque1 ) {
3264 opaq = add;
3265 } else {
3266 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
3267 assert( add == init, "" );
3268 opaq = cmp->in(2);
3269 }
3270 assert( opaq->Opcode() == Op_Opaque1, "" );
3271
3272 }
3273
3274 if (_child != NULL) _child->verify_tree(loop->_child, this);
3275 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
3276 // Innermost loops need to verify loop bodies,
3277 // but only if no 'major_progress'
3278 int fail = 0;
3279 if (!Compile::current()->major_progress() && _child == NULL) {
3280 for( uint i = 0; i < _body.size(); i++ ) {
3281 Node *n = _body.at(i);
3282 if (n->outcnt() == 0) continue; // Ignore dead
3283 uint j;
3284 for( j = 0; j < loop->_body.size(); j++ )
3285 if( loop->_body.at(j) == n )
3286 break;
3287 if( j == loop->_body.size() ) { // Not found in loop body
3288 // Last ditch effort to avoid assertion: Its possible that we
3289 // have some users (so outcnt not zero) but are still dead.
3290 // Try to find from root.
3291 if (Compile::current()->root()->find(n->_idx)) {
3292 fail++;
3293 tty->print("We have that verify does not: ");
3294 n->dump();
3295 }
3296 }
3297 }
3298 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
3299 Node *n = loop->_body.at(i2);
3300 if (n->outcnt() == 0) continue; // Ignore dead
3301 uint j;
3302 for( j = 0; j < _body.size(); j++ )
3303 if( _body.at(j) == n )
3304 break;
3305 if( j == _body.size() ) { // Not found in loop body
3306 // Last ditch effort to avoid assertion: Its possible that we
3307 // have some users (so outcnt not zero) but are still dead.
3308 // Try to find from root.
3309 if (Compile::current()->root()->find(n->_idx)) {
3310 fail++;
3311 tty->print("Verify has that we do not: ");
3312 n->dump();
3313 }
3314 }
3315 }
3316 assert( !fail, "loop body mismatch" );
3317 }
3318 }
3319
3320 #endif
3321
3322 //------------------------------set_idom---------------------------------------
3323 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
3324 uint idx = d->_idx;
3325 if (idx >= _idom_size) {
3326 uint newsize = _idom_size<<1;
3327 while( idx >= newsize ) {
3328 newsize <<= 1;
3329 }
3330 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
3331 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
3332 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
3333 _idom_size = newsize;
3334 }
3335 _idom[idx] = n;
3336 _dom_depth[idx] = dom_depth;
3337 }
3338
3339 //------------------------------recompute_dom_depth---------------------------------------
3340 // The dominator tree is constructed with only parent pointers.
3341 // This recomputes the depth in the tree by first tagging all
3342 // nodes as "no depth yet" marker. The next pass then runs up
3343 // the dom tree from each node marked "no depth yet", and computes
3344 // the depth on the way back down.
3345 void PhaseIdealLoop::recompute_dom_depth() {
3346 uint no_depth_marker = C->unique();
3347 uint i;
3348 // Initialize depth to "no depth yet" and realize all lazy updates
3349 for (i = 0; i < _idom_size; i++) {
3350 // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized).
3351 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
3352 _dom_depth[i] = no_depth_marker;
3353
3354 // heal _idom if it has a fwd mapping in _nodes
3355 if (_idom[i]->in(0) == NULL) {
3356 idom(i);
3357 }
3358 }
3359 }
3360 if (_dom_stk == NULL) {
3361 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
3362 if (init_size < 10) init_size = 10;
3363 _dom_stk = new GrowableArray<uint>(init_size);
3364 }
3365 // Compute new depth for each node.
3366 for (i = 0; i < _idom_size; i++) {
3367 uint j = i;
3368 // Run up the dom tree to find a node with a depth
3369 while (_dom_depth[j] == no_depth_marker) {
3370 _dom_stk->push(j);
3371 j = _idom[j]->_idx;
3372 }
3373 // Compute the depth on the way back down this tree branch
3374 uint dd = _dom_depth[j] + 1;
3375 while (_dom_stk->length() > 0) {
3376 uint j = _dom_stk->pop();
3377 _dom_depth[j] = dd;
3378 dd++;
3379 }
3380 }
3381 }
3382
3383 //------------------------------sort-------------------------------------------
3384 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
3385 // loop tree, not the root.
3386 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
3387 if( !innermost ) return loop; // New innermost loop
3388
3389 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
3390 assert( loop_preorder, "not yet post-walked loop" );
3391 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
3392 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
3393
3394 // Insert at start of list
3395 while( l ) { // Insertion sort based on pre-order
3396 if( l == loop ) return innermost; // Already on list!
3397 int l_preorder = get_preorder(l->_head); // Cache pre-order number
3398 assert( l_preorder, "not yet post-walked l" );
3399 // Check header pre-order number to figure proper nesting
3400 if( loop_preorder > l_preorder )
3401 break; // End of insertion
3402 // If headers tie (e.g., shared headers) check tail pre-order numbers.
3403 // Since I split shared headers, you'd think this could not happen.
3404 // BUT: I must first do the preorder numbering before I can discover I
3405 // have shared headers, so the split headers all get the same preorder
3406 // number as the RegionNode they split from.
3407 if( loop_preorder == l_preorder &&
3408 get_preorder(loop->_tail) < get_preorder(l->_tail) )
3409 break; // Also check for shared headers (same pre#)
3410 pp = &l->_parent; // Chain up list
3411 l = *pp;
3412 }
3413 // Link into list
3414 // Point predecessor to me
3415 *pp = loop;
3416 // Point me to successor
3417 IdealLoopTree *p = loop->_parent;
3418 loop->_parent = l; // Point me to successor
3419 if( p ) sort( p, innermost ); // Insert my parents into list as well
3420 return innermost;
3421 }
3422
3423 //------------------------------build_loop_tree--------------------------------
3424 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
3425 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
3426 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
3427 // tightest enclosing IdealLoopTree for post-walked.
