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
44 //=============================================================================
45 //------------------------------is_loop_iv-------------------------------------
46 // Determine if a node is Counted loop induction variable.
47 // The method is declared in node.hpp.
48 const Node* Node::is_loop_iv() const {
49 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
50 this->as_Phi()->region()->is_CountedLoop() &&
51 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
52 return this;
53 } else {
54 return NULL;
55 }
56 }
57
58 //=============================================================================
59 //------------------------------dump_spec--------------------------------------
60 // Dump special per-node info
61 #ifndef PRODUCT
62 void LoopNode::dump_spec(outputStream *st) const {
63 if (is_inner_loop()) st->print( "inner " );
64 if (is_partial_peel_loop()) st->print( "partial_peel " );
65 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
66 }
67 #endif
68
69 //------------------------------is_valid_counted_loop-------------------------
70 bool LoopNode::is_valid_counted_loop() const {
71 if (is_CountedLoop()) {
72 CountedLoopNode* l = as_CountedLoop();
73 CountedLoopEndNode* le = l->loopexit_or_null();
74 if (le != NULL &&
75 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
76 Node* phi = l->phi();
77 Node* exit = le->proj_out_or_null(0 /* false */);
78 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
79 phi != NULL && phi->is_Phi() &&
80 phi->in(LoopNode::LoopBackControl) == l->incr() &&
81 le->loopnode() == l && le->stride_is_con()) {
82 return true;
83 }
84 }
85 }
86 return false;
87 }
88
89 //------------------------------get_early_ctrl---------------------------------
90 // Compute earliest legal control
91 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
92 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
93 uint i;
94 Node *early;
95 if (n->in(0) && !n->is_expensive()) {
96 early = n->in(0);
97 if (!early->is_CFG()) // Might be a non-CFG multi-def
98 early = get_ctrl(early); // So treat input as a straight data input
99 i = 1;
100 } else {
101 early = get_ctrl(n->in(1));
102 i = 2;
103 }
104 uint e_d = dom_depth(early);
105 assert( early, "" );
106 for (; i < n->req(); i++) {
107 Node *cin = get_ctrl(n->in(i));
108 assert( cin, "" );
109 // Keep deepest dominator depth
110 uint c_d = dom_depth(cin);
111 if (c_d > e_d) { // Deeper guy?
112 early = cin; // Keep deepest found so far
113 e_d = c_d;
114 } else if (c_d == e_d && // Same depth?
115 early != cin) { // If not equal, must use slower algorithm
116 // If same depth but not equal, one _must_ dominate the other
117 // and we want the deeper (i.e., dominated) guy.
118 Node *n1 = early;
119 Node *n2 = cin;
120 while (1) {
121 n1 = idom(n1); // Walk up until break cycle
122 n2 = idom(n2);
123 if (n1 == cin || // Walked early up to cin
124 dom_depth(n2) < c_d)
125 break; // early is deeper; keep him
126 if (n2 == early || // Walked cin up to early
127 dom_depth(n1) < c_d) {
128 early = cin; // cin is deeper; keep him
129 break;
130 }
131 }
132 e_d = dom_depth(early); // Reset depth register cache
133 }
134 }
135
136 // Return earliest legal location
137 assert(early == find_non_split_ctrl(early), "unexpected early control");
138
139 if (n->is_expensive() && !_verify_only && !_verify_me) {
140 assert(n->in(0), "should have control input");
141 early = get_early_ctrl_for_expensive(n, early);
142 }
143
144 return early;
145 }
146
147 //------------------------------get_early_ctrl_for_expensive---------------------------------
148 // Move node up the dominator tree as high as legal while still beneficial
149 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
150 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
151 assert(OptimizeExpensiveOps, "optimization off?");
152
153 Node* ctl = n->in(0);
154 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
155 uint min_dom_depth = dom_depth(earliest);
156 #ifdef ASSERT
157 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
158 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
159 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
160 }
161 #endif
162 if (dom_depth(ctl) < min_dom_depth) {
163 return earliest;
164 }
165
166 while (1) {
167 Node *next = ctl;
168 // Moving the node out of a loop on the projection of a If
169 // confuses loop predication. So once we hit a Loop in a If branch
170 // that doesn't branch to an UNC, we stop. The code that process
171 // expensive nodes will notice the loop and skip over it to try to
172 // move the node further up.
173 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
174 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
175 break;
176 }
177 next = idom(ctl->in(1)->in(0));
178 } else if (ctl->is_Proj()) {
179 // We only move it up along a projection if the projection is
180 // the single control projection for its parent: same code path,
181 // if it's a If with UNC or fallthrough of a call.
182 Node* parent_ctl = ctl->in(0);
183 if (parent_ctl == NULL) {
184 break;
185 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
186 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
187 } else if (parent_ctl->is_If()) {
188 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
189 break;
190 }
191 assert(idom(ctl) == parent_ctl, "strange");
192 next = idom(parent_ctl);
193 } else if (ctl->is_CatchProj()) {
194 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
195 break;
196 }
197 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
198 next = parent_ctl->in(0)->in(0)->in(0);
199 } else {
200 // Check if parent control has a single projection (this
201 // control is the only possible successor of the parent
202 // control). If so, we can try to move the node above the
203 // parent control.
204 int nb_ctl_proj = 0;
205 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
206 Node *p = parent_ctl->fast_out(i);
207 if (p->is_Proj() && p->is_CFG()) {
208 nb_ctl_proj++;
209 if (nb_ctl_proj > 1) {
210 break;
211 }
212 }
213 }
214
215 if (nb_ctl_proj > 1) {
216 break;
217 }
218 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
219 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
220 assert(idom(ctl) == parent_ctl, "strange");
221 next = idom(parent_ctl);
222 }
223 } else {
224 next = idom(ctl);
225 }
226 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
227 break;
228 }
229 ctl = next;
230 }
231
232 if (ctl != n->in(0)) {
233 _igvn.replace_input_of(n, 0, ctl);
234 _igvn.hash_insert(n);
235 }
236
237 return ctl;
238 }
239
240
241 //------------------------------set_early_ctrl---------------------------------
242 // Set earliest legal control
243 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
244 Node *early = get_early_ctrl(n);
245
246 // Record earliest legal location
247 set_ctrl(n, early);
248 }
249
250 //------------------------------set_subtree_ctrl-------------------------------
251 // set missing _ctrl entries on new nodes
252 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
253 // Already set? Get out.
254 if( _nodes[n->_idx] ) return;
255 // Recursively set _nodes array to indicate where the Node goes
256 uint i;
257 for( i = 0; i < n->req(); ++i ) {
258 Node *m = n->in(i);
259 if( m && m != C->root() )
260 set_subtree_ctrl( m );
261 }
262
263 // Fixup self
264 set_early_ctrl( n );
265 }
266
267 // Create a skeleton strip mined outer loop: a Loop head before the
268 // inner strip mined loop, a safepoint and an exit condition guarded
269 // by an opaque node after the inner strip mined loop with a backedge
270 // to the loop head. The inner strip mined loop is left as it is. Only
271 // once loop optimizations are over, do we adjust the inner loop exit
272 // condition to limit its number of iterations, set the outer loop
273 // exit condition and add Phis to the outer loop head. Some loop
274 // optimizations that operate on the inner strip mined loop need to be
275 // aware of the outer strip mined loop: loop unswitching needs to
276 // clone the outer loop as well as the inner, unrolling needs to only
277 // clone the inner loop etc. No optimizations need to change the outer
278 // strip mined loop as it is only a skeleton.
279 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
280 IdealLoopTree* loop, float cl_prob, float le_fcnt,
281 Node*& entry_control, Node*& iffalse) {
282 Node* outer_test = _igvn.intcon(0);
283 set_ctrl(outer_test, C->root());
284 Node *orig = iffalse;
285 iffalse = iffalse->clone();
286 _igvn.register_new_node_with_optimizer(iffalse);
287 set_idom(iffalse, idom(orig), dom_depth(orig));
288
289 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
290 Node *outer_ift = new IfTrueNode (outer_le);
291 Node* outer_iff = orig;
292 _igvn.replace_input_of(outer_iff, 0, outer_le);
293
294 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
295 entry_control = outer_l;
296
297 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
298 IdealLoopTree* parent = loop->_parent;
299 IdealLoopTree* sibling = parent->_child;
300 if (sibling == loop) {
301 parent->_child = outer_ilt;
302 } else {
303 while (sibling->_next != loop) {
304 sibling = sibling->_next;
305 }
306 sibling->_next = outer_ilt;
307 }
308 outer_ilt->_next = loop->_next;
309 outer_ilt->_parent = parent;
310 outer_ilt->_child = loop;
311 outer_ilt->_nest = loop->_nest;
312 loop->_parent = outer_ilt;
313 loop->_next = NULL;
314 loop->_nest++;
315
316 set_loop(iffalse, outer_ilt);
317 register_control(outer_le, outer_ilt, iffalse);
318 register_control(outer_ift, outer_ilt, outer_le);
319 set_idom(outer_iff, outer_le, dom_depth(outer_le));
320 _igvn.register_new_node_with_optimizer(outer_l);
321 set_loop(outer_l, outer_ilt);
322 set_idom(outer_l, init_control, dom_depth(init_control)+1);
323
324 return outer_ilt;
325 }
326
327 void PhaseIdealLoop::insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol) {
328 Node* new_predicate_proj = create_new_if_for_predicate(limit_check_proj, NULL,
329 Deoptimization::Reason_loop_limit_check,
330 Op_If);
331 Node* iff = new_predicate_proj->in(0);
332 assert(iff->Opcode() == Op_If, "bad graph shape");
333 Node* conv = iff->in(1);
334 assert(conv->Opcode() == Op_Conv2B, "bad graph shape");
335 Node* opaq = conv->in(1);
336 assert(opaq->Opcode() == Op_Opaque1, "bad graph shape");
337 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
338 bol = _igvn.register_new_node_with_optimizer(bol);
339 set_subtree_ctrl(bol);
340 _igvn.replace_input_of(iff, 1, bol);
341
342 #ifndef PRODUCT
343 // report that the loop predication has been actually performed
344 // for this loop
345 if (TraceLoopLimitCheck) {
346 tty->print_cr("Counted Loop Limit Check generated:");
347 debug_only( bol->dump(2); )
348 }
349 #endif
350 }
351
352 //------------------------------is_counted_loop--------------------------------
353 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) {
354 PhaseGVN *gvn = &_igvn;
355
356 // Counted loop head must be a good RegionNode with only 3 not NULL
357 // control input edges: Self, Entry, LoopBack.
358 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
359 return false;
360 }
361 Node *init_control = x->in(LoopNode::EntryControl);
362 Node *back_control = x->in(LoopNode::LoopBackControl);
363 if (init_control == NULL || back_control == NULL) // Partially dead
364 return false;
365 // Must also check for TOP when looking for a dead loop
366 if (init_control->is_top() || back_control->is_top())
367 return false;
368
369 // Allow funny placement of Safepoint
370 if (back_control->Opcode() == Op_SafePoint) {
371 if (LoopStripMiningIter != 0) {
372 // Leaving the safepoint on the backedge and creating a
373 // CountedLoop will confuse optimizations. We can't move the
374 // safepoint around because its jvm state wouldn't match a new
375 // location. Give up on that loop.
376 return false;
377 }
378 back_control = back_control->in(TypeFunc::Control);
379 }
380
381 // Controlling test for loop
382 Node *iftrue = back_control;
383 uint iftrue_op = iftrue->Opcode();
384 if (iftrue_op != Op_IfTrue &&
385 iftrue_op != Op_IfFalse)
386 // I have a weird back-control. Probably the loop-exit test is in
387 // the middle of the loop and I am looking at some trailing control-flow
388 // merge point. To fix this I would have to partially peel the loop.
389 return false; // Obscure back-control
390
391 // Get boolean guarding loop-back test
392 Node *iff = iftrue->in(0);
393 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
394 return false;
395 BoolNode *test = iff->in(1)->as_Bool();
396 BoolTest::mask bt = test->_test._test;
397 float cl_prob = iff->as_If()->_prob;
398 if (iftrue_op == Op_IfFalse) {
399 bt = BoolTest(bt).negate();
400 cl_prob = 1.0 - cl_prob;
401 }
402 // Get backedge compare
403 Node *cmp = test->in(1);
404 int cmp_op = cmp->Opcode();
405 if (cmp_op != Op_CmpI)
406 return false; // Avoid pointer & float compares
407
408 // Find the trip-counter increment & limit. Limit must be loop invariant.
409 Node *incr = cmp->in(1);
410 Node *limit = cmp->in(2);
411
412 // ---------
413 // need 'loop()' test to tell if limit is loop invariant
414 // ---------
415
416 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
417 Node *tmp = incr; // Then reverse order into the CmpI
418 incr = limit;
419 limit = tmp;
420 bt = BoolTest(bt).commute(); // And commute the exit test
421 }
422 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
423 return false;
424 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
425 return false;
426
427 Node* phi_incr = NULL;
428 // Trip-counter increment must be commutative & associative.
429 if (incr->Opcode() == Op_CastII) {
430 incr = incr->in(1);
431 }
432 if (incr->is_Phi()) {
433 if (incr->as_Phi()->region() != x || incr->req() != 3)
434 return false; // Not simple trip counter expression
435 phi_incr = incr;
436 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
437 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
438 return false;
439 }
440
441 Node* trunc1 = NULL;
442 Node* trunc2 = NULL;
443 const TypeInt* iv_trunc_t = NULL;
444 Node* orig_incr = incr;
445 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
446 return false; // Funny increment opcode
447 }
448 assert(incr->Opcode() == Op_AddI, "wrong increment code");
449
450 const TypeInt* limit_t = gvn->type(limit)->is_int();
451 if (trunc1 != NULL) {
452 // When there is a truncation, we must be sure that after the truncation
453 // the trip counter will end up higher than the limit, otherwise we are looking
454 // at an endless loop. Can happen with range checks.
455
456 // Example:
457 // int i = 0;
458 // while (true)
459 // sum + = array[i];
460 // i++;
461 // i = i && 0x7fff;
462 // }
463 //
464 // If the array is shorter than 0x8000 this exits through a AIOOB
465 // - Counted loop transformation is ok
466 // If the array is longer then this is an endless loop
467 // - No transformation can be done.
468
469 const TypeInt* incr_t = gvn->type(orig_incr)->is_int();
470 if (limit_t->_hi > incr_t->_hi) {
471 // if the limit can have a higher value than the increment (before the phi)
472 return false;
473 }
474 }
475
476 // Get merge point
477 Node *xphi = incr->in(1);
478 Node *stride = incr->in(2);
479 if (!stride->is_Con()) { // Oops, swap these
480 if (!xphi->is_Con()) // Is the other guy a constant?
481 return false; // Nope, unknown stride, bail out
482 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
483 xphi = stride;
484 stride = tmp;
485 }
486 if (xphi->Opcode() == Op_CastII) {
487 xphi = xphi->in(1);
488 }
489 // Stride must be constant
490 int stride_con = stride->get_int();
491 if (stride_con == 0)
492 return false; // missed some peephole opt
493
494 if (!xphi->is_Phi())
495 return false; // Too much math on the trip counter
496 if (phi_incr != NULL && phi_incr != xphi)
497 return false;
498 PhiNode *phi = xphi->as_Phi();
499
500 // Phi must be of loop header; backedge must wrap to increment
501 if (phi->region() != x)
502 return false;
503 if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) ||
504 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) {
505 return false;
506 }
507 Node *init_trip = phi->in(LoopNode::EntryControl);
508
509 // If iv trunc type is smaller than int, check for possible wrap.
510 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
511 assert(trunc1 != NULL, "must have found some truncation");
512
513 // Get a better type for the phi (filtered thru if's)
514 const TypeInt* phi_ft = filtered_type(phi);
515
516 // Can iv take on a value that will wrap?
517 //
518 // Ensure iv's limit is not within "stride" of the wrap value.
519 //
520 // Example for "short" type
521 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
522 // If the stride is +10, then the last value of the induction
523 // variable before the increment (phi_ft->_hi) must be
524 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
525 // ensure no truncation occurs after the increment.
526
527 if (stride_con > 0) {
528 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
529 iv_trunc_t->_lo > phi_ft->_lo) {
530 return false; // truncation may occur
531 }
532 } else if (stride_con < 0) {
533 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
534 iv_trunc_t->_hi < phi_ft->_hi) {
535 return false; // truncation may occur
536 }
537 }
538 // No possibility of wrap so truncation can be discarded
539 // Promote iv type to Int
540 } else {
541 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
542 }
543
544 // If the condition is inverted and we will be rolling
545 // through MININT to MAXINT, then bail out.
546 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
547 // Odd stride
548 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
549 // Count down loop rolls through MAXINT
550 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
551 // Count up loop rolls through MININT
552 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
553 return false; // Bail out
554 }
555
556 const TypeInt* init_t = gvn->type(init_trip)->is_int();
557
558 if (stride_con > 0) {
559 jlong init_p = (jlong)init_t->_lo + stride_con;
560 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
561 return false; // cyclic loop or this loop trips only once
562 } else {
563 jlong init_p = (jlong)init_t->_hi + stride_con;
564 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
565 return false; // cyclic loop or this loop trips only once
566 }
567
568 if (phi_incr != NULL && bt != BoolTest::ne) {
569 // check if there is a possiblity of IV overflowing after the first increment
570 if (stride_con > 0) {
571 if (init_t->_hi > max_jint - stride_con) {
572 return false;
573 }
574 } else {
575 if (init_t->_lo < min_jint - stride_con) {
576 return false;
577 }
578 }
579 }
580
581 // =================================================
582 // ---- SUCCESS! Found A Trip-Counted Loop! -----
583 //
584 assert(x->Opcode() == Op_Loop, "regular loops only");
585 C->print_method(PHASE_BEFORE_CLOOPS, 3);
586
587 Node *hook = new Node(6);
588
589 // ===================================================
590 // Generate loop limit check to avoid integer overflow
591 // in cases like next (cyclic loops):
592 //
593 // for (i=0; i <= max_jint; i++) {}
594 // for (i=0; i < max_jint; i+=2) {}
595 //
596 //
597 // Limit check predicate depends on the loop test:
598 //
599 // for(;i != limit; i++) --> limit <= (max_jint)
600 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
601 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
602 //
603
604 // Check if limit is excluded to do more precise int overflow check.
605 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
606 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
607
608 // If compare points directly to the phi we need to adjust
609 // the compare so that it points to the incr. Limit have
610 // to be adjusted to keep trip count the same and the
611 // adjusted limit should be checked for int overflow.
612 if (phi_incr != NULL) {
613 stride_m += stride_con;
614 }
615
616 if (limit->is_Con()) {
617 int limit_con = limit->get_int();
618 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
619 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
620 // Bailout: it could be integer overflow.
621 return false;
622 }
623 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
624 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
625 // Limit's type may satisfy the condition, for example,
626 // when it is an array length.
627 } else {
628 // Generate loop's limit check.
629 // Loop limit check predicate should be near the loop.
630 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
631 if (!limit_check_proj) {
632 // The limit check predicate is not generated if this method trapped here before.
633 #ifdef ASSERT
634 if (TraceLoopLimitCheck) {
635 tty->print("missing loop limit check:");
636 loop->dump_head();
637 x->dump(1);
638 }
639 #endif
640 return false;
641 }
642
643 IfNode* check_iff = limit_check_proj->in(0)->as_If();
644
645 if (!is_dominator(get_ctrl(limit), check_iff->in(0))) {
646 return false;
647 }
648
649 Node* cmp_limit;
650 Node* bol;
651
652 if (stride_con > 0) {
653 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
654 bol = new BoolNode(cmp_limit, BoolTest::le);
655 } else {
656 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
657 bol = new BoolNode(cmp_limit, BoolTest::ge);
658 }
659
660 insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
661 }
662
663 // Now we need to canonicalize loop condition.
