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