3428 //
3429 // During my forward walk I do a short 1-layer lookahead to see if I can find
3430 // a loop backedge with that doesn't have any work on the backedge. This
3431 // helps me construct nested loops with shared headers better.
3432 //
3433 // Once I've done the forward recursion, I do the post-work. For each child
3434 // I check to see if there is a backedge. Backedges define a loop! I
3435 // insert an IdealLoopTree at the target of the backedge.
3436 //
3437 // During the post-work I also check to see if I have several children
3438 // belonging to different loops. If so, then this Node is a decision point
3439 // where control flow can choose to change loop nests. It is at this
3440 // decision point where I can figure out how loops are nested. At this
3441 // time I can properly order the different loop nests from my children.
3442 // Note that there may not be any backedges at the decision point!
3443 //
3444 // Since the decision point can be far removed from the backedges, I can't
3445 // order my loops at the time I discover them. Thus at the decision point
3446 // I need to inspect loop header pre-order numbers to properly nest my
3447 // loops. This means I need to sort my childrens' loops by pre-order.
3448 // The sort is of size number-of-control-children, which generally limits
3449 // it to size 2 (i.e., I just choose between my 2 target loops).
3450 void PhaseIdealLoop::build_loop_tree() {
3451 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3452 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
3453 Node *n = C->root();
3454 bltstack.push(n);
3455 int pre_order = 1;
3456 int stack_size;
3457
3458 while ( ( stack_size = bltstack.length() ) != 0 ) {
3459 n = bltstack.top(); // Leave node on stack
3460 if ( !is_visited(n) ) {
3461 // ---- Pre-pass Work ----
3462 // Pre-walked but not post-walked nodes need a pre_order number.
3463
3464 set_preorder_visited( n, pre_order ); // set as visited
3465
3466 // ---- Scan over children ----
3467 // Scan first over control projections that lead to loop headers.
3468 // This helps us find inner-to-outer loops with shared headers better.
3469
3470 // Scan children's children for loop headers.
3471 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
3472 Node* m = n->raw_out(i); // Child
3473 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
3474 // Scan over children's children to find loop
3475 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3476 Node* l = m->fast_out(j);
3477 if( is_visited(l) && // Been visited?
3478 !is_postvisited(l) && // But not post-visited
3479 get_preorder(l) < pre_order ) { // And smaller pre-order
3480 // Found! Scan the DFS down this path before doing other paths
3481 bltstack.push(m);
3482 break;
3483 }
3484 }
3485 }
3486 }
3487 pre_order++;
3488 }
3489 else if ( !is_postvisited(n) ) {
3490 // Note: build_loop_tree_impl() adds out edges on rare occasions,
3491 // such as com.sun.rsasign.am::a.
3492 // For non-recursive version, first, process current children.
3493 // On next iteration, check if additional children were added.
3494 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
3495 Node* u = n->raw_out(k);
3496 if ( u->is_CFG() && !is_visited(u) ) {
3497 bltstack.push(u);
3498 }
3499 }
3500 if ( bltstack.length() == stack_size ) {
3501 // There were no additional children, post visit node now
3502 (void)bltstack.pop(); // Remove node from stack
3503 pre_order = build_loop_tree_impl( n, pre_order );
3504 // Check for bailout
3505 if (C->failing()) {
3506 return;
3507 }
3508 // Check to grow _preorders[] array for the case when
3509 // build_loop_tree_impl() adds new nodes.
3510 check_grow_preorders();
3511 }
3512 }
3513 else {
3514 (void)bltstack.pop(); // Remove post-visited node from stack
3515 }
3516 }
3517 }
3518
3519 //------------------------------build_loop_tree_impl---------------------------
3520 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
3521 // ---- Post-pass Work ----
3522 // Pre-walked but not post-walked nodes need a pre_order number.
3523
3524 // Tightest enclosing loop for this Node
3525 IdealLoopTree *innermost = NULL;
3526
3527 // For all children, see if any edge is a backedge. If so, make a loop
3528 // for it. Then find the tightest enclosing loop for the self Node.
3529 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3530 Node* m = n->fast_out(i); // Child
3531 if( n == m ) continue; // Ignore control self-cycles
3532 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
3533
3534 IdealLoopTree *l; // Child's loop
3535 if( !is_postvisited(m) ) { // Child visited but not post-visited?
3536 // Found a backedge
3537 assert( get_preorder(m) < pre_order, "should be backedge" );
3538 // Check for the RootNode, which is already a LoopNode and is allowed
3539 // to have multiple "backedges".
3540 if( m == C->root()) { // Found the root?
3541 l = _ltree_root; // Root is the outermost LoopNode
3542 } else { // Else found a nested loop
3543 // Insert a LoopNode to mark this loop.
3544 l = new IdealLoopTree(this, m, n);
3545 } // End of Else found a nested loop
3546 if( !has_loop(m) ) // If 'm' does not already have a loop set
3547 set_loop(m, l); // Set loop header to loop now
3548
3549 } else { // Else not a nested loop
3550 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
3551 l = get_loop(m); // Get previously determined loop
3552 // If successor is header of a loop (nest), move up-loop till it
3553 // is a member of some outer enclosing loop. Since there are no
3554 // shared headers (I've split them already) I only need to go up
3555 // at most 1 level.
3556 while( l && l->_head == m ) // Successor heads loop?
3557 l = l->_parent; // Move up 1 for me
3558 // If this loop is not properly parented, then this loop
3559 // has no exit path out, i.e. its an infinite loop.
3560 if( !l ) {
3561 // Make loop "reachable" from root so the CFG is reachable. Basically
3562 // insert a bogus loop exit that is never taken. 'm', the loop head,
3563 // points to 'n', one (of possibly many) fall-in paths. There may be
3564 // many backedges as well.
3565
3566 // Here I set the loop to be the root loop. I could have, after
3567 // inserting a bogus loop exit, restarted the recursion and found my
3568 // new loop exit. This would make the infinite loop a first-class
3569 // loop and it would then get properly optimized. What's the use of
3570 // optimizing an infinite loop?