664 if (bt == BoolTest::ne) {
665 assert(stride_con == 1 || stride_con == -1, "simple increment only");
666 if (stride_con > 0 && init_t->_hi < limit_t->_lo) {
667 // 'ne' can be replaced with 'lt' only when init < limit.
668 bt = BoolTest::lt;
669 } else if (stride_con < 0 && init_t->_lo > limit_t->_hi) {
670 // 'ne' can be replaced with 'gt' only when init > limit.
671 bt = BoolTest::gt;
672 } else {
673 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
674 if (!limit_check_proj) {
675 // The limit check predicate is not generated if this method trapped here before.
676 #ifdef ASSERT
677 if (TraceLoopLimitCheck) {
678 tty->print("missing loop limit check:");
679 loop->dump_head();
680 x->dump(1);
681 }
682 #endif
683 return false;
684 }
685 IfNode* check_iff = limit_check_proj->in(0)->as_If();
686
687 if (!is_dominator(get_ctrl(limit), check_iff->in(0)) ||
688 !is_dominator(get_ctrl(init_trip), check_iff->in(0))) {
689 return false;
690 }
691
692 Node* cmp_limit;
693 Node* bol;
694
695 if (stride_con > 0) {
696 cmp_limit = new CmpINode(init_trip, limit);
697 bol = new BoolNode(cmp_limit, BoolTest::lt);
698 } else {
699 cmp_limit = new CmpINode(init_trip, limit);
700 bol = new BoolNode(cmp_limit, BoolTest::gt);
701 }
702
703 insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
704
705 if (stride_con > 0) {
706 // 'ne' can be replaced with 'lt' only when init < limit.
707 bt = BoolTest::lt;
708 } else if (stride_con < 0) {
709 // 'ne' can be replaced with 'gt' only when init > limit.
710 bt = BoolTest::gt;
711 }
712 }
713 }
714
715 if (phi_incr != NULL) {
716 // If compare points directly to the phi we need to adjust
717 // the compare so that it points to the incr. Limit have
718 // to be adjusted to keep trip count the same and we
719 // should avoid int overflow.
720 //
721 // i = init; do {} while(i++ < limit);
722 // is converted to
723 // i = init; do {} while(++i < limit+1);
724 //
725 limit = gvn->transform(new AddINode(limit, stride));
726 }
727
728 if (incl_limit) {
729 // The limit check guaranties that 'limit <= (max_jint - stride)' so
730 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
731 //
732 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
733 limit = gvn->transform(new AddINode(limit, one));
734 if (bt == BoolTest::le)
735 bt = BoolTest::lt;
736 else if (bt == BoolTest::ge)
737 bt = BoolTest::gt;
738 else
739 ShouldNotReachHere();
740 }
741 set_subtree_ctrl( limit );
742
743 if (LoopStripMiningIter == 0) {
744 // Check for SafePoint on backedge and remove
745 Node *sfpt = x->in(LoopNode::LoopBackControl);
746 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
747 lazy_replace( sfpt, iftrue );
748 if (loop->_safepts != NULL) {
749 loop->_safepts->yank(sfpt);
750 }
751 loop->_tail = iftrue;
752 }
753 }
754
755 // Build a canonical trip test.
756 // Clone code, as old values may be in use.
757 incr = incr->clone();
758 incr->set_req(1,phi);
759 incr->set_req(2,stride);
760 incr = _igvn.register_new_node_with_optimizer(incr);
761 set_early_ctrl( incr );
762 _igvn.rehash_node_delayed(phi);
763 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
764
765 // If phi type is more restrictive than Int, raise to
766 // Int to prevent (almost) infinite recursion in igvn
767 // which can only handle integer types for constants or minint..maxint.
768 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
769 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
770 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
771 nphi = _igvn.register_new_node_with_optimizer(nphi);
772 set_ctrl(nphi, get_ctrl(phi));
773 _igvn.replace_node(phi, nphi);
774 phi = nphi->as_Phi();
775 }
776 cmp = cmp->clone();
777 cmp->set_req(1,incr);
778 cmp->set_req(2,limit);
779 cmp = _igvn.register_new_node_with_optimizer(cmp);
780 set_ctrl(cmp, iff->in(0));
781
782 test = test->clone()->as_Bool();
783 (*(BoolTest*)&test->_test)._test = bt;
784 test->set_req(1,cmp);
785 _igvn.register_new_node_with_optimizer(test);
786 set_ctrl(test, iff->in(0));
787
788 // Replace the old IfNode with a new LoopEndNode
789 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
790 IfNode *le = lex->as_If();
791 uint dd = dom_depth(iff);
792 set_idom(le, le->in(0), dd); // Update dominance for loop exit
793 set_loop(le, loop);
794
795 // Get the loop-exit control
796 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
797
798 // Need to swap loop-exit and loop-back control?
799 if (iftrue_op == Op_IfFalse) {
800 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
801 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
802
803 loop->_tail = back_control = ift2;
804 set_loop(ift2, loop);
805 set_loop(iff2, get_loop(iffalse));
806
807 // Lazy update of 'get_ctrl' mechanism.
808 lazy_replace(iffalse, iff2);
809 lazy_replace(iftrue, ift2);
810
811 // Swap names
812 iffalse = iff2;
813 iftrue = ift2;
814 } else {
815 _igvn.rehash_node_delayed(iffalse);
816 _igvn.rehash_node_delayed(iftrue);
817 iffalse->set_req_X( 0, le, &_igvn );
818 iftrue ->set_req_X( 0, le, &_igvn );
819 }
820
821 set_idom(iftrue, le, dd+1);
822 set_idom(iffalse, le, dd+1);
823 assert(iff->outcnt() == 0, "should be dead now");
824 lazy_replace( iff, le ); // fix 'get_ctrl'
825
826 Node *sfpt2 = le->in(0);
827
828 Node* entry_control = init_control;
829 bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL &&
830 sfpt2->Opcode() == Op_SafePoint && !loop->_has_call;
831 IdealLoopTree* outer_ilt = NULL;
832 if (strip_mine_loop) {
833 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
834 cl_prob, le->_fcnt, entry_control,
835 iffalse);
836 }
837
838 // Now setup a new CountedLoopNode to replace the existing LoopNode
839 CountedLoopNode *l = new CountedLoopNode(entry_control, back_control);
840 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
841 // The following assert is approximately true, and defines the intention
842 // of can_be_counted_loop. It fails, however, because phase->type
843 // is not yet initialized for this loop and its parts.
844 //assert(l->can_be_counted_loop(this), "sanity");
845 _igvn.register_new_node_with_optimizer(l);
846 set_loop(l, loop);
847 loop->_head = l;
848 // Fix all data nodes placed at the old loop head.
849 // Uses the lazy-update mechanism of 'get_ctrl'.
850 lazy_replace( x, l );
851 set_idom(l, entry_control, dom_depth(entry_control) + 1);
852
853 if (LoopStripMiningIter == 0 || strip_mine_loop) {
854 // Check for immediately preceding SafePoint and remove
855 if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) {
856 if (strip_mine_loop) {
857 Node* outer_le = outer_ilt->_tail->in(0);
858 Node* sfpt = sfpt2->clone();
859 sfpt->set_req(0, iffalse);
860 outer_le->set_req(0, sfpt);
861 register_control(sfpt, outer_ilt, iffalse);
862 set_idom(outer_le, sfpt, dom_depth(sfpt));
863 }
864 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
865 if (loop->_safepts != NULL) {
866 loop->_safepts->yank(sfpt2);
867 }
868 }
869 }
870
871 // Free up intermediate goo
872 _igvn.remove_dead_node(hook);
873
874 #ifdef ASSERT
875 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
876 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
877 #endif
878 #ifndef PRODUCT
879 if (TraceLoopOpts) {
880 tty->print("Counted ");
881 loop->dump_head();
882 }
883 #endif
884
885 C->print_method(PHASE_AFTER_CLOOPS, 3);
886
887 // Capture bounds of the loop in the induction variable Phi before
888 // subsequent transformation (iteration splitting) obscures the
889 // bounds
890 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
891
892 if (strip_mine_loop) {
893 l->mark_strip_mined();
894 l->verify_strip_mined(1);
895 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
896 loop = outer_ilt;
897 }
898
899 return true;
900 }
901
902 //----------------------exact_limit-------------------------------------------
903 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
904 assert(loop->_head->is_CountedLoop(), "");
905 CountedLoopNode *cl = loop->_head->as_CountedLoop();
906 assert(cl->is_valid_counted_loop(), "");
907
908 if (ABS(cl->stride_con()) == 1 ||
909 cl->limit()->Opcode() == Op_LoopLimit) {
910 // Old code has exact limit (it could be incorrect in case of int overflow).
911 // Loop limit is exact with stride == 1. And loop may already have exact limit.
912 return cl->limit();
913 }
914 Node *limit = NULL;
915 #ifdef ASSERT
916 BoolTest::mask bt = cl->loopexit()->test_trip();
917 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
918 #endif
919 if (cl->has_exact_trip_count()) {
920 // Simple case: loop has constant boundaries.
921 // Use jlongs to avoid integer overflow.
922 int stride_con = cl->stride_con();
923 jlong init_con = cl->init_trip()->get_int();
924 jlong limit_con = cl->limit()->get_int();
925 julong trip_cnt = cl->trip_count();
926 jlong final_con = init_con + trip_cnt*stride_con;
927 int final_int = (int)final_con;
928 // The final value should be in integer range since the loop
929 // is counted and the limit was checked for overflow.
930 assert(final_con == (jlong)final_int, "final value should be integer");
931 limit = _igvn.intcon(final_int);
932 } else {
933 // Create new LoopLimit node to get exact limit (final iv value).
934 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
935 register_new_node(limit, cl->in(LoopNode::EntryControl));
936 }
937 assert(limit != NULL, "sanity");
938 return limit;
939 }
940
941 //------------------------------Ideal------------------------------------------
942 // Return a node which is more "ideal" than the current node.
943 // Attempt to convert into a counted-loop.
944 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
945 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
946 phase->C->set_major_progress();
947 }
948 return RegionNode::Ideal(phase, can_reshape);
949 }
950
951 void LoopNode::verify_strip_mined(int expect_skeleton) const {
952 #ifdef ASSERT
953 const OuterStripMinedLoopNode* outer = NULL;
954 const CountedLoopNode* inner = NULL;
955 if (is_strip_mined()) {
956 assert(is_CountedLoop(), "no Loop should be marked strip mined");
957 inner = as_CountedLoop();
958 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
959 } else if (is_OuterStripMinedLoop()) {
960 outer = this->as_OuterStripMinedLoop();
961 inner = outer->unique_ctrl_out()->as_CountedLoop();
962 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
963 }
964 if (inner != NULL || outer != NULL) {
965 assert(inner != NULL && outer != NULL, "missing loop in strip mined nest");
966 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
967 Node* outer_le = outer_tail->in(0);
968 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
969 Node* sfpt = outer_le->in(0);
970 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
971 Node* inner_out = sfpt->in(0);
972 if (inner_out->outcnt() != 1) {
973 ResourceMark rm;
974 Unique_Node_List wq;
975
976 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
977 Node* u = inner_out->fast_out(i);
978 if (u == sfpt) {
979 continue;
980 }
981 wq.clear();
982 wq.push(u);
983 bool found_sfpt = false;
984 for (uint next = 0; next < wq.size() && !found_sfpt; next++) {
985 Node *n = wq.at(next);
986 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) {
987 Node* u = n->fast_out(i);
988 if (u == sfpt) {
989 found_sfpt = true;
990 }
991 if (!u->is_CFG()) {
992 wq.push(u);
993 }
994 }
995 }
996 assert(found_sfpt, "no node in loop that's not input to safepoint");
997 }
998 }
999 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
1000 assert(cle == inner->loopexit_or_null(), "mismatch");
1001 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
1002 if (has_skeleton) {
1003 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
1004 assert(outer->outcnt() == 2, "only phis");
1005 } else {
1006 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
1007 uint phis = 0;
1008 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
1009 Node* u = inner->fast_out(i);
1010 if (u->is_Phi()) {
1011 phis++;
1012 }
1013 }
1014 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
1015 Node* u = outer->fast_out(i);
1016 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
1017 }
1018 uint stores = 0;
1019 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
1020 Node* u = inner_out->fast_out(i);
1021 if (u->is_Store()) {
1022 stores++;
1023 }
1024 }
1025 assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
1026 }
1027 assert(sfpt->outcnt() == 1, "no data node");
1028 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
1029 }
1030 #endif
1031 }
1032
1033 //=============================================================================
1034 //------------------------------Ideal------------------------------------------
1035 // Return a node which is more "ideal" than the current node.
1036 // Attempt to convert into a counted-loop.
1037 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1038 return RegionNode::Ideal(phase, can_reshape);
1039 }
1040
1041 //------------------------------dump_spec--------------------------------------
1042 // Dump special per-node info
1043 #ifndef PRODUCT
1044 void CountedLoopNode::dump_spec(outputStream *st) const {
1045 LoopNode::dump_spec(st);
1046 if (stride_is_con()) {
1047 st->print("stride: %d ",stride_con());
1048 }
1049 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
1050 if (is_main_loop()) st->print("main of N%d", _idx);
1051 if (is_post_loop()) st->print("post of N%d", _main_idx);
1052 if (is_strip_mined()) st->print(" strip mined");
1053 }
1054 #endif
1055
1056 //=============================================================================
1057 int CountedLoopEndNode::stride_con() const {
1058 return stride()->bottom_type()->is_int()->get_con();
1059 }
1060
1061 //=============================================================================
1062 //------------------------------Value-----------------------------------------
1063 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
1064 const Type* init_t = phase->type(in(Init));
1065 const Type* limit_t = phase->type(in(Limit));
1066 const Type* stride_t = phase->type(in(Stride));
1067 // Either input is TOP ==> the result is TOP
1068 if (init_t == Type::TOP) return Type::TOP;
1069 if (limit_t == Type::TOP) return Type::TOP;
1070 if (stride_t == Type::TOP) return Type::TOP;
1071
1072 int stride_con = stride_t->is_int()->get_con();
1073 if (stride_con == 1)
1074 return NULL; // Identity
1075
1076 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
1077 // Use jlongs to avoid integer overflow.
1078 jlong init_con = init_t->is_int()->get_con();
1079 jlong limit_con = limit_t->is_int()->get_con();
1080 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
1081 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
1082 jlong final_con = init_con + stride_con*trip_count;
1083 int final_int = (int)final_con;
1084 // The final value should be in integer range since the loop
1085 // is counted and the limit was checked for overflow.
1086 assert(final_con == (jlong)final_int, "final value should be integer");
1087 return TypeInt::make(final_int);
1088 }
1089
1090 return bottom_type(); // TypeInt::INT
1091 }
1092
1093 //------------------------------Ideal------------------------------------------
1094 // Return a node which is more "ideal" than the current node.
1095 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1096 if (phase->type(in(Init)) == Type::TOP ||
1097 phase->type(in(Limit)) == Type::TOP ||
1098 phase->type(in(Stride)) == Type::TOP)
1099 return NULL; // Dead
1100
1101 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1102 if (stride_con == 1)
1103 return NULL; // Identity
1104
1105 if (in(Init)->is_Con() && in(Limit)->is_Con())
1106 return NULL; // Value
1107
1108 // Delay following optimizations until all loop optimizations
1109 // done to keep Ideal graph simple.
1110 if (!can_reshape || phase->C->major_progress())
1111 return NULL;
1112
1113 const TypeInt* init_t = phase->type(in(Init) )->is_int();
1114 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
1115 int stride_p;
1116 jlong lim, ini;
1117 julong max;
1118 if (stride_con > 0) {
1119 stride_p = stride_con;
1120 lim = limit_t->_hi;
1121 ini = init_t->_lo;
1122 max = (julong)max_jint;
1123 } else {
1124 stride_p = -stride_con;
1125 lim = init_t->_hi;
1126 ini = limit_t->_lo;
1127 max = (julong)min_jint;
1128 }
1129 julong range = lim - ini + stride_p;
1130 if (range <= max) {
1131 // Convert to integer expression if it is not overflow.
1132 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
1133 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
1134 Node *bias = phase->transform(new AddINode(range, stride_m));
1135 Node *trip = phase->transform(new DivINode(0, bias, in(Stride)));
1136 Node *span = phase->transform(new MulINode(trip, in(Stride)));
1137 return new AddINode(span, in(Init)); // exact limit
1138 }
1139
1140 if (is_power_of_2(stride_p) || // divisor is 2^n
1141 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
1142 // Convert to long expression to avoid integer overflow
1143 // and let igvn optimizer convert this division.
1144 //
1145 Node* init = phase->transform( new ConvI2LNode(in(Init)));
1146 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
1147 Node* stride = phase->longcon(stride_con);
1148 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
1149
1150 Node *range = phase->transform(new SubLNode(limit, init));
1151 Node *bias = phase->transform(new AddLNode(range, stride_m));
1152 Node *span;
1153 if (stride_con > 0 && is_power_of_2(stride_p)) {
1154 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
1155 // and avoid generating rounding for division. Zero trip guard should
1156 // guarantee that init < limit but sometimes the guard is missing and
1157 // we can get situation when init > limit. Note, for the empty loop
1158 // optimization zero trip guard is generated explicitly which leaves
1159 // only RCE predicate where exact limit is used and the predicate
1160 // will simply fail forcing recompilation.
1161 Node* neg_stride = phase->longcon(-stride_con);
1162 span = phase->transform(new AndLNode(bias, neg_stride));
1163 } else {
1164 Node *trip = phase->transform(new DivLNode(0, bias, stride));
1165 span = phase->transform(new MulLNode(trip, stride));
1166 }
1167 // Convert back to int
1168 Node *span_int = phase->transform(new ConvL2INode(span));
1169 return new AddINode(span_int, in(Init)); // exact limit
1170 }
1171
1172 return NULL; // No progress
1173 }
1174
1175 //------------------------------Identity---------------------------------------
1176 // If stride == 1 return limit node.
1177 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1178 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1179 if (stride_con == 1 || stride_con == -1)
1180 return in(Limit);
1181 return this;
1182 }
1183
1184 //=============================================================================
1185 //----------------------match_incr_with_optional_truncation--------------------
1186 // Match increment with optional truncation:
1187 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1188 // Return NULL for failure. Success returns the increment node.
1189 Node* CountedLoopNode::match_incr_with_optional_truncation(
1190 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1191 // Quick cutouts:
1192 if (expr == NULL || expr->req() != 3) return NULL;
1193
1194 Node *t1 = NULL;
1195 Node *t2 = NULL;
1196 const TypeInt* trunc_t = TypeInt::INT;
1197 Node* n1 = expr;
1198 int n1op = n1->Opcode();
1199
1200 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1201 if (n1op == Op_AndI &&
1202 n1->in(2)->is_Con() &&
1203 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1204 // %%% This check should match any mask of 2**K-1.
1205 t1 = n1;
1206 n1 = t1->in(1);
1207 n1op = n1->Opcode();
1208 trunc_t = TypeInt::CHAR;
1209 } else if (n1op == Op_RShiftI &&
1210 n1->in(1) != NULL &&
1211 n1->in(1)->Opcode() == Op_LShiftI &&
1212 n1->in(2) == n1->in(1)->in(2) &&
1213 n1->in(2)->is_Con()) {
1214 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1215 // %%% This check should match any shift in [1..31].
