Wed, 10 Aug 2016 14:59:21 +0200
8152172: PPC64: Support AES intrinsics
Summary: Add support for AES intrinsics on PPC64.
Reviewed-by: kvn, mdoerr, simonis, zmajo
Contributed-by: Hiroshi H Horii <horii@jp.ibm.com>
1 /*
2 * Copyright (c) 1998, 2014, 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 */
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 "opto/addnode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/rootnode.hpp"
38 #include "opto/superword.hpp"
40 //=============================================================================
41 //------------------------------is_loop_iv-------------------------------------
42 // Determine if a node is Counted loop induction variable.
43 // The method is declared in node.hpp.
44 const Node* Node::is_loop_iv() const {
45 if (this->is_Phi() && !this->as_Phi()->is_copy() &&
46 this->as_Phi()->region()->is_CountedLoop() &&
47 this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
48 return this;
49 } else {
50 return NULL;
51 }
52 }
54 //=============================================================================
55 //------------------------------dump_spec--------------------------------------
56 // Dump special per-node info
57 #ifndef PRODUCT
58 void LoopNode::dump_spec(outputStream *st) const {
59 if (is_inner_loop()) st->print( "inner " );
60 if (is_partial_peel_loop()) st->print( "partial_peel " );
61 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
62 }
63 #endif
65 //------------------------------is_valid_counted_loop-------------------------
66 bool LoopNode::is_valid_counted_loop() const {
67 if (is_CountedLoop()) {
68 CountedLoopNode* l = as_CountedLoop();
69 CountedLoopEndNode* le = l->loopexit();
70 if (le != NULL &&
71 le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
72 Node* phi = l->phi();
73 Node* exit = le->proj_out(0 /* false */);
74 if (exit != NULL && exit->Opcode() == Op_IfFalse &&
75 phi != NULL && phi->is_Phi() &&
76 phi->in(LoopNode::LoopBackControl) == l->incr() &&
77 le->loopnode() == l && le->stride_is_con()) {
78 return true;
79 }
80 }
81 }
82 return false;
83 }
85 //------------------------------get_early_ctrl---------------------------------
86 // Compute earliest legal control
87 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
88 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
89 uint i;
90 Node *early;
91 if (n->in(0) && !n->is_expensive()) {
92 early = n->in(0);
93 if (!early->is_CFG()) // Might be a non-CFG multi-def
94 early = get_ctrl(early); // So treat input as a straight data input
95 i = 1;
96 } else {
97 early = get_ctrl(n->in(1));
98 i = 2;
99 }
100 uint e_d = dom_depth(early);
101 assert( early, "" );
102 for (; i < n->req(); i++) {
103 Node *cin = get_ctrl(n->in(i));
104 assert( cin, "" );
105 // Keep deepest dominator depth
106 uint c_d = dom_depth(cin);
107 if (c_d > e_d) { // Deeper guy?
108 early = cin; // Keep deepest found so far
109 e_d = c_d;
110 } else if (c_d == e_d && // Same depth?
111 early != cin) { // If not equal, must use slower algorithm
112 // If same depth but not equal, one _must_ dominate the other
113 // and we want the deeper (i.e., dominated) guy.
114 Node *n1 = early;
115 Node *n2 = cin;
116 while (1) {
117 n1 = idom(n1); // Walk up until break cycle
118 n2 = idom(n2);
119 if (n1 == cin || // Walked early up to cin
120 dom_depth(n2) < c_d)
121 break; // early is deeper; keep him
122 if (n2 == early || // Walked cin up to early
123 dom_depth(n1) < c_d) {
124 early = cin; // cin is deeper; keep him
125 break;
126 }
127 }
128 e_d = dom_depth(early); // Reset depth register cache
129 }
130 }
132 // Return earliest legal location
133 assert(early == find_non_split_ctrl(early), "unexpected early control");
135 if (n->is_expensive()) {
136 assert(n->in(0), "should have control input");
137 early = get_early_ctrl_for_expensive(n, early);
138 }
140 return early;
141 }
143 //------------------------------get_early_ctrl_for_expensive---------------------------------
144 // Move node up the dominator tree as high as legal while still beneficial
145 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
146 assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
147 assert(OptimizeExpensiveOps, "optimization off?");
149 Node* ctl = n->in(0);
150 assert(ctl->is_CFG(), "expensive input 0 must be cfg");
151 uint min_dom_depth = dom_depth(earliest);
152 #ifdef ASSERT
153 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
154 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
155 assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
156 }
157 #endif
158 if (dom_depth(ctl) < min_dom_depth) {
159 return earliest;
160 }
162 while (1) {
163 Node *next = ctl;
164 // Moving the node out of a loop on the projection of a If
165 // confuses loop predication. So once we hit a Loop in a If branch
166 // that doesn't branch to an UNC, we stop. The code that process
167 // expensive nodes will notice the loop and skip over it to try to
168 // move the node further up.