3571 l = _ltree_root; // Oops, found infinite loop
3572
3573 if (!_verify_only) {
3574 // Insert the NeverBranch between 'm' and it's control user.
3575 NeverBranchNode *iff = new NeverBranchNode( m );
3576 _igvn.register_new_node_with_optimizer(iff);
3577 set_loop(iff, l);
3578 Node *if_t = new CProjNode( iff, 0 );
3579 _igvn.register_new_node_with_optimizer(if_t);
3580 set_loop(if_t, l);
3581
3582 Node* cfg = NULL; // Find the One True Control User of m
3583 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3584 Node* x = m->fast_out(j);
3585 if (x->is_CFG() && x != m && x != iff)
3586 { cfg = x; break; }
3587 }
3588 assert(cfg != NULL, "must find the control user of m");
3589 uint k = 0; // Probably cfg->in(0)
3590 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
3591 cfg->set_req( k, if_t ); // Now point to NeverBranch
3592 _igvn._worklist.push(cfg);
3593
3594 // Now create the never-taken loop exit
3595 Node *if_f = new CProjNode( iff, 1 );
3596 _igvn.register_new_node_with_optimizer(if_f);
3597 set_loop(if_f, l);
3598 // Find frame ptr for Halt. Relies on the optimizer
3599 // V-N'ing. Easier and quicker than searching through
3600 // the program structure.
3601 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
3602 _igvn.register_new_node_with_optimizer(frame);
3603 // Halt & Catch Fire
3604 Node *halt = new HaltNode( if_f, frame );
3605 _igvn.register_new_node_with_optimizer(halt);
3606 set_loop(halt, l);
3607 C->root()->add_req(halt);
3608 }
3609 set_loop(C->root(), _ltree_root);
3610 }
3611 }
3612 // Weeny check for irreducible. This child was already visited (this
3613 // IS the post-work phase). Is this child's loop header post-visited
3614 // as well? If so, then I found another entry into the loop.
3615 if (!_verify_only) {
3616 while( is_postvisited(l->_head) ) {
3617 // found irreducible
3618 l->_irreducible = 1; // = true
3619 l = l->_parent;
3620 _has_irreducible_loops = true;
3621 // Check for bad CFG here to prevent crash, and bailout of compile
3622 if (l == NULL) {
3623 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3624 return pre_order;
3625 }
3626 }
3627 C->set_has_irreducible_loop(_has_irreducible_loops);
3628 }
3629
3630 // This Node might be a decision point for loops. It is only if
3631 // it's children belong to several different loops. The sort call
3632 // does a trivial amount of work if there is only 1 child or all
3633 // children belong to the same loop. If however, the children
3634 // belong to different loops, the sort call will properly set the
3635 // _parent pointers to show how the loops nest.
3636 //
3637 // In any case, it returns the tightest enclosing loop.
3638 innermost = sort( l, innermost );
3639 }
3640
3641 // Def-use info will have some dead stuff; dead stuff will have no
3642 // loop decided on.
3643
3644 // Am I a loop header? If so fix up my parent's child and next ptrs.
3645 if( innermost && innermost->_head == n ) {
3646 assert( get_loop(n) == innermost, "" );
3647 IdealLoopTree *p = innermost->_parent;
3648 IdealLoopTree *l = innermost;
3649 while( p && l->_head == n ) {
3650 l->_next = p->_child; // Put self on parents 'next child'
3651 p->_child = l; // Make self as first child of parent
3652 l = p; // Now walk up the parent chain
3653 p = l->_parent;
3654 }
3655 } else {
3656 // Note that it is possible for a LoopNode to reach here, if the
3657 // backedge has been made unreachable (hence the LoopNode no longer
3658 // denotes a Loop, and will eventually be removed).
3659
3660 // Record tightest enclosing loop for self. Mark as post-visited.
3661 set_loop(n, innermost);
3662 // Also record has_call flag early on
3663 if( innermost ) {
3664 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3665 // Do not count uncommon calls
3666 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3667 Node *iff = n->in(0)->in(0);
3668 // No any calls for vectorized loops.
3669 if( UseSuperWord || !iff->is_If() ||
3670 (n->in(0)->Opcode() == Op_IfFalse &&
3671 (1.0 - iff->as_If()->_prob) >= 0.01) ||
3672 (iff->as_If()->_prob >= 0.01) )
3673 innermost->_has_call = 1;
3674 }
3675 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3676 // Disable loop optimizations if the loop has a scalar replaceable
3677 // allocation. This disabling may cause a potential performance lost
3678 // if the allocation is not eliminated for some reason.
3679 innermost->_allow_optimizations = false;
3680 innermost->_has_call = 1; // = true
3681 } else if (n->Opcode() == Op_SafePoint) {
3682 // Record all safepoints in this loop.
3683 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3684 innermost->_safepts->push(n);
3685 }
3686 }
3687 }
3688
3689 // Flag as post-visited now
3690 set_postvisited(n);
3691 return pre_order;
3692 }
3693
3694
3695 //------------------------------build_loop_early-------------------------------
3696 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3697 // First pass computes the earliest controlling node possible. This is the
3698 // controlling input with the deepest dominating depth.
3699 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3700 while (worklist.size() != 0) {
3701 // Use local variables nstack_top_n & nstack_top_i to cache values
3702 // on nstack's top.
3703 Node *nstack_top_n = worklist.pop();
3704 uint nstack_top_i = 0;
3705 //while_nstack_nonempty:
3706 while (true) {
3707 // Get parent node and next input's index from stack's top.
3708 Node *n = nstack_top_n;
3709 uint i = nstack_top_i;
3710 uint cnt = n->req(); // Count of inputs
3711 if (i == 0) { // Pre-process the node.
3712 if( has_node(n) && // Have either loop or control already?
3713 !has_ctrl(n) ) { // Have loop picked out already?