1216 if (shift == 16 || shift == 8) {
1217 t1 = n1;
1218 t2 = t1->in(1);
1219 n1 = t2->in(1);
1220 n1op = n1->Opcode();
1221 if (shift == 16) {
1222 trunc_t = TypeInt::SHORT;
1223 } else if (shift == 8) {
1224 trunc_t = TypeInt::BYTE;
1225 }
1226 }
1227 }
1228
1229 // If (maybe after stripping) it is an AddI, we won:
1230 if (n1op == Op_AddI) {
1231 *trunc1 = t1;
1232 *trunc2 = t2;
1233 *trunc_type = trunc_t;
1234 return n1;
1235 }
1236
1237 // failed
1238 return NULL;
1239 }
1240
1241 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
1242 if (is_strip_mined()) {
1243 verify_strip_mined(expect_skeleton);
1244 return in(EntryControl)->as_Loop();
1245 }
1246 return this;
1247 }
1248
1249 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
1250 assert(is_strip_mined(), "not a strip mined loop");
1251 Node* c = in(EntryControl);
1252 if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) {
1253 return NULL;
1254 }
1255 return c->as_OuterStripMinedLoop();
1256 }
1257
1258 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
1259 Node* c = in(LoopBackControl);
1260 if (c == NULL || c->is_top()) {
1261 return NULL;
1262 }
1263 return c->as_IfTrue();
1264 }
1265
1266 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
1267 LoopNode* l = outer_loop();
1268 if (l == NULL) {
1269 return NULL;
1270 }
1271 return l->outer_loop_tail();
1272 }
1273
1274 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
1275 IfTrueNode* proj = outer_loop_tail();
1276 if (proj == NULL) {
1277 return NULL;
1278 }
1279 Node* c = proj->in(0);
1280 if (c == NULL || c->is_top() || c->outcnt() != 2) {
1281 return NULL;
1282 }
1283 return c->as_OuterStripMinedLoopEnd();
1284 }
1285
1286 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
1287 LoopNode* l = outer_loop();
1288 if (l == NULL) {
1289 return NULL;
1290 }
1291 return l->outer_loop_end();
1292 }
1293
1294 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
1295 IfNode* le = outer_loop_end();
1296 if (le == NULL) {
1297 return NULL;
1298 }
1299 Node* c = le->proj_out_or_null(false);
1300 if (c == NULL) {
1301 return NULL;
1302 }
1303 return c->as_IfFalse();
1304 }
1305
1306 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
1307 LoopNode* l = outer_loop();
1308 if (l == NULL) {
1309 return NULL;
1310 }
1311 return l->outer_loop_exit();
1312 }
1313
1314 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
1315 IfNode* le = outer_loop_end();
1316 if (le == NULL) {
1317 return NULL;
1318 }
1319 Node* c = le->in(0);
1320 if (c == NULL || c->is_top()) {
1321 return NULL;
1322 }
1323 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
1324 return c->as_SafePoint();
1325 }
1326
1327 SafePointNode* CountedLoopNode::outer_safepoint() const {
1328 LoopNode* l = outer_loop();
1329 if (l == NULL) {
1330 return NULL;
1331 }
1332 return l->outer_safepoint();
1333 }
1334
1335 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) {
1336 while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() &&
1337 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 &&
1338 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) {
1339 ctrl = ctrl->in(0)->in(0);
1340 }
1341
1342 return ctrl;
1343 }
1344
1345 Node* CountedLoopNode::skip_predicates() {
1346 if (is_main_loop()) {
1347 Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
1348
1349 return skip_predicates_from_entry(ctrl);
1350 }
1351 return in(LoopNode::EntryControl);
1352 }
1353
1354 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
1355 // Look for the outer & inner strip mined loop, reduce number of
1356 // iterations of the inner loop, set exit condition of outer loop,
1357 // construct required phi nodes for outer loop.
1358 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
1359 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
1360 Node* inner_iv_phi = inner_cl->phi();
1361 if (inner_iv_phi == NULL) {
1362 IfNode* outer_le = outer_loop_end();
1363 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1364 igvn->replace_node(outer_le, iff);
1365 inner_cl->clear_strip_mined();
1366 return;
1367 }
1368 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
1369
1370 int stride = inner_cl->stride_con();
1371 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride);
1372 int scaled_iters = (int)scaled_iters_long;
1373 int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride);
1374 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
1375 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
1376 assert(iter_estimate > 0, "broken");
1377 if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) {
1378 // Remove outer loop and safepoint (too few iterations)
1379 Node* outer_sfpt = outer_safepoint();
1380 Node* outer_out = outer_loop_exit();
1381 igvn->replace_node(outer_out, outer_sfpt->in(0));
1382 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
1383 inner_cl->clear_strip_mined();
1384 return;
1385 }
1386 if (iter_estimate <= scaled_iters_long) {
1387 // We would only go through one iteration of
1388 // the outer loop: drop the outer loop but
1389 // keep the safepoint so we don't run for
1390 // too long without a safepoint
1391 IfNode* outer_le = outer_loop_end();
1392 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1393 igvn->replace_node(outer_le, iff);
1394 inner_cl->clear_strip_mined();
1395 return;
1396 }
1397
1398 Node* cle_tail = inner_cle->proj_out(true);
1399 ResourceMark rm;
1400 Node_List old_new;
1401 if (cle_tail->outcnt() > 1) {
1402 // Look for nodes on backedge of inner loop and clone them
1403 Unique_Node_List backedge_nodes;
1404 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
1405 Node* u = cle_tail->fast_out(i);
1406 if (u != inner_cl) {
1407 assert(!u->is_CFG(), "control flow on the backedge?");
1408 backedge_nodes.push(u);
1409 }
1410 }
1411 uint last = igvn->C->unique();
1412 for (uint next = 0; next < backedge_nodes.size(); next++) {
1413 Node* n = backedge_nodes.at(next);
1414 old_new.map(n->_idx, n->clone());
1415 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1416 Node* u = n->fast_out(i);
1417 assert(!u->is_CFG(), "broken");
1418 if (u->_idx >= last) {
1419 continue;
1420 }
1421 if (!u->is_Phi()) {
1422 backedge_nodes.push(u);
1423 } else {
1424 assert(u->in(0) == inner_cl, "strange phi on the backedge");
1425 }
1426 }
1427 }
1428 // Put the clones on the outer loop backedge
1429 Node* le_tail = outer_loop_tail();
1430 for (uint next = 0; next < backedge_nodes.size(); next++) {
1431 Node *n = old_new[backedge_nodes.at(next)->_idx];
1432 for (uint i = 1; i < n->req(); i++) {
1433 if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) {
1434 n->set_req(i, old_new[n->in(i)->_idx]);
1435 }
1436 }
1437 if (n->in(0) != NULL && n->in(0) == cle_tail) {
1438 n->set_req(0, le_tail);
1439 }
1440 igvn->register_new_node_with_optimizer(n);
1441 }
1442 }
1443
1444 Node* iv_phi = NULL;
1445 // Make a clone of each phi in the inner loop
1446 // for the outer loop
1447 for (uint i = 0; i < inner_cl->outcnt(); i++) {
1448 Node* u = inner_cl->raw_out(i);
1449 if (u->is_Phi()) {
1450 assert(u->in(0) == inner_cl, "inconsistent");
1451 Node* phi = u->clone();
1452 phi->set_req(0, this);
1453 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
1454 if (be != NULL) {
1455 phi->set_req(LoopNode::LoopBackControl, be);
1456 }
1457 phi = igvn->transform(phi);
1458 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
1459 if (u == inner_iv_phi) {
1460 iv_phi = phi;
1461 }
1462 }
1463 }
1464 Node* cle_out = inner_cle->proj_out(false);
1465 if (cle_out->outcnt() > 1) {
1466 // Look for chains of stores that were sunk
1467 // out of the inner loop and are in the outer loop
1468 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
1469 Node* u = cle_out->fast_out(i);
1470 if (u->is_Store()) {
1471 Node* first = u;
1472 for(;;) {
1473 Node* next = first->in(MemNode::Memory);
1474 if (!next->is_Store() || next->in(0) != cle_out) {
1475 break;
1476 }
1477 first = next;
1478 }
1479 Node* last = u;
1480 for(;;) {
1481 Node* next = NULL;
1482 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
1483 Node* uu = last->fast_out(j);
1484 if (uu->is_Store() && uu->in(0) == cle_out) {
1485 assert(next == NULL, "only one in the outer loop");
1486 next = uu;
1487 }
1488 }
1489 if (next == NULL) {
1490 break;
1491 }
1492 last = next;
1493 }
1494 Node* phi = NULL;
1495 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1496 Node* uu = fast_out(j);
1497 if (uu->is_Phi()) {
1498 Node* be = uu->in(LoopNode::LoopBackControl);
1499 if (be->is_Store() && old_new[be->_idx] != NULL) {
1500 assert(false, "store on the backedge + sunk stores: unsupported");
1501 // drop outer loop
1502 IfNode* outer_le = outer_loop_end();
1503 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1504 igvn->replace_node(outer_le, iff);
1505 inner_cl->clear_strip_mined();
1506 return;
1507 }
1508 if (be == last || be == first->in(MemNode::Memory)) {
1509 assert(phi == NULL, "only one phi");
1510 phi = uu;
1511 }
1512 }
1513 }
1514 #ifdef ASSERT
1515 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1516 Node* uu = fast_out(j);
1517 if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) {
1518 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
1519 assert(phi == uu, "what's that phi?");
1520 } else if (uu->adr_type() == TypePtr::BOTTOM) {
1521 Node* n = uu->in(LoopNode::LoopBackControl);
1522 uint limit = igvn->C->live_nodes();
1523 uint i = 0;
1524 while (n != uu) {
1525 i++;
1526 assert(i < limit, "infinite loop");
1527 if (n->is_Proj()) {
1528 n = n->in(0);
1529 } else if (n->is_SafePoint() || n->is_MemBar()) {
1530 n = n->in(TypeFunc::Memory);
1531 } else if (n->is_Phi()) {
1532 n = n->in(1);
1533 } else if (n->is_MergeMem()) {
1534 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
1535 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
1536 n = n->in(MemNode::Memory);
1537 } else {
1538 n->dump();
1539 ShouldNotReachHere();
1540 }
1541 }
1542 }
1543 }
1544 }
1545 #endif
1546 if (phi == NULL) {
1547 // If the an entire chains was sunk, the
1548 // inner loop has no phi for that memory
1549 // slice, create one for the outer loop
1550 phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY,
1551 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
1552 phi->set_req(LoopNode::LoopBackControl, last);
1553 phi = igvn->transform(phi);
1554 igvn->replace_input_of(first, MemNode::Memory, phi);
1555 } else {
1556 // Or fix the outer loop fix to include
1557 // that chain of stores.
1558 Node* be = phi->in(LoopNode::LoopBackControl);
1559 assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported");
1560 if (be == first->in(MemNode::Memory)) {
1561 if (be == phi->in(LoopNode::LoopBackControl)) {
1562 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
1563 } else {
1564 igvn->replace_input_of(be, MemNode::Memory, last);
1565 }
1566 } else {
1567 #ifdef ASSERT
1568 if (be == phi->in(LoopNode::LoopBackControl)) {
1569 assert(phi->in(LoopNode::LoopBackControl) == last, "");
1570 } else {
1571 assert(be->in(MemNode::Memory) == last, "");
1572 }
1573 #endif
1574 }
1575 }
1576 }
1577 }
1578 }
1579
1580 if (iv_phi != NULL) {
1581 // Now adjust the inner loop's exit condition
1582 Node* limit = inner_cl->limit();
1583 Node* sub = NULL;
1584 if (stride > 0) {
1585 sub = igvn->transform(new SubINode(limit, iv_phi));
1586 } else {
1587 sub = igvn->transform(new SubINode(iv_phi, limit));
1588 }
1589 Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters)));
1590 Node* new_limit = NULL;
1591 if (stride > 0) {
1592 new_limit = igvn->transform(new AddINode(min, iv_phi));
1593 } else {
1594 new_limit = igvn->transform(new SubINode(iv_phi, min));
1595 }
1596 Node* cmp = inner_cle->cmp_node()->clone();
1597 igvn->replace_input_of(cmp, 2, new_limit);
1598 Node* bol = inner_cle->in(CountedLoopEndNode::TestValue)->clone();
1599 cmp->set_req(2, limit);
1600 bol->set_req(1, igvn->transform(cmp));
1601 igvn->replace_input_of(outer_loop_end(), 1, igvn->transform(bol));
1602 } else {
1603 assert(false, "should be able to adjust outer loop");
1604 IfNode* outer_le = outer_loop_end();
1605 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1606 igvn->replace_node(outer_le, iff);
1607 inner_cl->clear_strip_mined();
1608 }
1609 }
1610
1611 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
1612 if (!in(0)) return Type::TOP;
1613 if (phase->type(in(0)) == Type::TOP)
1614 return Type::TOP;
1615
1616 return TypeTuple::IFBOTH;
1617 }
1618
1619 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1620 if (remove_dead_region(phase, can_reshape)) return this;
1621
1622 return NULL;
1623 }
1624
1625 //------------------------------filtered_type--------------------------------
1626 // Return a type based on condition control flow
1627 // A successful return will be a type that is restricted due
1628 // to a series of dominating if-tests, such as:
1629 // if (i < 10) {
1630 // if (i > 0) {
1631 // here: "i" type is [1..10)
1632 // }
1633 // }
1634 // or a control flow merge
1635 // if (i < 10) {
1636 // do {
1637 // phi( , ) -- at top of loop type is [min_int..10)
1638 // i = ?
1639 // } while ( i < 10)
1640 //
1641 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1642 assert(n && n->bottom_type()->is_int(), "must be int");
1643 const TypeInt* filtered_t = NULL;
1644 if (!n->is_Phi()) {
1645 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1646 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1647
1648 } else {
1649 Node* phi = n->as_Phi();
1650 Node* region = phi->in(0);
1651 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1652 if (region && region != C->top()) {
1653 for (uint i = 1; i < phi->req(); i++) {
1654 Node* val = phi->in(i);
1655 Node* use_c = region->in(i);
1656 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1657 if (val_t != NULL) {
1658 if (filtered_t == NULL) {
1659 filtered_t = val_t;
1660 } else {
1661 filtered_t = filtered_t->meet(val_t)->is_int();
1662 }
1663 }
1664 }
1665 }
1666 }
1667 const TypeInt* n_t = _igvn.type(n)->is_int();
1668 if (filtered_t != NULL) {
1669 n_t = n_t->join(filtered_t)->is_int();
1670 }
1671 return n_t;
1672 }
1673
1674
1675 //------------------------------filtered_type_from_dominators--------------------------------
1676 // Return a possibly more restrictive type for val based on condition control flow of dominators
1677 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1678 if (val->is_Con()) {
1679 return val->bottom_type()->is_int();
1680 }
1681 uint if_limit = 10; // Max number of dominating if's visited
1682 const TypeInt* rtn_t = NULL;
1683
1684 if (use_ctrl && use_ctrl != C->top()) {
1685 Node* val_ctrl = get_ctrl(val);
1686 uint val_dom_depth = dom_depth(val_ctrl);
1687 Node* pred = use_ctrl;
1688 uint if_cnt = 0;
1689 while (if_cnt < if_limit) {
1690 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1691 if_cnt++;
1692 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1693 if (if_t != NULL) {
1694 if (rtn_t == NULL) {
1695 rtn_t = if_t;
1696 } else {
1697 rtn_t = rtn_t->join(if_t)->is_int();
1698 }
1699 }
1700 }
1701 pred = idom(pred);
1702 if (pred == NULL || pred == C->top()) {
1703 break;
1704 }
1705 // Stop if going beyond definition block of val
1706 if (dom_depth(pred) < val_dom_depth) {
1707 break;
1708 }
1709 }
1710 }
1711 return rtn_t;
1712 }
1713
1714
1715 //------------------------------dump_spec--------------------------------------
1716 // Dump special per-node info
1717 #ifndef PRODUCT
1718 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1719 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1720 BoolTest bt( test_trip()); // Added this for g++.
1721
1722 st->print("[");
1723 bt.dump_on(st);
1724 st->print("]");
1725 }
1726 st->print(" ");
1727 IfNode::dump_spec(st);
1728 }
1729 #endif
1730
1731 //=============================================================================
1732 //------------------------------is_member--------------------------------------
1733 // Is 'l' a member of 'this'?
1734 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1735 while( l->_nest > _nest ) l = l->_parent;
1736 return l == this;
1737 }
1738
1739 //------------------------------set_nest---------------------------------------
1740 // Set loop tree nesting depth. Accumulate _has_call bits.
1741 int IdealLoopTree::set_nest( uint depth ) {
1742 _nest = depth;
1743 int bits = _has_call;
1744 if( _child ) bits |= _child->set_nest(depth+1);
1745 if( bits ) _has_call = 1;
1746 if( _next ) bits |= _next ->set_nest(depth );
1747 return bits;
1748 }
1749
1750 //------------------------------split_fall_in----------------------------------
1751 // Split out multiple fall-in edges from the loop header. Move them to a
1752 // private RegionNode before the loop. This becomes the loop landing pad.
1753 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1754 PhaseIterGVN &igvn = phase->_igvn;
1755 uint i;
1756
1757 // Make a new RegionNode to be the landing pad.
1758 Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1759 phase->set_loop(landing_pad,_parent);
1760 // Gather all the fall-in control paths into the landing pad
1761 uint icnt = fall_in_cnt;
1762 uint oreq = _head->req();
1763 for( i = oreq-1; i>0; i-- )
1764 if( !phase->is_member( this, _head->in(i) ) )
1765 landing_pad->set_req(icnt--,_head->in(i));
1766
1767 // Peel off PhiNode edges as well
1768 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1769 Node *oj = _head->fast_out(j);
1770 if( oj->is_Phi() ) {
1771 PhiNode* old_phi = oj->as_Phi();
1772 assert( old_phi->region() == _head, "" );
1773 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1774 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1775 uint icnt = fall_in_cnt;
1776 for( i = oreq-1; i>0; i-- ) {
1777 if( !phase->is_member( this, _head->in(i) ) ) {
1778 p->init_req(icnt--, old_phi->in(i));
1779 // Go ahead and clean out old edges from old phi
1780 old_phi->del_req(i);
1781 }
1782 }
1783 // Search for CSE's here, because ZKM.jar does a lot of
1784 // loop hackery and we need to be a little incremental
1785 // with the CSE to avoid O(N^2) node blow-up.
1786 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1787 if( p2 ) { // Found CSE
1788 p->destruct(); // Recover useless new node
1789 p = p2; // Use old node
1790 } else {
1791 igvn.register_new_node_with_optimizer(p, old_phi);
1792 }
1793 // Make old Phi refer to new Phi.
1794 old_phi->add_req(p);
1795 // Check for the special case of making the old phi useless and
1796 // disappear it. In JavaGrande I have a case where this useless
1797 // Phi is the loop limit and prevents recognizing a CountedLoop
1798 // which in turn prevents removing an empty loop.
1799 Node *id_old_phi = igvn.apply_identity(old_phi);
1800 if( id_old_phi != old_phi ) { // Found a simple identity?
1801 // Note that I cannot call 'replace_node' here, because
1802 // that will yank the edge from old_phi to the Region and
1803 // I'm mid-iteration over the Region's uses.