169 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
170 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
171 break;
172 }
173 next = idom(ctl->in(1)->in(0));
174 } else if (ctl->is_Proj()) {
175 // We only move it up along a projection if the projection is
176 // the single control projection for its parent: same code path,
177 // if it's a If with UNC or fallthrough of a call.
178 Node* parent_ctl = ctl->in(0);
179 if (parent_ctl == NULL) {
180 break;
181 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
182 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
183 } else if (parent_ctl->is_If()) {
184 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
185 break;
186 }
187 assert(idom(ctl) == parent_ctl, "strange");
188 next = idom(parent_ctl);
189 } else if (ctl->is_CatchProj()) {
190 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
191 break;
192 }
193 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
194 next = parent_ctl->in(0)->in(0)->in(0);
195 } else {
196 // Check if parent control has a single projection (this
197 // control is the only possible successor of the parent
198 // control). If so, we can try to move the node above the
199 // parent control.
200 int nb_ctl_proj = 0;
201 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
202 Node *p = parent_ctl->fast_out(i);
203 if (p->is_Proj() && p->is_CFG()) {
204 nb_ctl_proj++;
205 if (nb_ctl_proj > 1) {
206 break;
207 }
208 }
209 }
211 if (nb_ctl_proj > 1) {
212 break;
213 }
214 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node");
215 assert(idom(ctl) == parent_ctl, "strange");
216 next = idom(parent_ctl);
217 }
218 } else {
219 next = idom(ctl);
220 }
221 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
222 break;
223 }
224 ctl = next;
225 }
227 if (ctl != n->in(0)) {
228 _igvn.hash_delete(n);
229 n->set_req(0, ctl);
230 _igvn.hash_insert(n);
231 }
233 return ctl;
234 }
237 //------------------------------set_early_ctrl---------------------------------
238 // Set earliest legal control
239 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
240 Node *early = get_early_ctrl(n);
242 // Record earliest legal location
243 set_ctrl(n, early);
244 }
246 //------------------------------set_subtree_ctrl-------------------------------
247 // set missing _ctrl entries on new nodes
248 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
249 // Already set? Get out.
250 if( _nodes[n->_idx] ) return;
251 // Recursively set _nodes array to indicate where the Node goes
252 uint i;
253 for( i = 0; i < n->req(); ++i ) {
254 Node *m = n->in(i);
255 if( m && m != C->root() )
256 set_subtree_ctrl( m );
257 }
259 // Fixup self
260 set_early_ctrl( n );
261 }
263 //------------------------------is_counted_loop--------------------------------
264 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
265 PhaseGVN *gvn = &_igvn;
267 // Counted loop head must be a good RegionNode with only 3 not NULL
268 // control input edges: Self, Entry, LoopBack.
269 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
270 return false;
271 }
272 Node *init_control = x->in(LoopNode::EntryControl);
273 Node *back_control = x->in(LoopNode::LoopBackControl);
274 if (init_control == NULL || back_control == NULL) // Partially dead
275 return false;
276 // Must also check for TOP when looking for a dead loop
277 if (init_control->is_top() || back_control->is_top())
278 return false;
280 // Allow funny placement of Safepoint
281 if (back_control->Opcode() == Op_SafePoint)
282 back_control = back_control->in(TypeFunc::Control);
284 // Controlling test for loop
285 Node *iftrue = back_control;
286 uint iftrue_op = iftrue->Opcode();
287 if (iftrue_op != Op_IfTrue &&
288 iftrue_op != Op_IfFalse)
289 // I have a weird back-control. Probably the loop-exit test is in
290 // the middle of the loop and I am looking at some trailing control-flow
291 // merge point. To fix this I would have to partially peel the loop.