3714 // During "merge_many_backedges" we fold up several nested loops
3715 // into a single loop. This makes the members of the original
3716 // loop bodies pointing to dead loops; they need to move up
3717 // to the new UNION'd larger loop. I set the _head field of these
3718 // dead loops to NULL and the _parent field points to the owning
3719 // loop. Shades of UNION-FIND algorithm.
3720 IdealLoopTree *ilt;
3721 while( !(ilt = get_loop(n))->_head ) {
3722 // Normally I would use a set_loop here. But in this one special
3723 // case, it is legal (and expected) to change what loop a Node
3724 // belongs to.
3725 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3726 }
3727 // Remove safepoints ONLY if I've already seen I don't need one.
3728 // (the old code here would yank a 2nd safepoint after seeing a
3729 // first one, even though the 1st did not dominate in the loop body
3730 // and thus could be avoided indefinitely)
3731 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3732 is_deleteable_safept(n)) {
3733 Node *in = n->in(TypeFunc::Control);
3734 lazy_replace(n,in); // Pull safepoint now
3735 if (ilt->_safepts != NULL) {
3736 ilt->_safepts->yank(n);
3737 }
3738 // Carry on with the recursion "as if" we are walking
3739 // only the control input
3740 if( !visited.test_set( in->_idx ) ) {
3741 worklist.push(in); // Visit this guy later, using worklist
3742 }
3743 // Get next node from nstack:
3744 // - skip n's inputs processing by setting i > cnt;
3745 // - we also will not call set_early_ctrl(n) since
3746 // has_node(n) == true (see the condition above).
3747 i = cnt + 1;
3748 }
3749 }
3750 } // if (i == 0)
3751
3752 // Visit all inputs
3753 bool done = true; // Assume all n's inputs will be processed
3754 while (i < cnt) {
3755 Node *in = n->in(i);
3756 ++i;
3757 if (in == NULL) continue;
3758 if (in->pinned() && !in->is_CFG())
3759 set_ctrl(in, in->in(0));
3760 int is_visited = visited.test_set( in->_idx );
3761 if (!has_node(in)) { // No controlling input yet?
3762 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3763 assert( !is_visited, "visit only once" );
3764 nstack.push(n, i); // Save parent node and next input's index.
3765 nstack_top_n = in; // Process current input now.
3766 nstack_top_i = 0;
3767 done = false; // Not all n's inputs processed.
3768 break; // continue while_nstack_nonempty;
3769 } else if (!is_visited) {
3770 // This guy has a location picked out for him, but has not yet
3771 // been visited. Happens to all CFG nodes, for instance.
3772 // Visit him using the worklist instead of recursion, to break
3773 // cycles. Since he has a location already we do not need to
3774 // find his location before proceeding with the current Node.
3775 worklist.push(in); // Visit this guy later, using worklist
3776 }
3777 }
3778 if (done) {
3779 // All of n's inputs have been processed, complete post-processing.
3780
3781 // Compute earliest point this Node can go.
3782 // CFG, Phi, pinned nodes already know their controlling input.
3783 if (!has_node(n)) {
3784 // Record earliest legal location
3785 set_early_ctrl( n );
3786 }
3787 if (nstack.is_empty()) {
3788 // Finished all nodes on stack.
3789 // Process next node on the worklist.
3790 break;
3791 }
3792 // Get saved parent node and next input's index.
3793 nstack_top_n = nstack.node();
3794 nstack_top_i = nstack.index();
3795 nstack.pop();
3796 }
3797 } // while (true)
3798 }
3799 }
3800
3801 //------------------------------dom_lca_internal--------------------------------
3802 // Pair-wise LCA
3803 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3804 if( !n1 ) return n2; // Handle NULL original LCA
3805 assert( n1->is_CFG(), "" );
3806 assert( n2->is_CFG(), "" );
3807 // find LCA of all uses
3808 uint d1 = dom_depth(n1);
3809 uint d2 = dom_depth(n2);
3810 while (n1 != n2) {
3811 if (d1 > d2) {
3812 n1 = idom(n1);
3813 d1 = dom_depth(n1);
3814 } else if (d1 < d2) {
3815 n2 = idom(n2);
3816 d2 = dom_depth(n2);
3817 } else {
3818 // Here d1 == d2. Due to edits of the dominator-tree, sections
3819 // of the tree might have the same depth. These sections have
3820 // to be searched more carefully.
3821
3822 // Scan up all the n1's with equal depth, looking for n2.
3823 Node *t1 = idom(n1);
3824 while (dom_depth(t1) == d1) {
3825 if (t1 == n2) return n2;
3826 t1 = idom(t1);
3827 }
3828 // Scan up all the n2's with equal depth, looking for n1.
3829 Node *t2 = idom(n2);
3830 while (dom_depth(t2) == d2) {
3831 if (t2 == n1) return n1;
3832 t2 = idom(t2);
3833 }
3834 // Move up to a new dominator-depth value as well as up the dom-tree.
3835 n1 = t1;
3836 n2 = t2;
3837 d1 = dom_depth(n1);
3838 d2 = dom_depth(n2);
3839 }
3840 }
3841 return n1;
3842 }
3843
3844 //------------------------------compute_idom-----------------------------------
3845 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3846 // IDOMs are correct.
3847 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3848 assert( region->is_Region(), "" );
3849 Node *LCA = NULL;
3850 for( uint i = 1; i < region->req(); i++ ) {
3851 if( region->in(i) != C->top() )
3852 LCA = dom_lca( LCA, region->in(i) );
3853 }
3854 return LCA;
3855 }
3856
3857 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3858 bool had_error = false;
3859 #ifdef ASSERT
3860 if (early != C->root()) {
3861 // Make sure that there's a dominance path from LCA to early
3862 Node* d = LCA;
3863 while (d != early) {
3864 if (d == C->root()) {
3865 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3866 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3867 had_error = true;
3868 break;
3869 }
3870 d = idom(d);
3871 }
3872 }
3873 #endif
3874 return had_error;
3875 }
3876
3877
3878 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3879 // Compute LCA over list of uses
3880 bool had_error = false;
3881 Node *LCA = NULL;
3882 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3883 Node* c = n->fast_out(i);
3884 if (_nodes[c->_idx] == NULL)
3885 continue; // Skip the occasional dead node
3886 if( c->is_Phi() ) { // For Phis, we must land above on the path
3887 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3888 if( c->in(j) == n ) { // Found matching input?