1804 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1805 Node* use = old_phi->last_out(i);
1806 igvn.rehash_node_delayed(use);
1807 uint uses_found = 0;
1808 for (uint j = 0; j < use->len(); j++) {
1809 if (use->in(j) == old_phi) {
1810 if (j < use->req()) use->set_req (j, id_old_phi);
1811 else use->set_prec(j, id_old_phi);
1812 uses_found++;
1813 }
1814 }
1815 i -= uses_found; // we deleted 1 or more copies of this edge
1816 }
1817 }
1818 igvn._worklist.push(old_phi);
1819 }
1820 }
1821 // Finally clean out the fall-in edges from the RegionNode
1822 for( i = oreq-1; i>0; i-- ) {
1823 if( !phase->is_member( this, _head->in(i) ) ) {
1824 _head->del_req(i);
1825 }
1826 }
1827 igvn.rehash_node_delayed(_head);
1828 // Transform landing pad
1829 igvn.register_new_node_with_optimizer(landing_pad, _head);
1830 // Insert landing pad into the header
1831 _head->add_req(landing_pad);
1832 }
1833
1834 //------------------------------split_outer_loop-------------------------------
1835 // Split out the outermost loop from this shared header.
1836 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1837 PhaseIterGVN &igvn = phase->_igvn;
1838
1839 // Find index of outermost loop; it should also be my tail.
1840 uint outer_idx = 1;
1841 while( _head->in(outer_idx) != _tail ) outer_idx++;
1842
1843 // Make a LoopNode for the outermost loop.
1844 Node *ctl = _head->in(LoopNode::EntryControl);
1845 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1846 outer = igvn.register_new_node_with_optimizer(outer, _head);
1847 phase->set_created_loop_node();
1848
1849 // Outermost loop falls into '_head' loop
1850 _head->set_req(LoopNode::EntryControl, outer);
1851 _head->del_req(outer_idx);
1852 // Split all the Phis up between '_head' loop and 'outer' loop.
1853 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1854 Node *out = _head->fast_out(j);
1855 if( out->is_Phi() ) {
1856 PhiNode *old_phi = out->as_Phi();
1857 assert( old_phi->region() == _head, "" );
1858 Node *phi = PhiNode::make_blank(outer, old_phi);
1859 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1860 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1861 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1862 // Make old Phi point to new Phi on the fall-in path
1863 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1864 old_phi->del_req(outer_idx);
1865 }
1866 }
1867
1868 // Use the new loop head instead of the old shared one
1869 _head = outer;
1870 phase->set_loop(_head, this);
1871 }
1872
1873 //------------------------------fix_parent-------------------------------------
1874 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1875 loop->_parent = parent;
1876 if( loop->_child ) fix_parent( loop->_child, loop );
1877 if( loop->_next ) fix_parent( loop->_next , parent );
1878 }
1879
1880 //------------------------------estimate_path_freq-----------------------------
1881 static float estimate_path_freq( Node *n ) {
1882 // Try to extract some path frequency info
1883 IfNode *iff;
1884 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1885 uint nop = n->Opcode();
1886 if( nop == Op_SafePoint ) { // Skip any safepoint
1887 n = n->in(0);
1888 continue;
1889 }
1890 if( nop == Op_CatchProj ) { // Get count from a prior call
1891 // Assume call does not always throw exceptions: means the call-site
1892 // count is also the frequency of the fall-through path.
1893 assert( n->is_CatchProj(), "" );
1894 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1895 return 0.0f; // Assume call exception path is rare
1896 Node *call = n->in(0)->in(0)->in(0);
1897 assert( call->is_Call(), "expect a call here" );
1898 const JVMState *jvms = ((CallNode*)call)->jvms();
1899 ciMethodData* methodData = jvms->method()->method_data();
1900 if (!methodData->is_mature()) return 0.0f; // No call-site data
1901 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1902 if ((data == NULL) || !data->is_CounterData()) {
1903 // no call profile available, try call's control input
1904 n = n->in(0);
1905 continue;
1906 }
1907 return data->as_CounterData()->count()/FreqCountInvocations;
1908 }
1909 // See if there's a gating IF test
1910 Node *n_c = n->in(0);
1911 if( !n_c->is_If() ) break; // No estimate available
1912 iff = n_c->as_If();
1913 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1914 // Compute how much count comes on this path
1915 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1916 // Have no count info. Skip dull uncommon-trap like branches.
1917 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1918 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1919 break;
1920 // Skip through never-taken branch; look for a real loop exit.
1921 n = iff->in(0);
1922 }
1923 return 0.0f; // No estimate available
1924 }
1925
1926 //------------------------------merge_many_backedges---------------------------
1927 // Merge all the backedges from the shared header into a private Region.
1928 // Feed that region as the one backedge to this loop.
1929 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1930 uint i;
1931
1932 // Scan for the top 2 hottest backedges
1933 float hotcnt = 0.0f;
1934 float warmcnt = 0.0f;
1935 uint hot_idx = 0;
1936 // Loop starts at 2 because slot 1 is the fall-in path
1937 for( i = 2; i < _head->req(); i++ ) {
1938 float cnt = estimate_path_freq(_head->in(i));
1939 if( cnt > hotcnt ) { // Grab hottest path
1940 warmcnt = hotcnt;
1941 hotcnt = cnt;
1942 hot_idx = i;
1943 } else if( cnt > warmcnt ) { // And 2nd hottest path
1944 warmcnt = cnt;
1945 }
1946 }
1947
1948 // See if the hottest backedge is worthy of being an inner loop
1949 // by being much hotter than the next hottest backedge.
1950 if( hotcnt <= 0.0001 ||
1951 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1952
1953 // Peel out the backedges into a private merge point; peel
1954 // them all except optionally hot_idx.
1955 PhaseIterGVN &igvn = phase->_igvn;
1956
1957 Node *hot_tail = NULL;
1958 // Make a Region for the merge point
1959 Node *r = new RegionNode(1);
1960 for( i = 2; i < _head->req(); i++ ) {
1961 if( i != hot_idx )
1962 r->add_req( _head->in(i) );
1963 else hot_tail = _head->in(i);
1964 }
1965 igvn.register_new_node_with_optimizer(r, _head);
1966 // Plug region into end of loop _head, followed by hot_tail
1967 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1968 igvn.replace_input_of(_head, 2, r);
1969 if( hot_idx ) _head->add_req(hot_tail);
1970
1971 // Split all the Phis up between '_head' loop and the Region 'r'
1972 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1973 Node *out = _head->fast_out(j);
1974 if( out->is_Phi() ) {
1975 PhiNode* n = out->as_Phi();
1976 igvn.hash_delete(n); // Delete from hash before hacking edges
1977 Node *hot_phi = NULL;
1978 Node *phi = new PhiNode(r, n->type(), n->adr_type());
1979 // Check all inputs for the ones to peel out
1980 uint j = 1;
1981 for( uint i = 2; i < n->req(); i++ ) {
1982 if( i != hot_idx )
1983 phi->set_req( j++, n->in(i) );
1984 else hot_phi = n->in(i);
1985 }
1986 // Register the phi but do not transform until whole place transforms
1987 igvn.register_new_node_with_optimizer(phi, n);
1988 // Add the merge phi to the old Phi
1989 while( n->req() > 3 ) n->del_req( n->req()-1 );
1990 igvn.replace_input_of(n, 2, phi);
1991 if( hot_idx ) n->add_req(hot_phi);
1992 }
1993 }
1994
1995
1996 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1997 // of self loop tree. Turn self into a loop headed by _head and with
1998 // tail being the new merge point.
1999 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
2000 phase->set_loop(_tail,ilt); // Adjust tail
2001 _tail = r; // Self's tail is new merge point
2002 phase->set_loop(r,this);
2003 ilt->_child = _child; // New guy has my children
2004 _child = ilt; // Self has new guy as only child
2005 ilt->_parent = this; // new guy has self for parent
2006 ilt->_nest = _nest; // Same nesting depth (for now)
2007
2008 // Starting with 'ilt', look for child loop trees using the same shared
2009 // header. Flatten these out; they will no longer be loops in the end.
2010 IdealLoopTree **pilt = &_child;
2011 while( ilt ) {
2012 if( ilt->_head == _head ) {
2013 uint i;
2014 for( i = 2; i < _head->req(); i++ )
2015 if( _head->in(i) == ilt->_tail )
2016 break; // Still a loop
2017 if( i == _head->req() ) { // No longer a loop
2018 // Flatten ilt. Hang ilt's "_next" list from the end of
2019 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
2020 IdealLoopTree **cp = &ilt->_child;
2021 while( *cp ) cp = &(*cp)->_next; // Find end of child list
2022 *cp = ilt->_next; // Hang next list at end of child list
2023 *pilt = ilt->_child; // Move child up to replace ilt
2024 ilt->_head = NULL; // Flag as a loop UNIONED into parent
2025 ilt = ilt->_child; // Repeat using new ilt
2026 continue; // do not advance over ilt->_child
2027 }
2028 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
2029 phase->set_loop(_head,ilt);
2030 }
2031 pilt = &ilt->_child; // Advance to next
2032 ilt = *pilt;
2033 }
2034
2035 if( _child ) fix_parent( _child, this );
2036 }
2037
2038 //------------------------------beautify_loops---------------------------------
2039 // Split shared headers and insert loop landing pads.
2040 // Insert a LoopNode to replace the RegionNode.
2041 // Return TRUE if loop tree is structurally changed.
2042 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
2043 bool result = false;
2044 // Cache parts in locals for easy
2045 PhaseIterGVN &igvn = phase->_igvn;
2046
2047 igvn.hash_delete(_head); // Yank from hash before hacking edges
2048
2049 // Check for multiple fall-in paths. Peel off a landing pad if need be.
2050 int fall_in_cnt = 0;
2051 for( uint i = 1; i < _head->req(); i++ )
2052 if( !phase->is_member( this, _head->in(i) ) )
2053 fall_in_cnt++;
2054 assert( fall_in_cnt, "at least 1 fall-in path" );
2055 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
2056 split_fall_in( phase, fall_in_cnt );
2057
2058 // Swap inputs to the _head and all Phis to move the fall-in edge to
2059 // the left.
2060 fall_in_cnt = 1;
2061 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
2062 fall_in_cnt++;
2063 if( fall_in_cnt > 1 ) {
2064 // Since I am just swapping inputs I do not need to update def-use info
2065 Node *tmp = _head->in(1);
2066 igvn.rehash_node_delayed(_head);
2067 _head->set_req( 1, _head->in(fall_in_cnt) );
2068 _head->set_req( fall_in_cnt, tmp );
2069 // Swap also all Phis
2070 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
2071 Node* phi = _head->fast_out(i);
2072 if( phi->is_Phi() ) {
2073 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
2074 tmp = phi->in(1);
2075 phi->set_req( 1, phi->in(fall_in_cnt) );
2076 phi->set_req( fall_in_cnt, tmp );
2077 }
2078 }
2079 }
2080 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
2081 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
2082
2083 // If I am a shared header (multiple backedges), peel off the many
2084 // backedges into a private merge point and use the merge point as
2085 // the one true backedge.
2086 if( _head->req() > 3 ) {
2087 // Merge the many backedges into a single backedge but leave
2088 // the hottest backedge as separate edge for the following peel.
2089 merge_many_backedges( phase );
2090 result = true;
2091 }
2092
2093 // If I have one hot backedge, peel off myself loop.
2094 // I better be the outermost loop.
2095 if (_head->req() > 3 && !_irreducible) {
2096 split_outer_loop( phase );
2097 result = true;
2098
2099 } else if (!_head->is_Loop() && !_irreducible) {
2100 // Make a new LoopNode to replace the old loop head
2101 Node *l = new LoopNode( _head->in(1), _head->in(2) );
2102 l = igvn.register_new_node_with_optimizer(l, _head);
2103 phase->set_created_loop_node();
2104 // Go ahead and replace _head
2105 phase->_igvn.replace_node( _head, l );
2106 _head = l;
2107 phase->set_loop(_head, this);
2108 }
2109
2110 // Now recursively beautify nested loops
2111 if( _child ) result |= _child->beautify_loops( phase );
2112 if( _next ) result |= _next ->beautify_loops( phase );
2113 return result;
2114 }
2115
2116 //------------------------------allpaths_check_safepts----------------------------
2117 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
2118 // encountered. Helper for check_safepts.
2119 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
2120 assert(stack.size() == 0, "empty stack");
2121 stack.push(_tail);
2122 visited.Clear();
2123 visited.set(_tail->_idx);
2124 while (stack.size() > 0) {
2125 Node* n = stack.pop();
2126 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2127 // Terminate this path
2128 } else if (n->Opcode() == Op_SafePoint) {
2129 if (_phase->get_loop(n) != this) {
2130 if (_required_safept == NULL) _required_safept = new Node_List();
2131 _required_safept->push(n); // save the one closest to the tail
2132 }
2133 // Terminate this path
2134 } else {
2135 uint start = n->is_Region() ? 1 : 0;
2136 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
2137 for (uint i = start; i < end; i++) {
2138 Node* in = n->in(i);
2139 assert(in->is_CFG(), "must be");
2140 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
2141 stack.push(in);
2142 }
2143 }
2144 }
2145 }
2146 }
2147
2148 //------------------------------check_safepts----------------------------
2149 // Given dominators, try to find loops with calls that must always be
2150 // executed (call dominates loop tail). These loops do not need non-call
2151 // safepoints (ncsfpt).
2152 //
2153 // A complication is that a safepoint in a inner loop may be needed
2154 // by an outer loop. In the following, the inner loop sees it has a
2155 // call (block 3) on every path from the head (block 2) to the
2156 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
2157 // in block 2, _but_ this leaves the outer loop without a safepoint.
2158 //
2159 // entry 0
2160 // |
2161 // v
2162 // outer 1,2 +->1
2163 // | |
2164 // | v
2165 // | 2<---+ ncsfpt in 2
2166 // |_/|\ |
2167 // | v |
2168 // inner 2,3 / 3 | call in 3
2169 // / | |
2170 // v +--+
2171 // exit 4
2172 //
2173 //
2174 // This method creates a list (_required_safept) of ncsfpt nodes that must
2175 // be protected is created for each loop. When a ncsfpt maybe deleted, it
2176 // is first looked for in the lists for the outer loops of the current loop.
2177 //
2178 // The insights into the problem:
2179 // A) counted loops are okay
2180 // B) innermost loops are okay (only an inner loop can delete
2181 // a ncsfpt needed by an outer loop)
2182 // C) a loop is immune from an inner loop deleting a safepoint
2183 // if the loop has a call on the idom-path
2184 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
2185 // idom-path that is not in a nested loop
2186 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
2187 // loop needs to be prevented from deletion by an inner loop
2188 //
2189 // There are two analyses:
2190 // 1) The first, and cheaper one, scans the loop body from
2191 // tail to head following the idom (immediate dominator)
2192 // chain, looking for the cases (C,D,E) above.
2193 // Since inner loops are scanned before outer loops, there is summary
2194 // information about inner loops. Inner loops can be skipped over
2195 // when the tail of an inner loop is encountered.
2196 //
2197 // 2) The second, invoked if the first fails to find a call or ncsfpt on
2198 // the idom path (which is rare), scans all predecessor control paths
2199 // from the tail to the head, terminating a path when a call or sfpt
2200 // is encountered, to find the ncsfpt's that are closest to the tail.
2201 //
2202 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
2203 // Bottom up traversal
2204 IdealLoopTree* ch = _child;
2205 if (_child) _child->check_safepts(visited, stack);
2206 if (_next) _next ->check_safepts(visited, stack);
2207
2208 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
2209 bool has_call = false; // call on dom-path
2210 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
2211 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
2212 // Scan the dom-path nodes from tail to head
2213 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
2214 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2215 has_call = true;
2216 _has_sfpt = 1; // Then no need for a safept!
2217 break;
2218 } else if (n->Opcode() == Op_SafePoint) {
2219 if (_phase->get_loop(n) == this) {
2220 has_local_ncsfpt = true;
2221 break;
2222 }
2223 if (nonlocal_ncsfpt == NULL) {
2224 nonlocal_ncsfpt = n; // save the one closest to the tail
2225 }
2226 } else {
2227 IdealLoopTree* nlpt = _phase->get_loop(n);
2228 if (this != nlpt) {
2229 // If at an inner loop tail, see if the inner loop has already
2230 // recorded seeing a call on the dom-path (and stop.) If not,
2231 // jump to the head of the inner loop.
2232 assert(is_member(nlpt), "nested loop");
2233 Node* tail = nlpt->_tail;
2234 if (tail->in(0)->is_If()) tail = tail->in(0);
2235 if (n == tail) {
2236 // If inner loop has call on dom-path, so does outer loop
2237 if (nlpt->_has_sfpt) {
2238 has_call = true;
2239 _has_sfpt = 1;
2240 break;
2241 }
2242 // Skip to head of inner loop
2243 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
2244 n = nlpt->_head;
2245 }
2246 }
2247 }
2248 }
2249 // Record safept's that this loop needs preserved when an
2250 // inner loop attempts to delete it's safepoints.
2251 if (_child != NULL && !has_call && !has_local_ncsfpt) {
2252 if (nonlocal_ncsfpt != NULL) {
2253 if (_required_safept == NULL) _required_safept = new Node_List();
2254 _required_safept->push(nonlocal_ncsfpt);
2255 } else {
2256 // Failed to find a suitable safept on the dom-path. Now use
2257 // an all paths walk from tail to head, looking for safepoints to preserve.
2258 allpaths_check_safepts(visited, stack);
2259 }
2260 }
2261 }
2262 }
2263
2264 //---------------------------is_deleteable_safept----------------------------
2265 // Is safept not required by an outer loop?
2266 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
2267 assert(sfpt->Opcode() == Op_SafePoint, "");
2268 IdealLoopTree* lp = get_loop(sfpt)->_parent;
2269 while (lp != NULL) {
2270 Node_List* sfpts = lp->_required_safept;
2271 if (sfpts != NULL) {
2272 for (uint i = 0; i < sfpts->size(); i++) {
2273 if (sfpt == sfpts->at(i))
2274 return false;
2275 }
2276 }
2277 lp = lp->_parent;
2278 }
2279 return true;
2280 }
2281
2282 //---------------------------replace_parallel_iv-------------------------------
2283 // Replace parallel induction variable (parallel to trip counter)
2284 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
2285 assert(loop->_head->is_CountedLoop(), "");
2286 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2287 if (!cl->is_valid_counted_loop())
2288 return; // skip malformed counted loop
2289 Node *incr = cl->incr();
2290 if (incr == NULL)
2291 return; // Dead loop?
2292 Node *init = cl->init_trip();
2293 Node *phi = cl->phi();
2294 int stride_con = cl->stride_con();
2295
2296 // Visit all children, looking for Phis
2297 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
2298 Node *out = cl->out(i);
2299 // Look for other phis (secondary IVs). Skip dead ones
2300 if (!out->is_Phi() || out == phi || !has_node(out))
2301 continue;
2302 PhiNode* phi2 = out->as_Phi();
2303 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
2304 // Look for induction variables of the form: X += constant
2305 if (phi2->region() != loop->_head ||
2306 incr2->req() != 3 ||
2307 incr2->in(1) != phi2 ||
2308 incr2 == incr ||
2309 incr2->Opcode() != Op_AddI ||
2310 !incr2->in(2)->is_Con())
2311 continue;
2312
2313 // Check for parallel induction variable (parallel to trip counter)
2314 // via an affine function. In particular, count-down loops with
2315 // count-up array indices are common. We only RCE references off
2316 // the trip-counter, so we need to convert all these to trip-counter
2317 // expressions.
2318 Node *init2 = phi2->in( LoopNode::EntryControl );
2319 int stride_con2 = incr2->in(2)->get_int();
2320
2321 // The ratio of the two strides cannot be represented as an int
2322 // if stride_con2 is min_int and stride_con is -1.
2323 if (stride_con2 == min_jint && stride_con == -1) {
2324 continue;
2325 }
2326
2327 // The general case here gets a little tricky. We want to find the
2328 // GCD of all possible parallel IV's and make a new IV using this
2329 // GCD for the loop. Then all possible IVs are simple multiples of
2330 // the GCD. In practice, this will cover very few extra loops.