292 return false; // Obscure back-control
294 // Get boolean guarding loop-back test
295 Node *iff = iftrue->in(0);
296 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
297 return false;
298 BoolNode *test = iff->in(1)->as_Bool();
299 BoolTest::mask bt = test->_test._test;
300 float cl_prob = iff->as_If()->_prob;
301 if (iftrue_op == Op_IfFalse) {
302 bt = BoolTest(bt).negate();
303 cl_prob = 1.0 - cl_prob;
304 }
305 // Get backedge compare
306 Node *cmp = test->in(1);
307 int cmp_op = cmp->Opcode();
308 if (cmp_op != Op_CmpI)
309 return false; // Avoid pointer & float compares
311 // Find the trip-counter increment & limit. Limit must be loop invariant.
312 Node *incr = cmp->in(1);
313 Node *limit = cmp->in(2);
315 // ---------
316 // need 'loop()' test to tell if limit is loop invariant
317 // ---------
319 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
320 Node *tmp = incr; // Then reverse order into the CmpI
321 incr = limit;
322 limit = tmp;
323 bt = BoolTest(bt).commute(); // And commute the exit test
324 }
325 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
326 return false;
327 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
328 return false;
330 Node* phi_incr = NULL;
331 // Trip-counter increment must be commutative & associative.
332 if (incr->is_Phi()) {
333 if (incr->as_Phi()->region() != x || incr->req() != 3)
334 return false; // Not simple trip counter expression
335 phi_incr = incr;
336 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
337 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
338 return false;
339 }
341 Node* trunc1 = NULL;
342 Node* trunc2 = NULL;
343 const TypeInt* iv_trunc_t = NULL;
344 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
345 return false; // Funny increment opcode
346 }
347 assert(incr->Opcode() == Op_AddI, "wrong increment code");
349 // Get merge point
350 Node *xphi = incr->in(1);
351 Node *stride = incr->in(2);
352 if (!stride->is_Con()) { // Oops, swap these
353 if (!xphi->is_Con()) // Is the other guy a constant?
354 return false; // Nope, unknown stride, bail out
355 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
356 xphi = stride;
357 stride = tmp;
358 }
359 // Stride must be constant
360 int stride_con = stride->get_int();
361 if (stride_con == 0)
362 return false; // missed some peephole opt
364 if (!xphi->is_Phi())
365 return false; // Too much math on the trip counter
366 if (phi_incr != NULL && phi_incr != xphi)
367 return false;
368 PhiNode *phi = xphi->as_Phi();
370 // Phi must be of loop header; backedge must wrap to increment
371 if (phi->region() != x)
372 return false;
373 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
374 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
375 return false;
376 }
377 Node *init_trip = phi->in(LoopNode::EntryControl);
379 // If iv trunc type is smaller than int, check for possible wrap.
380 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
381 assert(trunc1 != NULL, "must have found some truncation");
383 // Get a better type for the phi (filtered thru if's)
384 const TypeInt* phi_ft = filtered_type(phi);
386 // Can iv take on a value that will wrap?
387 //
388 // Ensure iv's limit is not within "stride" of the wrap value.
389 //
390 // Example for "short" type
391 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
392 // If the stride is +10, then the last value of the induction
393 // variable before the increment (phi_ft->_hi) must be
394 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
395 // ensure no truncation occurs after the increment.
397 if (stride_con > 0) {
398 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
399 iv_trunc_t->_lo > phi_ft->_lo) {
400 return false; // truncation may occur
401 }
402 } else if (stride_con < 0) {
403 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
404 iv_trunc_t->_hi < phi_ft->_hi) {
405 return false; // truncation may occur
406 }
407 }
408 // No possibility of wrap so truncation can be discarded
409 // Promote iv type to Int
410 } else {
411 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
412 }
414 // If the condition is inverted and we will be rolling
415 // through MININT to MAXINT, then bail out.