3889 Node *use = c->in(0)->in(j);
3890 if (_verify_only && use->is_top()) continue;
3891 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3892 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3893 }
3894 }
3895 } else {
3896 // For CFG data-users, use is in the block just prior
3897 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3898 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3899 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3900 }
3901 }
3902 assert(!had_error, "bad dominance");
3903 return LCA;
3904 }
3905
3906 // Check the shape of the graph at the loop entry. In some cases,
3907 // the shape of the graph does not match the shape outlined below.
3908 // That is caused by the Opaque1 node "protecting" the shape of
3909 // the graph being removed by, for example, the IGVN performed
3910 // in PhaseIdealLoop::build_and_optimize().
3911 //
3912 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3913 // loop unswitching, and IGVN, or a combination of them) can freely change
3914 // the graph's shape. As a result, the graph shape outlined below cannot
3915 // be guaranteed anymore.
3916 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) {
3917 if (!cl->is_main_loop() && !cl->is_post_loop()) {
3918 return false;
3919 }
3920 Node* ctrl = cl->skip_predicates();
3921
3922 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3923 return false;
3924 }
3925 Node* iffm = ctrl->in(0);
3926 if (iffm == NULL || !iffm->is_If()) {
3927 return false;
3928 }
3929 Node* bolzm = iffm->in(1);
3930 if (bolzm == NULL || !bolzm->is_Bool()) {
3931 return false;
3932 }
3933 Node* cmpzm = bolzm->in(1);
3934 if (cmpzm == NULL || !cmpzm->is_Cmp()) {
3935 return false;
3936 }
3937 // compares can get conditionally flipped
3938 bool found_opaque = false;
3939 for (uint i = 1; i < cmpzm->req(); i++) {
3940 Node* opnd = cmpzm->in(i);
3941 if (opnd && opnd->Opcode() == Op_Opaque1) {
3942 found_opaque = true;
3943 break;
3944 }
3945 }
3946 if (!found_opaque) {
3947 return false;
3948 }
3949 return true;
3950 }
3951
3952 //------------------------------get_late_ctrl----------------------------------
3953 // Compute latest legal control.
3954 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3955 assert(early != NULL, "early control should not be NULL");
3956
3957 Node* LCA = compute_lca_of_uses(n, early);
3958 #ifdef ASSERT
3959 if (LCA == C->root() && LCA != early) {
3960 // def doesn't dominate uses so print some useful debugging output
3961 compute_lca_of_uses(n, early, true);
3962 }
3963 #endif
3964
3965 // if this is a load, check for anti-dependent stores
3966 // We use a conservative algorithm to identify potential interfering
3967 // instructions and for rescheduling the load. The users of the memory
3968 // input of this load are examined. Any use which is not a load and is
3969 // dominated by early is considered a potentially interfering store.
3970 // This can produce false positives.
3971 if (n->is_Load() && LCA != early) {
3972 Node_List worklist;
3973
3974 Node *mem = n->in(MemNode::Memory);
3975 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3976 Node* s = mem->fast_out(i);
3977 worklist.push(s);
3978 }
3979 while(worklist.size() != 0 && LCA != early) {
3980 Node* s = worklist.pop();
3981 if (s->is_Load() || s->is_ShenandoahBarrier() || s->Opcode() == Op_SafePoint ||
3982 (UseShenandoahGC && s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)) {
3983 continue;
3984 } else if (s->is_MergeMem()) {
3985 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3986 Node* s1 = s->fast_out(i);
3987 worklist.push(s1);
3988 }
3989 } else {
3990 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3991 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3992 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3993 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3994 }
3995 }
3996 }
3997 }
3998
3999 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
4000 return LCA;
4001 }
4002
4003 // true if CFG node d dominates CFG node n
4004 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
4005 if (d == n)
4006 return true;
4007 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
4008 uint dd = dom_depth(d);
4009 while (dom_depth(n) >= dd) {
4010 if (n == d)
4011 return true;
4012 n = idom(n);
4013 }
4014 return false;
4015 }
4016
4017 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
4018 // Pair-wise LCA with tags.
4019 // Tag each index with the node 'tag' currently being processed
4020 // before advancing up the dominator chain using idom().
4021 // Later calls that find a match to 'tag' know that this path has already
4022 // been considered in the current LCA (which is input 'n1' by convention).
4023 // Since get_late_ctrl() is only called once for each node, the tag array
4024 // does not need to be cleared between calls to get_late_ctrl().
4025 // Algorithm trades a larger constant factor for better asymptotic behavior
4026 //
4027 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
4028 uint d1 = dom_depth(n1);
4029 uint d2 = dom_depth(n2);
4030
4031 do {
4032 if (d1 > d2) {
4033 // current lca is deeper than n2
4034 _dom_lca_tags.map(n1->_idx, tag);
4035 n1 = idom(n1);
4036 d1 = dom_depth(n1);
4037 } else if (d1 < d2) {
4038 // n2 is deeper than current lca
4039 Node *memo = _dom_lca_tags[n2->_idx];
4040 if( memo == tag ) {
4041 return n1; // Return the current LCA
4042 }
4043 _dom_lca_tags.map(n2->_idx, tag);
4044 n2 = idom(n2);
4045 d2 = dom_depth(n2);
4046 } else {
4047 // Here d1 == d2. Due to edits of the dominator-tree, sections
4048 // of the tree might have the same depth. These sections have
4049 // to be searched more carefully.
4050
4051 // Scan up all the n1's with equal depth, looking for n2.