2331 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
2332 // where +/-1 is the common case, but other integer multiples are
2333 // also easy to handle.
2334 int ratio_con = stride_con2/stride_con;
2335
2336 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
2337 #ifndef PRODUCT
2338 if (TraceLoopOpts) {
2339 tty->print("Parallel IV: %d ", phi2->_idx);
2340 loop->dump_head();
2341 }
2342 #endif
2343 // Convert to using the trip counter. The parallel induction
2344 // variable differs from the trip counter by a loop-invariant
2345 // amount, the difference between their respective initial values.
2346 // It is scaled by the 'ratio_con'.
2347 Node* ratio = _igvn.intcon(ratio_con);
2348 set_ctrl(ratio, C->root());
2349 Node* ratio_init = new MulINode(init, ratio);
2350 _igvn.register_new_node_with_optimizer(ratio_init, init);
2351 set_early_ctrl(ratio_init);
2352 Node* diff = new SubINode(init2, ratio_init);
2353 _igvn.register_new_node_with_optimizer(diff, init2);
2354 set_early_ctrl(diff);
2355 Node* ratio_idx = new MulINode(phi, ratio);
2356 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
2357 set_ctrl(ratio_idx, cl);
2358 Node* add = new AddINode(ratio_idx, diff);
2359 _igvn.register_new_node_with_optimizer(add);
2360 set_ctrl(add, cl);
2361 _igvn.replace_node( phi2, add );
2362 // Sometimes an induction variable is unused
2363 if (add->outcnt() == 0) {
2364 _igvn.remove_dead_node(add);
2365 }
2366 --i; // deleted this phi; rescan starting with next position
2367 continue;
2368 }
2369 }
2370 }
2371
2372 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
2373 Node* keep = NULL;
2374 if (keep_one) {
2375 // Look for a safepoint on the idom-path.
2376 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
2377 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
2378 keep = i;
2379 break; // Found one
2380 }
2381 }
2382 }
2383
2384 // Don't remove any safepoints if it is requested to keep a single safepoint and
2385 // no safepoint was found on idom-path. It is not safe to remove any safepoint
2386 // in this case since there's no safepoint dominating all paths in the loop body.
2387 bool prune = !keep_one || keep != NULL;
2388
2389 // Delete other safepoints in this loop.
2390 Node_List* sfpts = _safepts;
2391 if (prune && sfpts != NULL) {
2392 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
2393 for (uint i = 0; i < sfpts->size(); i++) {
2394 Node* n = sfpts->at(i);
2395 assert(phase->get_loop(n) == this, "");
2396 if (n != keep && phase->is_deleteable_safept(n)) {
2397 phase->lazy_replace(n, n->in(TypeFunc::Control));
2398 }
2399 }
2400 }
2401 }
2402
2403 //------------------------------counted_loop-----------------------------------
2404 // Convert to counted loops where possible
2405 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
2406
2407 // For grins, set the inner-loop flag here
2408 if (!_child) {
2409 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
2410 }
2411
2412 IdealLoopTree* loop = this;
2413 if (_head->is_CountedLoop() ||
2414 phase->is_counted_loop(_head, loop)) {
2415
2416 if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) {
2417 // Indicate we do not need a safepoint here
2418 _has_sfpt = 1;
2419 }
2420
2421 // Remove safepoints
2422 bool keep_one_sfpt = !(_has_call || _has_sfpt);
2423 remove_safepoints(phase, keep_one_sfpt);
2424
2425 // Look for induction variables
2426 phase->replace_parallel_iv(this);
2427
2428 } else if (_parent != NULL && !_irreducible) {
2429 // Not a counted loop. Keep one safepoint.
2430 bool keep_one_sfpt = true;
2431 remove_safepoints(phase, keep_one_sfpt);
2432 }
2433
2434 // Recursively
2435 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
2436 assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops");
2437 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
2438 if (loop->_next) loop->_next ->counted_loop(phase);
2439 }
2440
2441 #ifndef PRODUCT
2442 //------------------------------dump_head--------------------------------------
2443 // Dump 1 liner for loop header info
2444 void IdealLoopTree::dump_head( ) const {
2445 for (uint i=0; i<_nest; i++)
2446 tty->print(" ");
2447 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
2448 if (_irreducible) tty->print(" IRREDUCIBLE");
2449 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl);
2450 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
2451 if (predicate != NULL ) {
2452 tty->print(" limit_check");
2453 entry = PhaseIdealLoop::skip_loop_predicates(entry);
2454 }
2455 if (UseLoopPredicate) {
2456 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2457 if (entry != NULL) {
2458 tty->print(" predicated");
2459 entry = PhaseIdealLoop::skip_loop_predicates(entry);
2460 }
2461 }
2462 if (UseProfiledLoopPredicate) {
2463 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
2464 if (entry != NULL) {
2465 tty->print(" profile_predicated");
2466 }
2467 }
2468 if (_head->is_CountedLoop()) {
2469 CountedLoopNode *cl = _head->as_CountedLoop();
2470 tty->print(" counted");
2471
2472 Node* init_n = cl->init_trip();
2473 if (init_n != NULL && init_n->is_Con())
2474 tty->print(" [%d,", cl->init_trip()->get_int());
2475 else
2476 tty->print(" [int,");
2477 Node* limit_n = cl->limit();
2478 if (limit_n != NULL && limit_n->is_Con())
2479 tty->print("%d),", cl->limit()->get_int());
2480 else
2481 tty->print("int),");
2482 int stride_con = cl->stride_con();
2483 if (stride_con > 0) tty->print("+");
2484 tty->print("%d", stride_con);
2485
2486 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
2487
2488 if (cl->is_pre_loop ()) tty->print(" pre" );
2489 if (cl->is_main_loop()) tty->print(" main");
2490 if (cl->is_post_loop()) tty->print(" post");
2491 if (cl->is_vectorized_loop()) tty->print(" vector");
2492 if (cl->range_checks_present()) tty->print(" rc ");
2493 if (cl->is_multiversioned()) tty->print(" multi ");
2494 }
2495 if (_has_call) tty->print(" has_call");
2496 if (_has_sfpt) tty->print(" has_sfpt");
2497 if (_rce_candidate) tty->print(" rce");
2498 if (_safepts != NULL && _safepts->size() > 0) {
2499 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
2500 }
2501 if (_required_safept != NULL && _required_safept->size() > 0) {
2502 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
2503 }
2504 if (Verbose) {
2505 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
2506 }
2507 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
2508 tty->print(" strip_mined");
2509 }
2510 tty->cr();
2511 }
2512
2513 //------------------------------dump-------------------------------------------
2514 // Dump loops by loop tree
2515 void IdealLoopTree::dump( ) const {
2516 dump_head();
2517 if (_child) _child->dump();
2518 if (_next) _next ->dump();
2519 }
2520
2521 #endif
2522
2523 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
2524 if (loop == root) {
2525 if (loop->_child != NULL) {
2526 log->begin_head("loop_tree");
2527 log->end_head();
2528 if( loop->_child ) log_loop_tree(root, loop->_child, log);
2529 log->tail("loop_tree");
2530 assert(loop->_next == NULL, "what?");
2531 }
2532 } else {
2533 Node* head = loop->_head;
2534 log->begin_head("loop");
2535 log->print(" idx='%d' ", head->_idx);
2536 if (loop->_irreducible) log->print("irreducible='1' ");
2537 if (head->is_Loop()) {
2538 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
2539 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
2540 }
2541 if (head->is_CountedLoop()) {
2542 CountedLoopNode* cl = head->as_CountedLoop();
2543 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
2544 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
2545 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
2546 }
2547 log->end_head();
2548 if( loop->_child ) log_loop_tree(root, loop->_child, log);
2549 log->tail("loop");
2550 if( loop->_next ) log_loop_tree(root, loop->_next, log);
2551 }
2552 }
2553
2554 //---------------------collect_potentially_useful_predicates-----------------------
2555 // Helper function to collect potentially useful predicates to prevent them from
2556 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
2557 void PhaseIdealLoop::collect_potentially_useful_predicates(
2558 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
2559 if (loop->_child) { // child
2560 collect_potentially_useful_predicates(loop->_child, useful_predicates);
2561 }
2562
2563 // self (only loops that we can apply loop predication may use their predicates)
2564 if (loop->_head->is_Loop() &&
2565 !loop->_irreducible &&
2566 !loop->tail()->is_top()) {
2567 LoopNode* lpn = loop->_head->as_Loop();
2568 Node* entry = lpn->in(LoopNode::EntryControl);
2569 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2570 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
2571 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2572 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2573 entry = skip_loop_predicates(entry);
2574 }
2575 predicate_proj = find_predicate(entry); // Predicate
2576 if (predicate_proj != NULL ) {
2577 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2578 entry = skip_loop_predicates(entry);
2579 }
2580 if (UseProfiledLoopPredicate) {
2581 predicate_proj = find_predicate(entry); // Predicate
2582 if (predicate_proj != NULL ) {
2583 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2584 }
2585 }
2586 }
2587
2588 if (loop->_next) { // sibling
2589 collect_potentially_useful_predicates(loop->_next, useful_predicates);
2590 }
2591 }
2592
2593 //------------------------eliminate_useless_predicates-----------------------------
2594 // Eliminate all inserted predicates if they could not be used by loop predication.
2595 // Note: it will also eliminates loop limits check predicate since it also uses
2596 // Opaque1 node (see Parse::add_predicate()).
2597 void PhaseIdealLoop::eliminate_useless_predicates() {
2598 if (C->predicate_count() == 0)
2599 return; // no predicate left
2600
2601 Unique_Node_List useful_predicates; // to store useful predicates
2602 if (C->has_loops()) {
2603 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2604 }
2605
2606 for (int i = C->predicate_count(); i > 0; i--) {
2607 Node * n = C->predicate_opaque1_node(i-1);
2608 assert(n->Opcode() == Op_Opaque1, "must be");
2609 if (!useful_predicates.member(n)) { // not in the useful list
2610 _igvn.replace_node(n, n->in(1));
2611 }
2612 }
2613 }
2614
2615 //------------------------process_expensive_nodes-----------------------------
2616 // Expensive nodes have their control input set to prevent the GVN
2617 // from commoning them and as a result forcing the resulting node to
2618 // be in a more frequent path. Use CFG information here, to change the
2619 // control inputs so that some expensive nodes can be commoned while
2620 // not executed more frequently.
2621 bool PhaseIdealLoop::process_expensive_nodes() {
2622 assert(OptimizeExpensiveOps, "optimization off?");
2623
2624 // Sort nodes to bring similar nodes together
2625 C->sort_expensive_nodes();
2626
2627 bool progress = false;
2628
2629 for (int i = 0; i < C->expensive_count(); ) {
2630 Node* n = C->expensive_node(i);
2631 int start = i;
2632 // Find nodes similar to n
2633 i++;
2634 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2635 int end = i;
2636 // And compare them two by two
2637 for (int j = start; j < end; j++) {
2638 Node* n1 = C->expensive_node(j);
2639 if (is_node_unreachable(n1)) {
2640 continue;
2641 }
2642 for (int k = j+1; k < end; k++) {
2643 Node* n2 = C->expensive_node(k);
2644 if (is_node_unreachable(n2)) {
2645 continue;
2646 }
2647
2648 assert(n1 != n2, "should be pair of nodes");
2649
2650 Node* c1 = n1->in(0);
2651 Node* c2 = n2->in(0);
2652
2653 Node* parent_c1 = c1;
2654 Node* parent_c2 = c2;
2655
2656 // The call to get_early_ctrl_for_expensive() moves the
2657 // expensive nodes up but stops at loops that are in a if
2658 // branch. See whether we can exit the loop and move above the
2659 // If.
2660 if (c1->is_Loop()) {
2661 parent_c1 = c1->in(1);
2662 }
2663 if (c2->is_Loop()) {
2664 parent_c2 = c2->in(1);
2665 }
2666
2667 if (parent_c1 == parent_c2) {
2668 _igvn._worklist.push(n1);
2669 _igvn._worklist.push(n2);
2670 continue;
2671 }
2672
2673 // Look for identical expensive node up the dominator chain.
2674 if (is_dominator(c1, c2)) {
2675 c2 = c1;
2676 } else if (is_dominator(c2, c1)) {
2677 c1 = c2;
2678 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2679 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2680 // Both branches have the same expensive node so move it up
2681 // before the if.
2682 c1 = c2 = idom(parent_c1->in(0));
2683 }
2684 // Do the actual moves
2685 if (n1->in(0) != c1) {
2686 _igvn.hash_delete(n1);
2687 n1->set_req(0, c1);
2688 _igvn.hash_insert(n1);
2689 _igvn._worklist.push(n1);
2690 progress = true;
2691 }
2692 if (n2->in(0) != c2) {
2693 _igvn.hash_delete(n2);
2694 n2->set_req(0, c2);
2695 _igvn.hash_insert(n2);
2696 _igvn._worklist.push(n2);
2697 progress = true;
2698 }
2699 }
2700 }
2701 }
2702
2703 return progress;
2704 }
2705
2706
2707 //=============================================================================
2708 //----------------------------build_and_optimize-------------------------------
2709 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2710 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2711 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
2712 bool do_split_ifs = (mode == LoopOptsDefault || mode == LoopOptsLastRound);
2713 bool skip_loop_opts = (mode == LoopOptsNone);
2714
2715 int old_progress = C->major_progress();
2716 uint orig_worklist_size = _igvn._worklist.size();
2717
2718 // Reset major-progress flag for the driver's heuristics
2719 C->clear_major_progress();
2720
2721 #ifndef PRODUCT
2722 // Capture for later assert
2723 uint unique = C->unique();
2724 _loop_invokes++;
2725 _loop_work += unique;
2726 #endif
2727
2728 // True if the method has at least 1 irreducible loop
2729 _has_irreducible_loops = false;
2730
2731 _created_loop_node = false;
2732
2733 Arena *a = Thread::current()->resource_area();
2734 VectorSet visited(a);
2735 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2736 _nodes.map(C->unique(), NULL);
2737 memset(_nodes.adr(), 0, wordSize * C->unique());
2738
2739 // Pre-build the top-level outermost loop tree entry
2740 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2741 // Do not need a safepoint at the top level
2742 _ltree_root->_has_sfpt = 1;
2743
2744 // Initialize Dominators.
2745 // Checked in clone_loop_predicate() during beautify_loops().
2746 _idom_size = 0;
2747 _idom = NULL;
2748 _dom_depth = NULL;
2749 _dom_stk = NULL;
2750
2751 // Empty pre-order array
2752 allocate_preorders();
2753
2754 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2755 // IdealLoopTree entries. Data nodes are NOT walked.
2756 build_loop_tree();
2757 // Check for bailout, and return
2758 if (C->failing()) {
2759 return;
2760 }
2761
2762 // No loops after all
2763 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2764
2765 // There should always be an outer loop containing the Root and Return nodes.
2766 // If not, we have a degenerate empty program. Bail out in this case.
2767 if (!has_node(C->root())) {
2768 if (!_verify_only) {
2769 C->clear_major_progress();
2770 C->record_method_not_compilable("empty program detected during loop optimization");
2771 }
2772 return;
2773 }
2774
2775 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2776 // Nothing to do, so get out
2777 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only &&
2778 !bs->is_gc_specific_loop_opts_pass(mode);
2779 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2780 bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(mode);
2781 if (stop_early && !do_expensive_nodes) {
2782 _igvn.optimize(); // Cleanup NeverBranches
2783 return;
2784 }
2785
2786 // Set loop nesting depth
2787 _ltree_root->set_nest( 0 );
2788
2789 // Split shared headers and insert loop landing pads.
2790 // Do not bother doing this on the Root loop of course.
2791 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2792 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2793 if( _ltree_root->_child->beautify_loops( this ) ) {
2794 // Re-build loop tree!
2795 _ltree_root->_child = NULL;
2796 _nodes.clear();
2797 reallocate_preorders();
2798 build_loop_tree();
2799 // Check for bailout, and return
2800 if (C->failing()) {
2801 return;
2802 }
2803 // Reset loop nesting depth
2804 _ltree_root->set_nest( 0 );
2805
2806 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2807 }
2808 }
2809
2810 // Build Dominators for elision of NULL checks & loop finding.
2811 // Since nodes do not have a slot for immediate dominator, make
2812 // a persistent side array for that info indexed on node->_idx.
2813 _idom_size = C->unique();
2814 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2815 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2816 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2817 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2818
2819 Dominators();
2820
2821 if (!_verify_only) {
2822 // As a side effect, Dominators removed any unreachable CFG paths
2823 // into RegionNodes. It doesn't do this test against Root, so
2824 // we do it here.
2825 for( uint i = 1; i < C->root()->req(); i++ ) {
2826 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2827 _igvn.delete_input_of(C->root(), i);
2828 i--; // Rerun same iteration on compressed edges
2829 }
2830 }
2831
2832 // Given dominators, try to find inner loops with calls that must
2833 // always be executed (call dominates loop tail). These loops do
2834 // not need a separate safepoint.
2835 Node_List cisstack(a);
2836 _ltree_root->check_safepts(visited, cisstack);
2837 }
2838
2839 // Walk the DATA nodes and place into loops. Find earliest control
2840 // node. For CFG nodes, the _nodes array starts out and remains
2841 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2842 // _nodes array holds the earliest legal controlling CFG node.
2843
2844 // Allocate stack with enough space to avoid frequent realloc
2845 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2846 Node_Stack nstack( a, stack_size );
2847
2848 visited.Clear();
2849 Node_List worklist(a);
2850 // Don't need C->root() on worklist since
2851 // it will be processed among C->top() inputs
2852 worklist.push( C->top() );
2853 visited.set( C->top()->_idx ); // Set C->top() as visited now
2854 build_loop_early( visited, worklist, nstack );
2855
2856 // Given early legal placement, try finding counted loops. This placement
2857 // is good enough to discover most loop invariants.
2858 if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
2859 _ltree_root->counted_loop( this );
2860 }
2861
2862 // Find latest loop placement. Find ideal loop placement.
2863 visited.Clear();
2864 init_dom_lca_tags();
2865 // Need C->root() on worklist when processing outs
2866 worklist.push( C->root() );
2867 NOT_PRODUCT( C->verify_graph_edges(); )
2868 worklist.push( C->top() );
2869 build_loop_late( visited, worklist, nstack );
2870
2871 if (_verify_only) {
2872 // restore major progress flag
2873 for (int i = 0; i < old_progress; i++)
2874 C->set_major_progress();
2875 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2876 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2877 return;
2878 }
2879
2880 // clear out the dead code after build_loop_late
2881 while (_deadlist.size()) {
2882 _igvn.remove_globally_dead_node(_deadlist.pop());
2883 }
2884
2885 if (stop_early) {
2886 assert(do_expensive_nodes, "why are we here?");
2887 if (process_expensive_nodes()) {
2888 // If we made some progress when processing expensive nodes then
2889 // the IGVN may modify the graph in a way that will allow us to
2890 // make some more progress: we need to try processing expensive
2891 // nodes again.
2892 C->set_major_progress();
2893 }
2894 _igvn.optimize();
2895 return;
2896 }
2897
2898 // Some parser-inserted loop predicates could never be used by loop
2899 // predication or they were moved away from loop during some optimizations.
2900 // For example, peeling. Eliminate them before next loop optimizations.
2901 eliminate_useless_predicates();
2902
2903 #ifndef PRODUCT
2904 C->verify_graph_edges();
2905 if (_verify_me) { // Nested verify pass?