416 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
417 // Odd stride
418 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
419 // Count down loop rolls through MAXINT
420 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
421 // Count up loop rolls through MININT
422 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
423 return false; // Bail out
424 }
426 const TypeInt* init_t = gvn->type(init_trip)->is_int();
427 const TypeInt* limit_t = gvn->type(limit)->is_int();
429 if (stride_con > 0) {
430 jlong init_p = (jlong)init_t->_lo + stride_con;
431 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
432 return false; // cyclic loop or this loop trips only once
433 } else {
434 jlong init_p = (jlong)init_t->_hi + stride_con;
435 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
436 return false; // cyclic loop or this loop trips only once
437 }
439 if (phi_incr != NULL) {
440 // check if there is a possiblity of IV overflowing after the first increment
441 if (stride_con > 0) {
442 if (init_t->_hi > max_jint - stride_con) {
443 return false;
444 }
445 } else {
446 if (init_t->_lo < min_jint - stride_con) {
447 return false;
448 }
449 }
450 }
452 // =================================================
453 // ---- SUCCESS! Found A Trip-Counted Loop! -----
454 //
455 assert(x->Opcode() == Op_Loop, "regular loops only");
456 C->print_method(PHASE_BEFORE_CLOOPS, 3);
458 Node *hook = new (C) Node(6);
460 if (LoopLimitCheck) {
462 // ===================================================
463 // Generate loop limit check to avoid integer overflow
464 // in cases like next (cyclic loops):
465 //
466 // for (i=0; i <= max_jint; i++) {}
467 // for (i=0; i < max_jint; i+=2) {}
468 //
469 //
470 // Limit check predicate depends on the loop test:
471 //
472 // for(;i != limit; i++) --> limit <= (max_jint)
473 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
474 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
475 //
477 // Check if limit is excluded to do more precise int overflow check.
478 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
479 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
481 // If compare points directly to the phi we need to adjust
482 // the compare so that it points to the incr. Limit have
483 // to be adjusted to keep trip count the same and the
484 // adjusted limit should be checked for int overflow.
485 if (phi_incr != NULL) {
486 stride_m += stride_con;
487 }
489 if (limit->is_Con()) {
490 int limit_con = limit->get_int();
491 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
492 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
493 // Bailout: it could be integer overflow.
494 return false;
495 }
496 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
497 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
498 // Limit's type may satisfy the condition, for example,
499 // when it is an array length.
500 } else {
501 // Generate loop's limit check.
502 // Loop limit check predicate should be near the loop.
503 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
504 if (!limit_check_proj) {
505 // The limit check predicate is not generated if this method trapped here before.
506 #ifdef ASSERT
507 if (TraceLoopLimitCheck) {
508 tty->print("missing loop limit check:");
509 loop->dump_head();
510 x->dump(1);
511 }
512 #endif
513 return false;
514 }
516 IfNode* check_iff = limit_check_proj->in(0)->as_If();
517 Node* cmp_limit;
518 Node* bol;
520 if (stride_con > 0) {
521 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(max_jint - stride_m));
522 bol = new (C) BoolNode(cmp_limit, BoolTest::le);
523 } else {
524 cmp_limit = new (C) CmpINode(limit, _igvn.intcon(min_jint - stride_m));
525 bol = new (C) BoolNode(cmp_limit, BoolTest::ge);
526 }
527 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
528 bol = _igvn.register_new_node_with_optimizer(bol);
529 set_subtree_ctrl(bol);
531 // Replace condition in original predicate but preserve Opaque node
532 // so that previous predicates could be found.
533 assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
534 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
535 Node* opq = check_iff->in(1)->in(1);
536 _igvn.hash_delete(opq);
537 opq->set_req(1, bol);
538 // Update ctrl.
539 set_ctrl(opq, check_iff->in(0));
540 set_ctrl(check_iff->in(1), check_iff->in(0));
542 #ifndef PRODUCT
543 // report that the loop predication has been actually performed
544 // for this loop
545 if (TraceLoopLimitCheck) {
546 tty->print_cr("Counted Loop Limit Check generated:");
547 debug_only( bol->dump(2); )
548 }
549 #endif
550 }
552 if (phi_incr != NULL) {
553 // If compare points directly to the phi we need to adjust
554 // the compare so that it points to the incr. Limit have
555 // to be adjusted to keep trip count the same and we
556 // should avoid int overflow.