4052 _dom_lca_tags.map(n1->_idx, tag);
4053 Node *t1 = idom(n1);
4054 while (dom_depth(t1) == d1) {
4055 if (t1 == n2) return n2;
4056 _dom_lca_tags.map(t1->_idx, tag);
4057 t1 = idom(t1);
4058 }
4059 // Scan up all the n2's with equal depth, looking for n1.
4060 _dom_lca_tags.map(n2->_idx, tag);
4061 Node *t2 = idom(n2);
4062 while (dom_depth(t2) == d2) {
4063 if (t2 == n1) return n1;
4064 _dom_lca_tags.map(t2->_idx, tag);
4065 t2 = idom(t2);
4066 }
4067 // Move up to a new dominator-depth value as well as up the dom-tree.
4068 n1 = t1;
4069 n2 = t2;
4070 d1 = dom_depth(n1);
4071 d2 = dom_depth(n2);
4072 }
4073 } while (n1 != n2);
4074 return n1;
4075 }
4076
4077 //------------------------------init_dom_lca_tags------------------------------
4078 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4079 // Intended use does not involve any growth for the array, so it could
4080 // be of fixed size.
4081 void PhaseIdealLoop::init_dom_lca_tags() {
4082 uint limit = C->unique() + 1;
4083 _dom_lca_tags.map( limit, NULL );
4084 #ifdef ASSERT
4085 for( uint i = 0; i < limit; ++i ) {
4086 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4087 }
4088 #endif // ASSERT
4089 }
4090
4091 //------------------------------clear_dom_lca_tags------------------------------
4092 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4093 // Intended use does not involve any growth for the array, so it could
4094 // be of fixed size.
4095 void PhaseIdealLoop::clear_dom_lca_tags() {
4096 uint limit = C->unique() + 1;
4097 _dom_lca_tags.map( limit, NULL );
4098 _dom_lca_tags.clear();
4099 #ifdef ASSERT
4100 for( uint i = 0; i < limit; ++i ) {
4101 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4102 }
4103 #endif // ASSERT
4104 }
4105
4106 //------------------------------build_loop_late--------------------------------
4107 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4108 // Second pass finds latest legal placement, and ideal loop placement.
4109 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
4110 while (worklist.size() != 0) {
4111 Node *n = worklist.pop();
4112 // Only visit once
4113 if (visited.test_set(n->_idx)) continue;
4114 uint cnt = n->outcnt();
4115 uint i = 0;
4116 while (true) {
4117 assert( _nodes[n->_idx], "no dead nodes" );
4118 // Visit all children
4119 if (i < cnt) {
4120 Node* use = n->raw_out(i);
4121 ++i;
4122 // Check for dead uses. Aggressively prune such junk. It might be
4123 // dead in the global sense, but still have local uses so I cannot
4124 // easily call 'remove_dead_node'.
4125 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
4126 // Due to cycles, we might not hit the same fixed point in the verify
4127 // pass as we do in the regular pass. Instead, visit such phis as
4128 // simple uses of the loop head.
4129 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
4130 if( !visited.test(use->_idx) )
4131 worklist.push(use);
4132 } else if( !visited.test_set(use->_idx) ) {
4133 nstack.push(n, i); // Save parent and next use's index.
4134 n = use; // Process all children of current use.
4135 cnt = use->outcnt();
4136 i = 0;
4137 }
4138 } else {
4139 // Do not visit around the backedge of loops via data edges.
4140 // push dead code onto a worklist
4141 _deadlist.push(use);
4142 }
4143 } else {
4144 // All of n's children have been processed, complete post-processing.
4145 build_loop_late_post(n);
4146 if (nstack.is_empty()) {
4147 // Finished all nodes on stack.
4148 // Process next node on the worklist.
4149 break;
4150 }
4151 // Get saved parent node and next use's index. Visit the rest of uses.
4152 n = nstack.node();
4153 cnt = n->outcnt();
4154 i = nstack.index();
4155 nstack.pop();
4156 }
4157 }
4158 }
4159 }
4160
4161 // Verify that no data node is schedules in the outer loop of a strip
4162 // mined loop.
4163 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
4164 #ifdef ASSERT
4165 if (get_loop(least)->_nest == 0) {
4166 return;
4167 }
4168 IdealLoopTree* loop = get_loop(least);
4169 Node* head = loop->_head;
4170 if (head->is_OuterStripMinedLoop() &&
4171 // Verification can't be applied to fully built strip mined loops
4172 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
4173 Node* sfpt = head->as_Loop()->outer_safepoint();
4174 ResourceMark rm;
4175 Unique_Node_List wq;
4176 wq.push(sfpt);
4177 for (uint i = 0; i < wq.size(); i++) {
4178 Node *m = wq.at(i);
4179 for (uint i = 1; i < m->req(); i++) {
4180 Node* nn = m->in(i);
4181 if (nn == n) {
4182 return;
4183 }
4184 if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
4185 wq.push(nn);
4186 }
4187 }
4188 }
4189 ShouldNotReachHere();
4190 }
4191 #endif
4192 }
4193
4194
4195 //------------------------------build_loop_late_post---------------------------
4196 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4197 // Second pass finds latest legal placement, and ideal loop placement.
4198 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
4199
4200 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
4201 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
4202 }
4203
4204 #ifdef ASSERT
4205 if (_verify_only && !n->is_CFG()) {
4206 // Check def-use domination.
4207 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
4208 }
4209 #endif
4210
4211 // CFG and pinned nodes already handled
4212 if( n->in(0) ) {
4213 if( n->in(0)->is_top() ) return; // Dead?
4214
4215 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
4216 // _must_ be pinned (they have to observe their control edge of course).
4217 // Unlike Stores (which modify an unallocable resource, the memory
4218 // state), Mods/Loads can float around. So free them up.
4219 bool pinned = true;
4220 switch( n->Opcode() ) {
4221 case Op_DivI:
4222 case Op_DivF:
4223 case Op_DivD:
4224 case Op_ModI:
4225 case Op_ModF:
4226 case Op_ModD:
4227 case Op_LoadB: // Same with Loads; they can sink
4228 case Op_LoadUB: // during loop optimizations.