2906 // Check to see if the verify mode is broken
2907 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2908 return;
2909 }
2910 if(VerifyLoopOptimizations) verify();
2911 if(TraceLoopOpts && C->has_loops()) {
2912 _ltree_root->dump();
2913 }
2914 #endif
2915
2916 if (skip_loop_opts) {
2917 // restore major progress flag
2918 for (int i = 0; i < old_progress; i++) {
2919 C->set_major_progress();
2920 }
2921
2922 // Cleanup any modified bits
2923 _igvn.optimize();
2924
2925 if (C->log() != NULL) {
2926 log_loop_tree(_ltree_root, _ltree_root, C->log());
2927 }
2928 return;
2929 }
2930
2931 if (bs->optimize_loops(this, mode, visited, nstack, worklist)) {
2932 _igvn.optimize();
2933 if (C->log() != NULL) {
2934 log_loop_tree(_ltree_root, _ltree_root, C->log());
2935 }
2936 return;
2937 }
2938
2939 if (ReassociateInvariants) {
2940 // Reassociate invariants and prep for split_thru_phi
2941 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2942 IdealLoopTree* lpt = iter.current();
2943 bool is_counted = lpt->is_counted();
2944 if (!is_counted || !lpt->is_inner()) continue;
2945
2946 // check for vectorized loops, any reassociation of invariants was already done
2947 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue;
2948
2949 lpt->reassociate_invariants(this);
2950
2951 // Because RCE opportunities can be masked by split_thru_phi,
2952 // look for RCE candidates and inhibit split_thru_phi
2953 // on just their loop-phi's for this pass of loop opts
2954 if (SplitIfBlocks && do_split_ifs) {
2955 if (lpt->policy_range_check(this)) {
2956 lpt->_rce_candidate = 1; // = true
2957 }
2958 }
2959 }
2960 }
2961
2962 // Check for aggressive application of split-if and other transforms
2963 // that require basic-block info (like cloning through Phi's)
2964 if( SplitIfBlocks && do_split_ifs ) {
2965 visited.Clear();
2966 split_if_with_blocks( visited, nstack, mode == LoopOptsLastRound );
2967 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2968 if (mode == LoopOptsLastRound) {
2969 C->set_major_progress();
2970 }
2971 }
2972
2973 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2974 C->set_major_progress();
2975 }
2976
2977 // Perform loop predication before iteration splitting
2978 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2979 _ltree_root->_child->loop_predication(this);
2980 }
2981
2982 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2983 if (do_intrinsify_fill()) {
2984 C->set_major_progress();
2985 }
2986 }
2987
2988 // Perform iteration-splitting on inner loops. Split iterations to avoid
2989 // range checks or one-shot null checks.
2990
2991 // If split-if's didn't hack the graph too bad (no CFG changes)
2992 // then do loop opts.
2993 if (C->has_loops() && !C->major_progress()) {
2994 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2995 _ltree_root->_child->iteration_split( this, worklist );
2996 // No verify after peeling! GCM has hoisted code out of the loop.
2997 // After peeling, the hoisted code could sink inside the peeled area.
2998 // The peeling code does not try to recompute the best location for
2999 // all the code before the peeled area, so the verify pass will always
3000 // complain about it.
3001 }
3002 // Do verify graph edges in any case
3003 NOT_PRODUCT( C->verify_graph_edges(); );
3004
3005 if (!do_split_ifs) {
3006 // We saw major progress in Split-If to get here. We forced a
3007 // pass with unrolling and not split-if, however more split-if's
3008 // might make progress. If the unrolling didn't make progress
3009 // then the major-progress flag got cleared and we won't try
3010 // another round of Split-If. In particular the ever-common
3011 // instance-of/check-cast pattern requires at least 2 rounds of
3012 // Split-If to clear out.
3013 C->set_major_progress();
3014 }
3015
3016 // Repeat loop optimizations if new loops were seen
3017 if (created_loop_node()) {
3018 C->set_major_progress();
3019 }
3020
3021 // Keep loop predicates and perform optimizations with them
3022 // until no more loop optimizations could be done.
3023 // After that switch predicates off and do more loop optimizations.
3024 if (!C->major_progress() && (C->predicate_count() > 0)) {
3025 C->cleanup_loop_predicates(_igvn);
3026 if (TraceLoopOpts) {
3027 tty->print_cr("PredicatesOff");
3028 }
3029 C->set_major_progress();
3030 }
3031
3032 // Convert scalar to superword operations at the end of all loop opts.
3033 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
3034 // SuperWord transform
3035 SuperWord sw(this);
3036 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3037 IdealLoopTree* lpt = iter.current();
3038 if (lpt->is_counted()) {
3039 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
3040
3041 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
3042 // Check that the rce'd post loop is encountered first, multiversion after all
3043 // major main loop optimization are concluded
3044 if (!C->major_progress()) {
3045 IdealLoopTree *lpt_next = lpt->_next;
3046 if (lpt_next && lpt_next->is_counted()) {
3047 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
3048 has_range_checks(lpt_next);
3049 if (cl->is_post_loop() && cl->range_checks_present()) {
3050 if (!cl->is_multiversioned()) {
3051 if (multi_version_post_loops(lpt, lpt_next) == false) {
3052 // Cause the rce loop to be optimized away if we fail
3053 cl->mark_is_multiversioned();
3054 cl->set_slp_max_unroll(0);
3055 poison_rce_post_loop(lpt);
3056 }
3057 }
3058 }
3059 }
3060 sw.transform_loop(lpt, true);
3061 }
3062 } else if (cl->is_main_loop()) {
3063 sw.transform_loop(lpt, true);
3064 }
3065 }
3066 }
3067 }
3068
3069 // Cleanup any modified bits
3070 _igvn.optimize();
3071
3072 // disable assert until issue with split_flow_path is resolved (6742111)
3073 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
3074 // "shouldn't introduce irreducible loops");
3075
3076 if (C->log() != NULL) {
3077 log_loop_tree(_ltree_root, _ltree_root, C->log());
3078 }
3079 }
3080
3081 #ifndef PRODUCT
3082 //------------------------------print_statistics-------------------------------
3083 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
3084 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
3085 void PhaseIdealLoop::print_statistics() {
3086 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
3087 }
3088
3089 //------------------------------verify-----------------------------------------
3090 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
3091 static int fail; // debug only, so its multi-thread dont care
3092 void PhaseIdealLoop::verify() const {
3093 int old_progress = C->major_progress();
3094 ResourceMark rm;
3095 PhaseIdealLoop loop_verify( _igvn, this );
3096 VectorSet visited(Thread::current()->resource_area());
3097
3098 fail = 0;
3099 verify_compare( C->root(), &loop_verify, visited );
3100 assert( fail == 0, "verify loops failed" );
3101 // Verify loop structure is the same
3102 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
3103 // Reset major-progress. It was cleared by creating a verify version of
3104 // PhaseIdealLoop.
3105 for( int i=0; i<old_progress; i++ )
3106 C->set_major_progress();
3107 }
3108
3109 //------------------------------verify_compare---------------------------------
3110 // Make sure me and the given PhaseIdealLoop agree on key data structures
3111 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
3112 if( !n ) return;
3113 if( visited.test_set( n->_idx ) ) return;
3114 if( !_nodes[n->_idx] ) { // Unreachable
3115 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
3116 return;
3117 }
3118
3119 uint i;
3120 for( i = 0; i < n->req(); i++ )
3121 verify_compare( n->in(i), loop_verify, visited );
3122
3123 // Check the '_nodes' block/loop structure
3124 i = n->_idx;
3125 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
3126 if( _nodes[i] != loop_verify->_nodes[i] &&
3127 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
3128 tty->print("Mismatched control setting for: ");
3129 n->dump();
3130 if( fail++ > 10 ) return;
3131 Node *c = get_ctrl_no_update(n);
3132 tty->print("We have it as: ");
3133 if( c->in(0) ) c->dump();
3134 else tty->print_cr("N%d",c->_idx);
3135 tty->print("Verify thinks: ");
3136 if( loop_verify->has_ctrl(n) )
3137 loop_verify->get_ctrl_no_update(n)->dump();
3138 else
3139 loop_verify->get_loop_idx(n)->dump();
3140 tty->cr();
3141 }
3142 } else { // We have a loop
3143 IdealLoopTree *us = get_loop_idx(n);
3144 if( loop_verify->has_ctrl(n) ) {
3145 tty->print("Mismatched loop setting for: ");
3146 n->dump();
3147 if( fail++ > 10 ) return;
3148 tty->print("We have it as: ");
3149 us->dump();
3150 tty->print("Verify thinks: ");
3151 loop_verify->get_ctrl_no_update(n)->dump();
3152 tty->cr();
3153 } else if (!C->major_progress()) {
3154 // Loop selection can be messed up if we did a major progress
3155 // operation, like split-if. Do not verify in that case.
3156 IdealLoopTree *them = loop_verify->get_loop_idx(n);
3157 if( us->_head != them->_head || us->_tail != them->_tail ) {
3158 tty->print("Unequals loops for: ");
3159 n->dump();
3160 if( fail++ > 10 ) return;
3161 tty->print("We have it as: ");
3162 us->dump();
3163 tty->print("Verify thinks: ");
3164 them->dump();
3165 tty->cr();
3166 }
3167 }
3168 }
3169
3170 // Check for immediate dominators being equal
3171 if( i >= _idom_size ) {
3172 if( !n->is_CFG() ) return;
3173 tty->print("CFG Node with no idom: ");
3174 n->dump();
3175 return;
3176 }
3177 if( !n->is_CFG() ) return;
3178 if( n == C->root() ) return; // No IDOM here
3179
3180 assert(n->_idx == i, "sanity");
3181 Node *id = idom_no_update(n);
3182 if( id != loop_verify->idom_no_update(n) ) {
3183 tty->print("Unequals idoms for: ");
3184 n->dump();
3185 if( fail++ > 10 ) return;
3186 tty->print("We have it as: ");
3187 id->dump();
3188 tty->print("Verify thinks: ");
3189 loop_verify->idom_no_update(n)->dump();
3190 tty->cr();
3191 }
3192
3193 }
3194
3195 //------------------------------verify_tree------------------------------------
3196 // Verify that tree structures match. Because the CFG can change, siblings
3197 // within the loop tree can be reordered. We attempt to deal with that by
3198 // reordering the verify's loop tree if possible.
3199 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
3200 assert( _parent == parent, "Badly formed loop tree" );
3201
3202 // Siblings not in same order? Attempt to re-order.
3203 if( _head != loop->_head ) {
3204 // Find _next pointer to update
3205 IdealLoopTree **pp = &loop->_parent->_child;
3206 while( *pp != loop )
3207 pp = &((*pp)->_next);
3208 // Find proper sibling to be next
3209 IdealLoopTree **nn = &loop->_next;
3210 while( (*nn) && (*nn)->_head != _head )
3211 nn = &((*nn)->_next);
3212
3213 // Check for no match.
3214 if( !(*nn) ) {
3215 // Annoyingly, irreducible loops can pick different headers
3216 // after a major_progress operation, so the rest of the loop
3217 // tree cannot be matched.
3218 if (_irreducible && Compile::current()->major_progress()) return;
3219 assert( 0, "failed to match loop tree" );
3220 }
3221
3222 // Move (*nn) to (*pp)
3223 IdealLoopTree *hit = *nn;
3224 *nn = hit->_next;
3225 hit->_next = loop;
3226 *pp = loop;
3227 loop = hit;
3228 // Now try again to verify
3229 }
3230
3231 assert( _head == loop->_head , "mismatched loop head" );
3232 Node *tail = _tail; // Inline a non-updating version of
3233 while( !tail->in(0) ) // the 'tail()' call.
3234 tail = tail->in(1);
3235 assert( tail == loop->_tail, "mismatched loop tail" );
3236
3237 // Counted loops that are guarded should be able to find their guards
3238 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
3239 CountedLoopNode *cl = _head->as_CountedLoop();
3240 Node *init = cl->init_trip();
3241 Node *ctrl = cl->in(LoopNode::EntryControl);
3242 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
3243 Node *iff = ctrl->in(0);
3244 assert( iff->Opcode() == Op_If, "" );
3245 Node *bol = iff->in(1);
3246 assert( bol->Opcode() == Op_Bool, "" );
3247 Node *cmp = bol->in(1);
3248 assert( cmp->Opcode() == Op_CmpI, "" );
3249 Node *add = cmp->in(1);
3250 Node *opaq;
3251 if( add->Opcode() == Op_Opaque1 ) {
3252 opaq = add;
3253 } else {
3254 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
3255 assert( add == init, "" );
3256 opaq = cmp->in(2);
3257 }
3258 assert( opaq->Opcode() == Op_Opaque1, "" );
3259
3260 }
3261
3262 if (_child != NULL) _child->verify_tree(loop->_child, this);
3263 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
3264 // Innermost loops need to verify loop bodies,
3265 // but only if no 'major_progress'
3266 int fail = 0;
3267 if (!Compile::current()->major_progress() && _child == NULL) {
3268 for( uint i = 0; i < _body.size(); i++ ) {
3269 Node *n = _body.at(i);
3270 if (n->outcnt() == 0) continue; // Ignore dead
3271 uint j;
3272 for( j = 0; j < loop->_body.size(); j++ )
3273 if( loop->_body.at(j) == n )
3274 break;
3275 if( j == loop->_body.size() ) { // Not found in loop body
3276 // Last ditch effort to avoid assertion: Its possible that we
3277 // have some users (so outcnt not zero) but are still dead.
3278 // Try to find from root.
3279 if (Compile::current()->root()->find(n->_idx)) {
3280 fail++;
3281 tty->print("We have that verify does not: ");
3282 n->dump();
3283 }
3284 }
3285 }
3286 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
3287 Node *n = loop->_body.at(i2);
3288 if (n->outcnt() == 0) continue; // Ignore dead
3289 uint j;
3290 for( j = 0; j < _body.size(); j++ )
3291 if( _body.at(j) == n )
3292 break;
3293 if( j == _body.size() ) { // Not found in loop body
3294 // Last ditch effort to avoid assertion: Its possible that we
3295 // have some users (so outcnt not zero) but are still dead.
3296 // Try to find from root.
3297 if (Compile::current()->root()->find(n->_idx)) {
3298 fail++;
3299 tty->print("Verify has that we do not: ");
3300 n->dump();
3301 }
3302 }
3303 }
3304 assert( !fail, "loop body mismatch" );
3305 }
3306 }
3307
3308 #endif
3309
3310 //------------------------------set_idom---------------------------------------
3311 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
3312 uint idx = d->_idx;
3313 if (idx >= _idom_size) {
3314 uint newsize = _idom_size<<1;
3315 while( idx >= newsize ) {
3316 newsize <<= 1;
3317 }
3318 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
3319 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
3320 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
3321 _idom_size = newsize;
3322 }
3323 _idom[idx] = n;
3324 _dom_depth[idx] = dom_depth;
3325 }
3326
3327 //------------------------------recompute_dom_depth---------------------------------------
3328 // The dominator tree is constructed with only parent pointers.
3329 // This recomputes the depth in the tree by first tagging all
3330 // nodes as "no depth yet" marker. The next pass then runs up
3331 // the dom tree from each node marked "no depth yet", and computes
3332 // the depth on the way back down.
3333 void PhaseIdealLoop::recompute_dom_depth() {
3334 uint no_depth_marker = C->unique();
3335 uint i;
3336 // Initialize depth to "no depth yet" and realize all lazy updates
3337 for (i = 0; i < _idom_size; i++) {
3338 // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized).
3339 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
3340 _dom_depth[i] = no_depth_marker;
3341
3342 // heal _idom if it has a fwd mapping in _nodes
3343 if (_idom[i]->in(0) == NULL) {
3344 idom(i);
3345 }
3346 }
3347 }
3348 if (_dom_stk == NULL) {
3349 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
3350 if (init_size < 10) init_size = 10;
3351 _dom_stk = new GrowableArray<uint>(init_size);
3352 }
3353 // Compute new depth for each node.
3354 for (i = 0; i < _idom_size; i++) {
3355 uint j = i;
3356 // Run up the dom tree to find a node with a depth
3357 while (_dom_depth[j] == no_depth_marker) {
3358 _dom_stk->push(j);
3359 j = _idom[j]->_idx;
3360 }
3361 // Compute the depth on the way back down this tree branch
3362 uint dd = _dom_depth[j] + 1;
3363 while (_dom_stk->length() > 0) {
3364 uint j = _dom_stk->pop();
3365 _dom_depth[j] = dd;
3366 dd++;
3367 }
3368 }
3369 }
3370
3371 //------------------------------sort-------------------------------------------
3372 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
3373 // loop tree, not the root.
3374 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
3375 if( !innermost ) return loop; // New innermost loop
3376
3377 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
3378 assert( loop_preorder, "not yet post-walked loop" );
3379 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
3380 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
3381
3382 // Insert at start of list
3383 while( l ) { // Insertion sort based on pre-order
3384 if( l == loop ) return innermost; // Already on list!
3385 int l_preorder = get_preorder(l->_head); // Cache pre-order number
3386 assert( l_preorder, "not yet post-walked l" );
3387 // Check header pre-order number to figure proper nesting
3388 if( loop_preorder > l_preorder )
3389 break; // End of insertion
3390 // If headers tie (e.g., shared headers) check tail pre-order numbers.
3391 // Since I split shared headers, you'd think this could not happen.
3392 // BUT: I must first do the preorder numbering before I can discover I
3393 // have shared headers, so the split headers all get the same preorder
3394 // number as the RegionNode they split from.
3395 if( loop_preorder == l_preorder &&
3396 get_preorder(loop->_tail) < get_preorder(l->_tail) )
3397 break; // Also check for shared headers (same pre#)
3398 pp = &l->_parent; // Chain up list
3399 l = *pp;
3400 }
3401 // Link into list
3402 // Point predecessor to me
3403 *pp = loop;
3404 // Point me to successor
3405 IdealLoopTree *p = loop->_parent;
3406 loop->_parent = l; // Point me to successor
3407 if( p ) sort( p, innermost ); // Insert my parents into list as well
3408 return innermost;
3409 }
3410
3411 //------------------------------build_loop_tree--------------------------------
3412 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
3413 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
3414 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
3415 // tightest enclosing IdealLoopTree for post-walked.
3416 //
3417 // During my forward walk I do a short 1-layer lookahead to see if I can find
3418 // a loop backedge with that doesn't have any work on the backedge. This
3419 // helps me construct nested loops with shared headers better.
3420 //
3421 // Once I've done the forward recursion, I do the post-work. For each child
3422 // I check to see if there is a backedge. Backedges define a loop! I
3423 // insert an IdealLoopTree at the target of the backedge.
3424 //
3425 // During the post-work I also check to see if I have several children
3426 // belonging to different loops. If so, then this Node is a decision point
3427 // where control flow can choose to change loop nests. It is at this
3428 // decision point where I can figure out how loops are nested. At this
3429 // time I can properly order the different loop nests from my children.
3430 // Note that there may not be any backedges at the decision point!
3431 //
3432 // Since the decision point can be far removed from the backedges, I can't
3433 // order my loops at the time I discover them. Thus at the decision point
3434 // I need to inspect loop header pre-order numbers to properly nest my
3435 // loops. This means I need to sort my childrens' loops by pre-order.
3436 // The sort is of size number-of-control-children, which generally limits
3437 // it to size 2 (i.e., I just choose between my 2 target loops).
3438 void PhaseIdealLoop::build_loop_tree() {
3439 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3440 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
3441 Node *n = C->root();
3442 bltstack.push(n);
3443 int pre_order = 1;
3444 int stack_size;
3445
3446 while ( ( stack_size = bltstack.length() ) != 0 ) {
3447 n = bltstack.top(); // Leave node on stack
3448 if ( !is_visited(n) ) {
3449 // ---- Pre-pass Work ----
3450 // Pre-walked but not post-walked nodes need a pre_order number.