557 //
558 // i = init; do {} while(i++ < limit);
559 // is converted to
560 // i = init; do {} while(++i < limit+1);
561 //
562 limit = gvn->transform(new (C) AddINode(limit, stride));
563 }
565 // Now we need to canonicalize loop condition.
566 if (bt == BoolTest::ne) {
567 assert(stride_con == 1 || stride_con == -1, "simple increment only");
568 // 'ne' can be replaced with 'lt' only when init < limit.
569 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
570 bt = BoolTest::lt;
571 // 'ne' can be replaced with 'gt' only when init > limit.
572 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
573 bt = BoolTest::gt;
574 }
576 if (incl_limit) {
577 // The limit check guaranties that 'limit <= (max_jint - stride)' so
578 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
579 //
580 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
581 limit = gvn->transform(new (C) AddINode(limit, one));
582 if (bt == BoolTest::le)
583 bt = BoolTest::lt;
584 else if (bt == BoolTest::ge)
585 bt = BoolTest::gt;
586 else
587 ShouldNotReachHere();
588 }
589 set_subtree_ctrl( limit );
591 } else { // LoopLimitCheck
593 // If compare points to incr, we are ok. Otherwise the compare
594 // can directly point to the phi; in this case adjust the compare so that
595 // it points to the incr by adjusting the limit.
596 if (cmp->in(1) == phi || cmp->in(2) == phi)
597 limit = gvn->transform(new (C) AddINode(limit,stride));
599 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
600 // Final value for iterator should be: trip_count * stride + init_trip.
601 Node *one_p = gvn->intcon( 1);
602 Node *one_m = gvn->intcon(-1);
604 Node *trip_count = NULL;
605 switch( bt ) {
606 case BoolTest::eq:
607 ShouldNotReachHere();
608 case BoolTest::ne: // Ahh, the case we desire
609 if (stride_con == 1)
610 trip_count = gvn->transform(new (C) SubINode(limit,init_trip));
611 else if (stride_con == -1)
612 trip_count = gvn->transform(new (C) SubINode(init_trip,limit));
613 else
614 ShouldNotReachHere();
615 set_subtree_ctrl(trip_count);
616 //_loop.map(trip_count->_idx,loop(limit));
617 break;
618 case BoolTest::le: // Maybe convert to '<' case
619 limit = gvn->transform(new (C) AddINode(limit,one_p));
620 set_subtree_ctrl( limit );
621 hook->init_req(4, limit);
623 bt = BoolTest::lt;
624 // Make the new limit be in the same loop nest as the old limit
625 //_loop.map(limit->_idx,limit_loop);
626 // Fall into next case
627 case BoolTest::lt: { // Maybe convert to '!=' case
628 if (stride_con < 0) // Count down loop rolls through MAXINT
629 ShouldNotReachHere();
630 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
631 set_subtree_ctrl( range );
632 hook->init_req(0, range);
634 Node *bias = gvn->transform(new (C) AddINode(range,stride));
635 set_subtree_ctrl( bias );
636 hook->init_req(1, bias);
638 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_m));
639 set_subtree_ctrl( bias1 );
640 hook->init_req(2, bias1);
642 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
643 set_subtree_ctrl( trip_count );
644 hook->init_req(3, trip_count);
645 break;
646 }
648 case BoolTest::ge: // Maybe convert to '>' case
649 limit = gvn->transform(new (C) AddINode(limit,one_m));
650 set_subtree_ctrl( limit );
651 hook->init_req(4 ,limit);
653 bt = BoolTest::gt;
654 // Make the new limit be in the same loop nest as the old limit
655 //_loop.map(limit->_idx,limit_loop);
656 // Fall into next case
657 case BoolTest::gt: { // Maybe convert to '!=' case
658 if (stride_con > 0) // count up loop rolls through MININT
659 ShouldNotReachHere();
660 Node *range = gvn->transform(new (C) SubINode(limit,init_trip));