4229 case Op_LoadUS:
4230 case Op_LoadD:
4231 case Op_LoadF:
4232 case Op_LoadI:
4233 case Op_LoadKlass:
4234 case Op_LoadNKlass:
4235 case Op_LoadL:
4236 case Op_LoadS:
4237 case Op_LoadP:
4238 case Op_LoadBarrierSlowReg:
4239 case Op_LoadBarrierWeakSlowReg:
4240 case Op_LoadN:
4241 case Op_LoadRange:
4242 case Op_LoadD_unaligned:
4243 case Op_LoadL_unaligned:
4244 case Op_StrComp: // Does a bunch of load-like effects
4245 case Op_StrEquals:
4246 case Op_StrIndexOf:
4247 case Op_StrIndexOfChar:
4248 case Op_AryEq:
4249 case Op_HasNegatives:
4250 pinned = false;
4251 }
4252 if (UseShenandoahGC && n->is_CMove()) {
4253 pinned = false;
4254 }
4255 if( pinned ) {
4256 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
4257 if( !chosen_loop->_child ) // Inner loop?
4258 chosen_loop->_body.push(n); // Collect inner loops
4259 return;
4260 }
4261 } else { // No slot zero
4262 if( n->is_CFG() ) { // CFG with no slot 0 is dead
4263 _nodes.map(n->_idx,0); // No block setting, it's globally dead
4264 return;
4265 }
4266 assert(!n->is_CFG() || n->outcnt() == 0, "");
4267 }
4268
4269 // Do I have a "safe range" I can select over?
4270 Node *early = get_ctrl(n);// Early location already computed
4271
4272 // Compute latest point this Node can go
4273 Node *LCA = get_late_ctrl( n, early );
4274 // LCA is NULL due to uses being dead
4275 if( LCA == NULL ) {
4276 #ifdef ASSERT
4277 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
4278 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
4279 }
4280 #endif
4281 _nodes.map(n->_idx, 0); // This node is useless
4282 _deadlist.push(n);
4283 return;
4284 }
4285 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
4286
4287 Node *legal = LCA; // Walk 'legal' up the IDOM chain
4288 Node *least = legal; // Best legal position so far
4289 while( early != legal ) { // While not at earliest legal
4290 #ifdef ASSERT
4291 if (legal->is_Start() && !early->is_Root()) {
4292 // Bad graph. Print idom path and fail.
4293 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
4294 assert(false, "Bad graph detected in build_loop_late");
4295 }
4296 #endif
4297 // Find least loop nesting depth
4298 legal = idom(legal); // Bump up the IDOM tree
4299 // Check for lower nesting depth
4300 if( get_loop(legal)->_nest < get_loop(least)->_nest )
4301 least = legal;
4302 }
4303 assert(early == legal || legal != C->root(), "bad dominance of inputs");
4304
4305 // Try not to place code on a loop entry projection
4306 // which can inhibit range check elimination.
4307 if (least != early) {
4308 Node* ctrl_out = least->unique_ctrl_out();
4309 if (ctrl_out && ctrl_out->is_Loop() &&
4310 least == ctrl_out->in(LoopNode::EntryControl)) {
4311 // Move the node above predicates as far up as possible so a
4312 // following pass of loop predication doesn't hoist a predicate
4313 // that depends on it above that node.
4314 Node* new_ctrl = least;
4315 for (;;) {
4316 if (!new_ctrl->is_Proj()) {
4317 break;
4318 }
4319 CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
4320 if (call == NULL) {
4321 break;
4322 }
4323 int req = call->uncommon_trap_request();
4324 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
4325 if (trap_reason != Deoptimization::Reason_loop_limit_check &&
4326 trap_reason != Deoptimization::Reason_predicate &&
4327 trap_reason != Deoptimization::Reason_profile_predicate) {
4328 break;
4329 }
4330 Node* c = new_ctrl->in(0)->in(0);
4331 if (is_dominator(c, early) && c != early) {
4332 break;
4333 }
4334 new_ctrl = c;
4335 }
4336 least = new_ctrl;
4337 }
4338 }
4339
4340 #ifdef ASSERT
4341 // If verifying, verify that 'verify_me' has a legal location
4342 // and choose it as our location.
4343 if( _verify_me ) {
4344 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
4345 Node *legal = LCA;
4346 while( early != legal ) { // While not at earliest legal
4347 if( legal == v_ctrl ) break; // Check for prior good location
4348 legal = idom(legal) ;// Bump up the IDOM tree
4349 }
4350 // Check for prior good location
4351 if( legal == v_ctrl ) least = legal; // Keep prior if found
4352 }
4353 #endif
4354
4355 // Assign discovered "here or above" point
4356 least = find_non_split_ctrl(least);
4357 verify_strip_mined_scheduling(n, least);
4358 set_ctrl(n, least);
4359
4360 // Collect inner loop bodies
4361 IdealLoopTree *chosen_loop = get_loop(least);
4362 if( !chosen_loop->_child ) // Inner loop?