3451
3452 set_preorder_visited( n, pre_order ); // set as visited
3453
3454 // ---- Scan over children ----
3455 // Scan first over control projections that lead to loop headers.
3456 // This helps us find inner-to-outer loops with shared headers better.
3457
3458 // Scan children's children for loop headers.
3459 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
3460 Node* m = n->raw_out(i); // Child
3461 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
3462 // Scan over children's children to find loop
3463 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3464 Node* l = m->fast_out(j);
3465 if( is_visited(l) && // Been visited?
3466 !is_postvisited(l) && // But not post-visited
3467 get_preorder(l) < pre_order ) { // And smaller pre-order
3468 // Found! Scan the DFS down this path before doing other paths
3469 bltstack.push(m);
3470 break;
3471 }
3472 }
3473 }
3474 }
3475 pre_order++;
3476 }
3477 else if ( !is_postvisited(n) ) {
3478 // Note: build_loop_tree_impl() adds out edges on rare occasions,
3479 // such as com.sun.rsasign.am::a.
3480 // For non-recursive version, first, process current children.
3481 // On next iteration, check if additional children were added.
3482 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
3483 Node* u = n->raw_out(k);
3484 if ( u->is_CFG() && !is_visited(u) ) {
3485 bltstack.push(u);
3486 }
3487 }
3488 if ( bltstack.length() == stack_size ) {
3489 // There were no additional children, post visit node now
3490 (void)bltstack.pop(); // Remove node from stack
3491 pre_order = build_loop_tree_impl( n, pre_order );
3492 // Check for bailout
3493 if (C->failing()) {
3494 return;
3495 }
3496 // Check to grow _preorders[] array for the case when
3497 // build_loop_tree_impl() adds new nodes.
3498 check_grow_preorders();
3499 }
3500 }
3501 else {
3502 (void)bltstack.pop(); // Remove post-visited node from stack
3503 }
3504 }
3505 }
3506
3507 //------------------------------build_loop_tree_impl---------------------------
3508 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
3509 // ---- Post-pass Work ----
3510 // Pre-walked but not post-walked nodes need a pre_order number.
3511
3512 // Tightest enclosing loop for this Node
3513 IdealLoopTree *innermost = NULL;
3514
3515 // For all children, see if any edge is a backedge. If so, make a loop
3516 // for it. Then find the tightest enclosing loop for the self Node.
3517 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3518 Node* m = n->fast_out(i); // Child
3519 if( n == m ) continue; // Ignore control self-cycles
3520 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
3521
3522 IdealLoopTree *l; // Child's loop
3523 if( !is_postvisited(m) ) { // Child visited but not post-visited?
3524 // Found a backedge
3525 assert( get_preorder(m) < pre_order, "should be backedge" );
3526 // Check for the RootNode, which is already a LoopNode and is allowed
3527 // to have multiple "backedges".
3528 if( m == C->root()) { // Found the root?
3529 l = _ltree_root; // Root is the outermost LoopNode
3530 } else { // Else found a nested loop
3531 // Insert a LoopNode to mark this loop.
3532 l = new IdealLoopTree(this, m, n);
3533 } // End of Else found a nested loop
3534 if( !has_loop(m) ) // If 'm' does not already have a loop set
3535 set_loop(m, l); // Set loop header to loop now
3536
3537 } else { // Else not a nested loop
3538 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
3539 l = get_loop(m); // Get previously determined loop
3540 // If successor is header of a loop (nest), move up-loop till it
3541 // is a member of some outer enclosing loop. Since there are no
3542 // shared headers (I've split them already) I only need to go up
3543 // at most 1 level.
3544 while( l && l->_head == m ) // Successor heads loop?
3545 l = l->_parent; // Move up 1 for me
3546 // If this loop is not properly parented, then this loop
3547 // has no exit path out, i.e. its an infinite loop.
3548 if( !l ) {
3549 // Make loop "reachable" from root so the CFG is reachable. Basically
3550 // insert a bogus loop exit that is never taken. 'm', the loop head,
3551 // points to 'n', one (of possibly many) fall-in paths. There may be
3552 // many backedges as well.
3553
3554 // Here I set the loop to be the root loop. I could have, after
3555 // inserting a bogus loop exit, restarted the recursion and found my
3556 // new loop exit. This would make the infinite loop a first-class
3557 // loop and it would then get properly optimized. What's the use of
3558 // optimizing an infinite loop?
3559 l = _ltree_root; // Oops, found infinite loop
3560
3561 if (!_verify_only) {
3562 // Insert the NeverBranch between 'm' and it's control user.
3563 NeverBranchNode *iff = new NeverBranchNode( m );
3564 _igvn.register_new_node_with_optimizer(iff);
3565 set_loop(iff, l);
3566 Node *if_t = new CProjNode( iff, 0 );
3567 _igvn.register_new_node_with_optimizer(if_t);
3568 set_loop(if_t, l);
3569
3570 Node* cfg = NULL; // Find the One True Control User of m
3571 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3572 Node* x = m->fast_out(j);
3573 if (x->is_CFG() && x != m && x != iff)
3574 { cfg = x; break; }
3575 }
3576 assert(cfg != NULL, "must find the control user of m");
3577 uint k = 0; // Probably cfg->in(0)
3578 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
3579 cfg->set_req( k, if_t ); // Now point to NeverBranch
3580 _igvn._worklist.push(cfg);
3581
3582 // Now create the never-taken loop exit
3583 Node *if_f = new CProjNode( iff, 1 );
3584 _igvn.register_new_node_with_optimizer(if_f);
3585 set_loop(if_f, l);
3586 // Find frame ptr for Halt. Relies on the optimizer
3587 // V-N'ing. Easier and quicker than searching through
3588 // the program structure.
3589 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
3590 _igvn.register_new_node_with_optimizer(frame);
3591 // Halt & Catch Fire
3592 Node *halt = new HaltNode( if_f, frame );
3593 _igvn.register_new_node_with_optimizer(halt);
3594 set_loop(halt, l);
3595 C->root()->add_req(halt);
3596 }
3597 set_loop(C->root(), _ltree_root);
3598 }
3599 }
3600 // Weeny check for irreducible. This child was already visited (this
3601 // IS the post-work phase). Is this child's loop header post-visited
3602 // as well? If so, then I found another entry into the loop.
3603 if (!_verify_only) {
3604 while( is_postvisited(l->_head) ) {
3605 // found irreducible
3606 l->_irreducible = 1; // = true
3607 l = l->_parent;
3608 _has_irreducible_loops = true;
3609 // Check for bad CFG here to prevent crash, and bailout of compile
3610 if (l == NULL) {
3611 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3612 return pre_order;
3613 }
3614 }
3615 C->set_has_irreducible_loop(_has_irreducible_loops);
3616 }
3617
3618 // This Node might be a decision point for loops. It is only if
3619 // it's children belong to several different loops. The sort call
3620 // does a trivial amount of work if there is only 1 child or all
3621 // children belong to the same loop. If however, the children
3622 // belong to different loops, the sort call will properly set the
3623 // _parent pointers to show how the loops nest.
3624 //
3625 // In any case, it returns the tightest enclosing loop.
3626 innermost = sort( l, innermost );
3627 }
3628
3629 // Def-use info will have some dead stuff; dead stuff will have no
3630 // loop decided on.
3631
3632 // Am I a loop header? If so fix up my parent's child and next ptrs.
3633 if( innermost && innermost->_head == n ) {
3634 assert( get_loop(n) == innermost, "" );
3635 IdealLoopTree *p = innermost->_parent;
3636 IdealLoopTree *l = innermost;
3637 while( p && l->_head == n ) {
3638 l->_next = p->_child; // Put self on parents 'next child'
3639 p->_child = l; // Make self as first child of parent
3640 l = p; // Now walk up the parent chain
3641 p = l->_parent;
3642 }
3643 } else {
3644 // Note that it is possible for a LoopNode to reach here, if the
3645 // backedge has been made unreachable (hence the LoopNode no longer
3646 // denotes a Loop, and will eventually be removed).
3647
3648 // Record tightest enclosing loop for self. Mark as post-visited.
3649 set_loop(n, innermost);
3650 // Also record has_call flag early on
3651 if( innermost ) {
3652 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3653 // Do not count uncommon calls
3654 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3655 Node *iff = n->in(0)->in(0);
3656 // No any calls for vectorized loops.
3657 if( UseSuperWord || !iff->is_If() ||
3658 (n->in(0)->Opcode() == Op_IfFalse &&
3659 (1.0 - iff->as_If()->_prob) >= 0.01) ||
3660 (iff->as_If()->_prob >= 0.01) )
3661 innermost->_has_call = 1;
3662 }
3663 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3664 // Disable loop optimizations if the loop has a scalar replaceable
3665 // allocation. This disabling may cause a potential performance lost
3666 // if the allocation is not eliminated for some reason.
3667 innermost->_allow_optimizations = false;
3668 innermost->_has_call = 1; // = true
3669 } else if (n->Opcode() == Op_SafePoint) {
3670 // Record all safepoints in this loop.
3671 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3672 innermost->_safepts->push(n);
3673 }
3674 }
3675 }
3676
3677 // Flag as post-visited now
3678 set_postvisited(n);
3679 return pre_order;
3680 }
3681
3682
3683 //------------------------------build_loop_early-------------------------------
3684 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3685 // First pass computes the earliest controlling node possible. This is the
3686 // controlling input with the deepest dominating depth.
3687 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3688 while (worklist.size() != 0) {
3689 // Use local variables nstack_top_n & nstack_top_i to cache values
3690 // on nstack's top.
3691 Node *nstack_top_n = worklist.pop();
3692 uint nstack_top_i = 0;
3693 //while_nstack_nonempty:
3694 while (true) {
3695 // Get parent node and next input's index from stack's top.
3696 Node *n = nstack_top_n;
3697 uint i = nstack_top_i;
3698 uint cnt = n->req(); // Count of inputs
3699 if (i == 0) { // Pre-process the node.
3700 if( has_node(n) && // Have either loop or control already?
3701 !has_ctrl(n) ) { // Have loop picked out already?
3702 // During "merge_many_backedges" we fold up several nested loops
3703 // into a single loop. This makes the members of the original
3704 // loop bodies pointing to dead loops; they need to move up
3705 // to the new UNION'd larger loop. I set the _head field of these
3706 // dead loops to NULL and the _parent field points to the owning
3707 // loop. Shades of UNION-FIND algorithm.
3708 IdealLoopTree *ilt;
3709 while( !(ilt = get_loop(n))->_head ) {
3710 // Normally I would use a set_loop here. But in this one special
3711 // case, it is legal (and expected) to change what loop a Node
3712 // belongs to.
3713 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3714 }
3715 // Remove safepoints ONLY if I've already seen I don't need one.
3716 // (the old code here would yank a 2nd safepoint after seeing a
3717 // first one, even though the 1st did not dominate in the loop body
3718 // and thus could be avoided indefinitely)
3719 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3720 is_deleteable_safept(n)) {
3721 Node *in = n->in(TypeFunc::Control);
3722 lazy_replace(n,in); // Pull safepoint now
3723 if (ilt->_safepts != NULL) {
3724 ilt->_safepts->yank(n);
3725 }
3726 // Carry on with the recursion "as if" we are walking
3727 // only the control input
3728 if( !visited.test_set( in->_idx ) ) {
3729 worklist.push(in); // Visit this guy later, using worklist
3730 }
3731 // Get next node from nstack:
3732 // - skip n's inputs processing by setting i > cnt;
3733 // - we also will not call set_early_ctrl(n) since
3734 // has_node(n) == true (see the condition above).
3735 i = cnt + 1;
3736 }
3737 }
3738 } // if (i == 0)
3739
3740 // Visit all inputs
3741 bool done = true; // Assume all n's inputs will be processed
3742 while (i < cnt) {
3743 Node *in = n->in(i);
3744 ++i;
3745 if (in == NULL) continue;
3746 if (in->pinned() && !in->is_CFG())
3747 set_ctrl(in, in->in(0));
3748 int is_visited = visited.test_set( in->_idx );
3749 if (!has_node(in)) { // No controlling input yet?
3750 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3751 assert( !is_visited, "visit only once" );
3752 nstack.push(n, i); // Save parent node and next input's index.
3753 nstack_top_n = in; // Process current input now.
3754 nstack_top_i = 0;
3755 done = false; // Not all n's inputs processed.
3756 break; // continue while_nstack_nonempty;
3757 } else if (!is_visited) {
3758 // This guy has a location picked out for him, but has not yet
3759 // been visited. Happens to all CFG nodes, for instance.
3760 // Visit him using the worklist instead of recursion, to break
3761 // cycles. Since he has a location already we do not need to
3762 // find his location before proceeding with the current Node.
3763 worklist.push(in); // Visit this guy later, using worklist
3764 }
3765 }
3766 if (done) {
3767 // All of n's inputs have been processed, complete post-processing.
3768
3769 // Compute earliest point this Node can go.
3770 // CFG, Phi, pinned nodes already know their controlling input.
3771 if (!has_node(n)) {
3772 // Record earliest legal location
3773 set_early_ctrl( n );
3774 }
3775 if (nstack.is_empty()) {
3776 // Finished all nodes on stack.
3777 // Process next node on the worklist.
3778 break;
3779 }
3780 // Get saved parent node and next input's index.
3781 nstack_top_n = nstack.node();
3782 nstack_top_i = nstack.index();
3783 nstack.pop();
3784 }
3785 } // while (true)
3786 }
3787 }
3788
3789 //------------------------------dom_lca_internal--------------------------------
3790 // Pair-wise LCA
3791 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3792 if( !n1 ) return n2; // Handle NULL original LCA
3793 assert( n1->is_CFG(), "" );
3794 assert( n2->is_CFG(), "" );
3795 // find LCA of all uses
3796 uint d1 = dom_depth(n1);
3797 uint d2 = dom_depth(n2);
3798 while (n1 != n2) {
3799 if (d1 > d2) {
3800 n1 = idom(n1);
3801 d1 = dom_depth(n1);
3802 } else if (d1 < d2) {
3803 n2 = idom(n2);
3804 d2 = dom_depth(n2);
3805 } else {
3806 // Here d1 == d2. Due to edits of the dominator-tree, sections
3807 // of the tree might have the same depth. These sections have
3808 // to be searched more carefully.
3809
3810 // Scan up all the n1's with equal depth, looking for n2.
3811 Node *t1 = idom(n1);
3812 while (dom_depth(t1) == d1) {
3813 if (t1 == n2) return n2;
3814 t1 = idom(t1);
3815 }
3816 // Scan up all the n2's with equal depth, looking for n1.
3817 Node *t2 = idom(n2);
3818 while (dom_depth(t2) == d2) {
3819 if (t2 == n1) return n1;
3820 t2 = idom(t2);
3821 }
3822 // Move up to a new dominator-depth value as well as up the dom-tree.
3823 n1 = t1;
3824 n2 = t2;
3825 d1 = dom_depth(n1);
3826 d2 = dom_depth(n2);
3827 }
3828 }
3829 return n1;
3830 }
3831
3832 //------------------------------compute_idom-----------------------------------
3833 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3834 // IDOMs are correct.
3835 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3836 assert( region->is_Region(), "" );
3837 Node *LCA = NULL;
3838 for( uint i = 1; i < region->req(); i++ ) {
3839 if( region->in(i) != C->top() )
3840 LCA = dom_lca( LCA, region->in(i) );
3841 }
3842 return LCA;
3843 }
3844
3845 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3846 bool had_error = false;
3847 #ifdef ASSERT
3848 if (early != C->root()) {
3849 // Make sure that there's a dominance path from LCA to early
3850 Node* d = LCA;
3851 while (d != early) {
3852 if (d == C->root()) {
3853 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3854 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3855 had_error = true;
3856 break;
3857 }
3858 d = idom(d);
3859 }
3860 }
3861 #endif
3862 return had_error;
3863 }
3864
3865
3866 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3867 // Compute LCA over list of uses
3868 bool had_error = false;
3869 Node *LCA = NULL;
3870 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3871 Node* c = n->fast_out(i);
3872 if (_nodes[c->_idx] == NULL)
3873 continue; // Skip the occasional dead node
3874 if( c->is_Phi() ) { // For Phis, we must land above on the path
3875 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3876 if( c->in(j) == n ) { // Found matching input?
3877 Node *use = c->in(0)->in(j);
3878 if (_verify_only && use->is_top()) continue;
3879 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3880 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3881 }
3882 }
3883 } else {
3884 // For CFG data-users, use is in the block just prior
3885 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3886 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3887 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3888 }
3889 }
3890 assert(!had_error, "bad dominance");
3891 return LCA;
3892 }
3893
3894 // Check the shape of the graph at the loop entry. In some cases,
3895 // the shape of the graph does not match the shape outlined below.
3896 // That is caused by the Opaque1 node "protecting" the shape of
3897 // the graph being removed by, for example, the IGVN performed
3898 // in PhaseIdealLoop::build_and_optimize().
3899 //
3900 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3901 // loop unswitching, and IGVN, or a combination of them) can freely change
3902 // the graph's shape. As a result, the graph shape outlined below cannot
3903 // be guaranteed anymore.
3904 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) {
3905 if (!cl->is_main_loop() && !cl->is_post_loop()) {
3906 return false;
3907 }
3908 Node* ctrl = cl->skip_predicates();
3909
3910 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3911 return false;
3912 }
3913 Node* iffm = ctrl->in(0);
3914 if (iffm == NULL || !iffm->is_If()) {
3915 return false;
3916 }
3917 Node* bolzm = iffm->in(1);
3918 if (bolzm == NULL || !bolzm->is_Bool()) {
3919 return false;
3920 }
3921 Node* cmpzm = bolzm->in(1);
3922 if (cmpzm == NULL || !cmpzm->is_Cmp()) {
3923 return false;
3924 }
3925 // compares can get conditionally flipped
3926 bool found_opaque = false;
3927 for (uint i = 1; i < cmpzm->req(); i++) {
3928 Node* opnd = cmpzm->in(i);
3929 if (opnd && opnd->Opcode() == Op_Opaque1) {
3930 found_opaque = true;
3931 break;
3932 }
3933 }
3934 if (!found_opaque) {
3935 return false;
3936 }
3937 return true;
3938 }
3939
3940 //------------------------------get_late_ctrl----------------------------------
3941 // Compute latest legal control.
3942 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3943 assert(early != NULL, "early control should not be NULL");
3944
3945 Node* LCA = compute_lca_of_uses(n, early);
3946 #ifdef ASSERT
3947 if (LCA == C->root() && LCA != early) {
3948 // def doesn't dominate uses so print some useful debugging output
3949 compute_lca_of_uses(n, early, true);
3950 }
3951 #endif
3952
3953 // if this is a load, check for anti-dependent stores
3954 // We use a conservative algorithm to identify potential interfering
3955 // instructions and for rescheduling the load. The users of the memory
3956 // input of this load are examined. Any use which is not a load and is
3957 // dominated by early is considered a potentially interfering store.
3958 // This can produce false positives.
3959 if (n->is_Load() && LCA != early) {
3960 Node_List worklist;
3961
3962 Node *mem = n->in(MemNode::Memory);
3963 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3964 Node* s = mem->fast_out(i);
3965 worklist.push(s);
3966 }
3967 while(worklist.size() != 0 && LCA != early) {
3968 Node* s = worklist.pop();
3969 if (s->is_Load() || s->is_ShenandoahBarrier() || s->Opcode() == Op_SafePoint ||
3970 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)) {
3971 continue;
3972 } else if (s->is_MergeMem()) {
3973 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3974 Node* s1 = s->fast_out(i);
3975 worklist.push(s1);
3976 }
3977 } else {
3978 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3979 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3980 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3981 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3982 }
3983 }
3984 }
3985 }
3986
3987 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3988 return LCA;
3989 }
3990
3991 // true if CFG node d dominates CFG node n
3992 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3993 if (d == n)
3994 return true;
3995 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3996 uint dd = dom_depth(d);
3997 while (dom_depth(n) >= dd) {
3998 if (n == d)
3999 return true;
4000 n = idom(n);
4001 }
4002 return false;
4003 }
4004
4005 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
4006 // Pair-wise LCA with tags.
4007 // Tag each index with the node 'tag' currently being processed
4008 // before advancing up the dominator chain using idom().
4009 // Later calls that find a match to 'tag' know that this path has already
4010 // been considered in the current LCA (which is input 'n1' by convention).
4011 // Since get_late_ctrl() is only called once for each node, the tag array
4012 // does not need to be cleared between calls to get_late_ctrl().
4013 // Algorithm trades a larger constant factor for better asymptotic behavior
4014 //
4015 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
4016 uint d1 = dom_depth(n1);
4017 uint d2 = dom_depth(n2);
4018
4019 do {
4020 if (d1 > d2) {
4021 // current lca is deeper than n2
4022 _dom_lca_tags.map(n1->_idx, tag);
4023 n1 = idom(n1);
4024 d1 = dom_depth(n1);
4025 } else if (d1 < d2) {
4026 // n2 is deeper than current lca
4027 Node *memo = _dom_lca_tags[n2->_idx];
4028 if( memo == tag ) {
4029 return n1; // Return the current LCA
4030 }
4031 _dom_lca_tags.map(n2->_idx, tag);
4032 n2 = idom(n2);
4033 d2 = dom_depth(n2);
4034 } else {
4035 // Here d1 == d2. Due to edits of the dominator-tree, sections
4036 // of the tree might have the same depth. These sections have
4037 // to be searched more carefully.
4038
4039 // Scan up all the n1's with equal depth, looking for n2.
4040 _dom_lca_tags.map(n1->_idx, tag);
4041 Node *t1 = idom(n1);
4042 while (dom_depth(t1) == d1) {
4043 if (t1 == n2) return n2;
4044 _dom_lca_tags.map(t1->_idx, tag);
4045 t1 = idom(t1);
4046 }
4047 // Scan up all the n2's with equal depth, looking for n1.
4048 _dom_lca_tags.map(n2->_idx, tag);
4049 Node *t2 = idom(n2);
4050 while (dom_depth(t2) == d2) {
4051 if (t2 == n1) return n1;
4052 _dom_lca_tags.map(t2->_idx, tag);
4053 t2 = idom(t2);
4054 }
4055 // Move up to a new dominator-depth value as well as up the dom-tree.
4056 n1 = t1;
4057 n2 = t2;
4058 d1 = dom_depth(n1);
4059 d2 = dom_depth(n2);
4060 }
4061 } while (n1 != n2);
4062 return n1;
4063 }
4064
4065 //------------------------------init_dom_lca_tags------------------------------
4066 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4067 // Intended use does not involve any growth for the array, so it could
4068 // be of fixed size.
4069 void PhaseIdealLoop::init_dom_lca_tags() {
4070 uint limit = C->unique() + 1;
4071 _dom_lca_tags.map( limit, NULL );
4072 #ifdef ASSERT
4073 for( uint i = 0; i < limit; ++i ) {
4074 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4075 }
4076 #endif // ASSERT
4077 }
4078
4079 //------------------------------clear_dom_lca_tags------------------------------
4080 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4081 // Intended use does not involve any growth for the array, so it could
4082 // be of fixed size.
4083 void PhaseIdealLoop::clear_dom_lca_tags() {
4084 uint limit = C->unique() + 1;
4085 _dom_lca_tags.map( limit, NULL );
4086 _dom_lca_tags.clear();
4087 #ifdef ASSERT
4088 for( uint i = 0; i < limit; ++i ) {
4089 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4090 }
4091 #endif // ASSERT
4092 }
4093
4094 //------------------------------build_loop_late--------------------------------
4095 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4096 // Second pass finds latest legal placement, and ideal loop placement.
4097 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
4098 while (worklist.size() != 0) {
4099 Node *n = worklist.pop();
4100 // Only visit once
4101 if (visited.test_set(n->_idx)) continue;
4102 uint cnt = n->outcnt();
4103 uint i = 0;
4104 while (true) {
4105 assert( _nodes[n->_idx], "no dead nodes" );
4106 // Visit all children
4107 if (i < cnt) {
4108 Node* use = n->raw_out(i);
4109 ++i;
4110 // Check for dead uses. Aggressively prune such junk. It might be
4111 // dead in the global sense, but still have local uses so I cannot
4112 // easily call 'remove_dead_node'.
4113 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
4114 // Due to cycles, we might not hit the same fixed point in the verify
4115 // pass as we do in the regular pass. Instead, visit such phis as
4116 // simple uses of the loop head.
4117 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
4118 if( !visited.test(use->_idx) )
4119 worklist.push(use);
4120 } else if( !visited.test_set(use->_idx) ) {
4121 nstack.push(n, i); // Save parent and next use's index.
4122 n = use; // Process all children of current use.
4123 cnt = use->outcnt();
4124 i = 0;
4125 }
4126 } else {
4127 // Do not visit around the backedge of loops via data edges.
4128 // push dead code onto a worklist
4129 _deadlist.push(use);
4130 }
4131 } else {
4132 // All of n's children have been processed, complete post-processing.
4133 build_loop_late_post(n);
4134 if (nstack.is_empty()) {
4135 // Finished all nodes on stack.
4136 // Process next node on the worklist.
4137 break;
4138 }
4139 // Get saved parent node and next use's index. Visit the rest of uses.
4140 n = nstack.node();
4141 cnt = n->outcnt();
4142 i = nstack.index();
4143 nstack.pop();
4144 }
4145 }
4146 }
4147 }
4148
4149 // Verify that no data node is scheduled in the outer loop of a strip
4150 // mined loop.
4151 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
4152 #ifdef ASSERT
4153 if (get_loop(least)->_nest == 0) {
4154 return;
4155 }
4156 IdealLoopTree* loop = get_loop(least);
4157 Node* head = loop->_head;
4158 if (head->is_OuterStripMinedLoop() &&
4159 // Verification can't be applied to fully built strip mined loops
4160 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
4161 Node* sfpt = head->as_Loop()->outer_safepoint();
4162 ResourceMark rm;
4163 Unique_Node_List wq;
4164 wq.push(sfpt);
4165 for (uint i = 0; i < wq.size(); i++) {
4166 Node *m = wq.at(i);
4167 for (uint i = 1; i < m->req(); i++) {
4168 Node* nn = m->in(i);
4169 if (nn == n) {
4170 return;
4171 }
4172 if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
4173 wq.push(nn);
4174 }
4175 }
4176 }
4177 ShouldNotReachHere();
4178 }
4179 #endif
4180 }
4181
4182
4183 //------------------------------build_loop_late_post---------------------------
4184 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4185 // Second pass finds latest legal placement, and ideal loop placement.
4186 void PhaseIdealLoop::build_loop_late_post(Node *n) {
4187 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4188
4189 if (bs->build_loop_late_post(this, n)) {
4190 return;
4191 }
4192
4193 build_loop_late_post_work(n, true);
4194 }
4195
4196 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
4197
4198 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
4199 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
4200 }
4201
4202 #ifdef ASSERT
4203 if (_verify_only && !n->is_CFG()) {
4204 // Check def-use domination.
4205 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
4206 }
4207 #endif
4208
4209 // CFG and pinned nodes already handled
4210 if( n->in(0) ) {
4211 if( n->in(0)->is_top() ) return; // Dead?
4212
4213 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
4214 // _must_ be pinned (they have to observe their control edge of course).
4215 // Unlike Stores (which modify an unallocable resource, the memory
4216 // state), Mods/Loads can float around. So free them up.
4217 switch( n->Opcode() ) {
4218 case Op_DivI:
4219 case Op_DivF:
4220 case Op_DivD:
4221 case Op_ModI:
4222 case Op_ModF:
4223 case Op_ModD:
4224 case Op_LoadB: // Same with Loads; they can sink
4225 case Op_LoadUB: // during loop optimizations.
4226 case Op_LoadUS:
4227 case Op_LoadD:
4228 case Op_LoadF:
4229 case Op_LoadI:
4230 case Op_LoadKlass:
4231 case Op_LoadNKlass:
4232 case Op_LoadL:
4233 case Op_LoadS:
4234 case Op_LoadP:
4235 case Op_LoadBarrierSlowReg:
4236 case Op_LoadBarrierWeakSlowReg:
4237 case Op_LoadN:
4238 case Op_LoadRange:
4239 case Op_LoadD_unaligned:
4240 case Op_LoadL_unaligned:
4241 case Op_StrComp: // Does a bunch of load-like effects
4242 case Op_StrEquals:
4243 case Op_StrIndexOf:
4244 case Op_StrIndexOfChar:
4245 case Op_AryEq:
4246 case Op_HasNegatives:
4247 pinned = false;
4248 }
4249 if( pinned ) {
4250 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
4251 if( !chosen_loop->_child ) // Inner loop?
4252 chosen_loop->_body.push(n); // Collect inner loops
4253 return;
4254 }
4255 } else { // No slot zero
4256 if( n->is_CFG() ) { // CFG with no slot 0 is dead
4257 _nodes.map(n->_idx,0); // No block setting, it's globally dead
4258 return;
4259 }
4260 assert(!n->is_CFG() || n->outcnt() == 0, "");
4261 }
4262
4263 // Do I have a "safe range" I can select over?
4264 Node *early = get_ctrl(n);// Early location already computed
4265
4266 // Compute latest point this Node can go
4267 Node *LCA = get_late_ctrl( n, early );
4268 // LCA is NULL due to uses being dead
4269 if( LCA == NULL ) {
4270 #ifdef ASSERT
4271 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
4272 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
4273 }
4274 #endif
4275 _nodes.map(n->_idx, 0); // This node is useless
4276 _deadlist.push(n);
4277 return;
4278 }
4279 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
4280
4281 Node *legal = LCA; // Walk 'legal' up the IDOM chain
4282 Node *least = legal; // Best legal position so far
4283 while( early != legal ) { // While not at earliest legal
4284 #ifdef ASSERT
4285 if (legal->is_Start() && !early->is_Root()) {
4286 // Bad graph. Print idom path and fail.
4287 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
4288 assert(false, "Bad graph detected in build_loop_late");
4289 }
4290 #endif
4291 // Find least loop nesting depth
4292 legal = idom(legal); // Bump up the IDOM tree
4293 // Check for lower nesting depth
4294 if( get_loop(legal)->_nest < get_loop(least)->_nest )
4295 least = legal;
4296 }
4297 assert(early == legal || legal != C->root(), "bad dominance of inputs");
4298
4299 // Try not to place code on a loop entry projection
4300 // which can inhibit range check elimination.
4301 if (least != early) {
4302 Node* ctrl_out = least->unique_ctrl_out();
4303 if (ctrl_out && ctrl_out->is_Loop() &&
4304 least == ctrl_out->in(LoopNode::EntryControl)) {
4305 // Move the node above predicates as far up as possible so a
4306 // following pass of loop predication doesn't hoist a predicate
4307 // that depends on it above that node.
4308 Node* new_ctrl = least;
4309 for (;;) {
4310 if (!new_ctrl->is_Proj()) {
4311 break;
4312 }
4313 CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
4314 if (call == NULL) {
4315 break;
4316 }
4317 int req = call->uncommon_trap_request();
4318 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
4319 if (trap_reason != Deoptimization::Reason_loop_limit_check &&
4320 trap_reason != Deoptimization::Reason_predicate &&
4321 trap_reason != Deoptimization::Reason_profile_predicate) {
4322 break;
4323 }
4324 Node* c = new_ctrl->in(0)->in(0);
4325 if (is_dominator(c, early) && c != early) {
4326 break;
4327 }
4328 new_ctrl = c;
4329 }
4330 least = new_ctrl;
4331 }
4332 }
4333
4334 #ifdef ASSERT
4335 // If verifying, verify that 'verify_me' has a legal location
4336 // and choose it as our location.
4337 if( _verify_me ) {
4338 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
4339 Node *legal = LCA;
4340 while( early != legal ) { // While not at earliest legal
4341 if( legal == v_ctrl ) break; // Check for prior good location
4342 legal = idom(legal) ;// Bump up the IDOM tree
4343 }
4344 // Check for prior good location
4345 if( legal == v_ctrl ) least = legal; // Keep prior if found
4346 }
4347 #endif
4348
4349 // Assign discovered "here or above" point
4350 least = find_non_split_ctrl(least);
4351 verify_strip_mined_scheduling(n, least);
4352 set_ctrl(n, least);
4353
4354 // Collect inner loop bodies
4355 IdealLoopTree *chosen_loop = get_loop(least);
4356 if( !chosen_loop->_child ) // Inner loop?
4357 chosen_loop->_body.push(n);// Collect inner loops
4358 }
4359
4360 #ifdef ASSERT
4361 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
4362 tty->print_cr("%s", msg);
4363 tty->print("n: "); n->dump();
4364 tty->print("early(n): "); early->dump();
4365 if (n->in(0) != NULL && !n->in(0)->is_top() &&
4366 n->in(0) != early && !n->in(0)->is_Root()) {
4367 tty->print("n->in(0): "); n->in(0)->dump();
4368 }
4369 for (uint i = 1; i < n->req(); i++) {
4370 Node* in1 = n->in(i);
4371 if (in1 != NULL && in1 != n && !in1->is_top()) {
4372 tty->print("n->in(%d): ", i); in1->dump();
4373 Node* in1_early = get_ctrl(in1);
4374 tty->print("early(n->in(%d)): ", i); in1_early->dump();
4375 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
4376 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
4377 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
4378 }
4379 for (uint j = 1; j < in1->req(); j++) {
4380 Node* in2 = in1->in(j);
4381 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
4382 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
4383 Node* in2_early = get_ctrl(in2);
4384 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
4385 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
4386 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
4387 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
4388 }
4389 }
4390 }
4391 }
4392 }
4393 tty->cr();
4394 tty->print("LCA(n): "); LCA->dump();
4395 for (uint i = 0; i < n->outcnt(); i++) {
4396 Node* u1 = n->raw_out(i);
4397 if (u1 == n)
4398 continue;
4399 tty->print("n->out(%d): ", i); u1->dump();
4400 if (u1->is_CFG()) {
4401 for (uint j = 0; j < u1->outcnt(); j++) {
4402 Node* u2 = u1->raw_out(j);
4403 if (u2 != u1 && u2 != n && u2->is_CFG()) {
4404 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4405 }
4406 }
4407 } else {
4408 Node* u1_later = get_ctrl(u1);
4409 tty->print("later(n->out(%d)): ", i); u1_later->dump();
4410 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
4411 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
4412 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
4413 }
4414 for (uint j = 0; j < u1->outcnt(); j++) {
4415 Node* u2 = u1->raw_out(j);
4416 if (u2 == n || u2 == u1)
4417 continue;
4418 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4419 if (!u2->is_CFG()) {
4420 Node* u2_later = get_ctrl(u2);
4421 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
4422 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
4423 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
4424 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
4425 }
4426 }
4427 }
4428 }
4429 }
4430 tty->cr();
4431 int ct = 0;
4432 Node *dbg_legal = LCA;
4433 while(!dbg_legal->is_Start() && ct < 100) {
4434 tty->print("idom[%d] ",ct); dbg_legal->dump();
4435 ct++;
4436 dbg_legal = idom(dbg_legal);
4437 }
4438 tty->cr();
4439 }
4440 #endif
4441
4442 #ifndef PRODUCT
4443 //------------------------------dump-------------------------------------------
4444 void PhaseIdealLoop::dump( ) const {
4445 ResourceMark rm;
4446 Arena* arena = Thread::current()->resource_area();
4447 Node_Stack stack(arena, C->live_nodes() >> 2);
4448 Node_List rpo_list;
4449 VectorSet visited(arena);
4450 visited.set(C->top()->_idx);
4451 rpo( C->root(), stack, visited, rpo_list );
4452 // Dump root loop indexed by last element in PO order
4453 dump( _ltree_root, rpo_list.size(), rpo_list );
4454 }
4455
4456 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
4457 loop->dump_head();
4458
4459 // Now scan for CFG nodes in the same loop
4460 for( uint j=idx; j > 0; j-- ) {
4461 Node *n = rpo_list[j-1];
4462 if( !_nodes[n->_idx] ) // Skip dead nodes
4463 continue;
4464 if( get_loop(n) != loop ) { // Wrong loop nest
4465 if( get_loop(n)->_head == n && // Found nested loop?
4466 get_loop(n)->_parent == loop )
4467 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
4468 continue;
4469 }
4470
4471 // Dump controlling node
4472 for( uint x = 0; x < loop->_nest; x++ )
4473 tty->print(" ");
4474 tty->print("C");
4475 if( n == C->root() ) {
4476 n->dump();
4477 } else {
4478 Node* cached_idom = idom_no_update(n);
4479 Node *computed_idom = n->in(0);
4480 if( n->is_Region() ) {
4481 computed_idom = compute_idom(n);
4482 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
4483 // any MultiBranch ctrl node), so apply a similar transform to
4484 // the cached idom returned from idom_no_update.
4485 cached_idom = find_non_split_ctrl(cached_idom);
4486 }
4487 tty->print(" ID:%d",computed_idom->_idx);
4488 n->dump();
4489 if( cached_idom != computed_idom ) {
4490 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
4491 computed_idom->_idx, cached_idom->_idx);
4492 }
4493 }
4494 // Dump nodes it controls
4495 for( uint k = 0; k < _nodes.Size(); k++ ) {
4496 // (k < C->unique() && get_ctrl(find(k)) == n)
4497 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
4498 Node *m = C->root()->find(k);
4499 if( m && m->outcnt() > 0 ) {
4500 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
4501 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
4502 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
4503 }
4504 for( uint j = 0; j < loop->_nest; j++ )
4505 tty->print(" ");
4506 tty->print(" ");
4507 m->dump();
4508 }
4509 }
4510 }
4511 }
4512 }
4513 #endif
4514
4515 // Collect a R-P-O for the whole CFG.
4516 // Result list is in post-order (scan backwards for RPO)
4517 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
4518 stk.push(start, 0);
4519 visited.set(start->_idx);
4520
4521 while (stk.is_nonempty()) {
4522 Node* m = stk.node();
4523 uint idx = stk.index();
4524 if (idx < m->outcnt()) {
4525 stk.set_index(idx + 1);
4526 Node* n = m->raw_out(idx);
4527 if (n->is_CFG() && !visited.test_set(n->_idx)) {
4528 stk.push(n, 0);
4529 }
4530 } else {
4531 rpo_list.push(m);
4532 stk.pop();
4533 }
4534 }
4535 }
4536
4537
4538 //=============================================================================
4539 //------------------------------LoopTreeIterator-----------------------------------
4540
4541 // Advance to next loop tree using a preorder, left-to-right traversal.
4542 void LoopTreeIterator::next() {
4543 assert(!done(), "must not be done.");
4544 if (_curnt->_child != NULL) {
4545 _curnt = _curnt->_child;
4546 } else if (_curnt->_next != NULL) {
4547 _curnt = _curnt->_next;
4548 } else {
4549 while (_curnt != _root && _curnt->_next == NULL) {
4550 _curnt = _curnt->_parent;
4551 }
4552 if (_curnt == _root) {
4553 _curnt = NULL;
4554 assert(done(), "must be done.");
4555 } else {
4556 assert(_curnt->_next != NULL, "must be more to do");
4557 _curnt = _curnt->_next;
4558 }
4559 }
4560 }