4363 chosen_loop->_body.push(n);// Collect inner loops
4364 }
4365
4366 #ifdef ASSERT
4367 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
4368 tty->print_cr("%s", msg);
4369 tty->print("n: "); n->dump();
4370 tty->print("early(n): "); early->dump();
4371 if (n->in(0) != NULL && !n->in(0)->is_top() &&
4372 n->in(0) != early && !n->in(0)->is_Root()) {
4373 tty->print("n->in(0): "); n->in(0)->dump();
4374 }
4375 for (uint i = 1; i < n->req(); i++) {
4376 Node* in1 = n->in(i);
4377 if (in1 != NULL && in1 != n && !in1->is_top()) {
4378 tty->print("n->in(%d): ", i); in1->dump();
4379 Node* in1_early = get_ctrl(in1);
4380 tty->print("early(n->in(%d)): ", i); in1_early->dump();
4381 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
4382 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
4383 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
4384 }
4385 for (uint j = 1; j < in1->req(); j++) {
4386 Node* in2 = in1->in(j);
4387 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
4388 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
4389 Node* in2_early = get_ctrl(in2);
4390 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
4391 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
4392 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
4393 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
4394 }
4395 }
4396 }
4397 }
4398 }
4399 tty->cr();
4400 tty->print("LCA(n): "); LCA->dump();
4401 for (uint i = 0; i < n->outcnt(); i++) {
4402 Node* u1 = n->raw_out(i);
4403 if (u1 == n)
4404 continue;
4405 tty->print("n->out(%d): ", i); u1->dump();
4406 if (u1->is_CFG()) {
4407 for (uint j = 0; j < u1->outcnt(); j++) {
4408 Node* u2 = u1->raw_out(j);
4409 if (u2 != u1 && u2 != n && u2->is_CFG()) {
4410 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4411 }
4412 }
4413 } else {
4414 Node* u1_later = get_ctrl(u1);
4415 tty->print("later(n->out(%d)): ", i); u1_later->dump();
4416 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
4417 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
4418 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
4419 }
4420 for (uint j = 0; j < u1->outcnt(); j++) {
4421 Node* u2 = u1->raw_out(j);
4422 if (u2 == n || u2 == u1)
4423 continue;
4424 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4425 if (!u2->is_CFG()) {
4426 Node* u2_later = get_ctrl(u2);
4427 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
4428 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
4429 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
4430 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
4431 }
4432 }
4433 }
4434 }
4435 }
4436 tty->cr();
4437 int ct = 0;
4438 Node *dbg_legal = LCA;
4439 while(!dbg_legal->is_Start() && ct < 100) {
4440 tty->print("idom[%d] ",ct); dbg_legal->dump();
4441 ct++;
4442 dbg_legal = idom(dbg_legal);
4443 }
4444 tty->cr();
4445 }
4446 #endif
4447
4448 #ifndef PRODUCT
4449 //------------------------------dump-------------------------------------------
4450 void PhaseIdealLoop::dump( ) const {
4451 ResourceMark rm;
4452 Arena* arena = Thread::current()->resource_area();
4453 Node_Stack stack(arena, C->live_nodes() >> 2);
4454 Node_List rpo_list;
4455 VectorSet visited(arena);
4456 visited.set(C->top()->_idx);
4457 rpo( C->root(), stack, visited, rpo_list );
4458 // Dump root loop indexed by last element in PO order
4459 dump( _ltree_root, rpo_list.size(), rpo_list );
4460 }
4461
4462 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
4463 loop->dump_head();
4464
4465 // Now scan for CFG nodes in the same loop
4466 for( uint j=idx; j > 0; j-- ) {
4467 Node *n = rpo_list[j-1];
4468 if( !_nodes[n->_idx] ) // Skip dead nodes
4469 continue;
4470 if( get_loop(n) != loop ) { // Wrong loop nest
4471 if( get_loop(n)->_head == n && // Found nested loop?
4472 get_loop(n)->_parent == loop )
4473 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
4474 continue;
4475 }
4476
4477 // Dump controlling node
4478 for( uint x = 0; x < loop->_nest; x++ )
4479 tty->print(" ");
4480 tty->print("C");
4481 if( n == C->root() ) {
4482 n->dump();
4483 } else {
4484 Node* cached_idom = idom_no_update(n);
4485 Node *computed_idom = n->in(0);
4486 if( n->is_Region() ) {
4487 computed_idom = compute_idom(n);
4488 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
4489 // any MultiBranch ctrl node), so apply a similar transform to
4490 // the cached idom returned from idom_no_update.
4491 cached_idom = find_non_split_ctrl(cached_idom);
4492 }
4493 tty->print(" ID:%d",computed_idom->_idx);
4494 n->dump();
4495 if( cached_idom != computed_idom ) {
4496 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
4497 computed_idom->_idx, cached_idom->_idx);
4498 }
4499 }
4500 // Dump nodes it controls
4501 for( uint k = 0; k < _nodes.Size(); k++ ) {
4502 // (k < C->unique() && get_ctrl(find(k)) == n)
4503 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
4504 Node *m = C->root()->find(k);
4505 if( m && m->outcnt() > 0 ) {
4506 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
4507 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
4508 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
4509 }
4510 for( uint j = 0; j < loop->_nest; j++ )
4511 tty->print(" ");
4512 tty->print(" ");
4513 m->dump();
4514 }
4515 }
4516 }
4517 }
4518 }
4519 #endif
4520
4521 // Collect a R-P-O for the whole CFG.
4522 // Result list is in post-order (scan backwards for RPO)
4523 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
4524 stk.push(start, 0);
4525 visited.set(start->_idx);
4526
4527 while (stk.is_nonempty()) {
4528 Node* m = stk.node();
4529 uint idx = stk.index();
4530 if (idx < m->outcnt()) {
4531 stk.set_index(idx + 1);
4532 Node* n = m->raw_out(idx);
4533 if (n->is_CFG() && !visited.test_set(n->_idx)) {
4534 stk.push(n, 0);
4535 }
4536 } else {
4537 rpo_list.push(m);
4538 stk.pop();
4539 }
4540 }
4541 }
4542
4543
4544 //=============================================================================
4545 //------------------------------LoopTreeIterator-----------------------------------
4546
4547 // Advance to next loop tree using a preorder, left-to-right traversal.
4548 void LoopTreeIterator::next() {
4549 assert(!done(), "must not be done.");
4550 if (_curnt->_child != NULL) {
4551 _curnt = _curnt->_child;
4552 } else if (_curnt->_next != NULL) {
4553 _curnt = _curnt->_next;
4554 } else {
4555 while (_curnt != _root && _curnt->_next == NULL) {
4556 _curnt = _curnt->_parent;
4557 }
4558 if (_curnt == _root) {
4559 _curnt = NULL;
4560 assert(done(), "must be done.");
4561 } else {
4562 assert(_curnt->_next != NULL, "must be more to do");
4563 _curnt = _curnt->_next;
4564 }
4565 }
4566 }