661 set_subtree_ctrl( range );
662 hook->init_req(0, range);
664 Node *bias = gvn->transform(new (C) AddINode(range,stride));
665 set_subtree_ctrl( bias );
666 hook->init_req(1, bias);
668 Node *bias1 = gvn->transform(new (C) AddINode(bias,one_p));
669 set_subtree_ctrl( bias1 );
670 hook->init_req(2, bias1);
672 trip_count = gvn->transform(new (C) DivINode(0,bias1,stride));
673 set_subtree_ctrl( trip_count );
674 hook->init_req(3, trip_count);
675 break;
676 }
677 } // switch( bt )
679 Node *span = gvn->transform(new (C) MulINode(trip_count,stride));
680 set_subtree_ctrl( span );
681 hook->init_req(5, span);
683 limit = gvn->transform(new (C) AddINode(span,init_trip));
684 set_subtree_ctrl( limit );
686 } // LoopLimitCheck
688 if (!UseCountedLoopSafepoints) {
689 // Check for SafePoint on backedge and remove
690 Node *sfpt = x->in(LoopNode::LoopBackControl);
691 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
692 lazy_replace( sfpt, iftrue );
693 if (loop->_safepts != NULL) {
694 loop->_safepts->yank(sfpt);
695 }
696 loop->_tail = iftrue;
697 }
698 }
700 // Build a canonical trip test.
701 // Clone code, as old values may be in use.
702 incr = incr->clone();
703 incr->set_req(1,phi);
704 incr->set_req(2,stride);
705 incr = _igvn.register_new_node_with_optimizer(incr);
706 set_early_ctrl( incr );
707 _igvn.hash_delete(phi);
708 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
710 // If phi type is more restrictive than Int, raise to
711 // Int to prevent (almost) infinite recursion in igvn
712 // which can only handle integer types for constants or minint..maxint.
713 if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
714 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
715 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
716 nphi = _igvn.register_new_node_with_optimizer(nphi);
717 set_ctrl(nphi, get_ctrl(phi));
718 _igvn.replace_node(phi, nphi);
719 phi = nphi->as_Phi();
720 }
721 cmp = cmp->clone();
722 cmp->set_req(1,incr);
723 cmp->set_req(2,limit);
724 cmp = _igvn.register_new_node_with_optimizer(cmp);
725 set_ctrl(cmp, iff->in(0));
727 test = test->clone()->as_Bool();
728 (*(BoolTest*)&test->_test)._test = bt;
729 test->set_req(1,cmp);
730 _igvn.register_new_node_with_optimizer(test);
731 set_ctrl(test, iff->in(0));
733 // Replace the old IfNode with a new LoopEndNode
734 Node *lex = _igvn.register_new_node_with_optimizer(new (C) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
735 IfNode *le = lex->as_If();
736 uint dd = dom_depth(iff);
737 set_idom(le, le->in(0), dd); // Update dominance for loop exit
738 set_loop(le, loop);
740 // Get the loop-exit control
741 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
743 // Need to swap loop-exit and loop-back control?
744 if (iftrue_op == Op_IfFalse) {
745 Node *ift2=_igvn.register_new_node_with_optimizer(new (C) IfTrueNode (le));
746 Node *iff2=_igvn.register_new_node_with_optimizer(new (C) IfFalseNode(le));
748 loop->_tail = back_control = ift2;
749 set_loop(ift2, loop);
750 set_loop(iff2, get_loop(iffalse));
752 // Lazy update of 'get_ctrl' mechanism.
753 lazy_replace(iffalse, iff2);
754 lazy_replace(iftrue, ift2);
756 // Swap names
757 iffalse = iff2;
758 iftrue = ift2;
759 } else {
760 _igvn.hash_delete(iffalse);
761 _igvn.hash_delete(iftrue);
762 iffalse->set_req_X( 0, le, &_igvn );
763 iftrue ->set_req_X( 0, le, &_igvn );
764 }
766 set_idom(iftrue, le, dd+1);
767 set_idom(iffalse, le, dd+1);
768 assert(iff->outcnt() == 0, "should be dead now");
769 lazy_replace( iff, le ); // fix 'get_ctrl'
771 // Now setup a new CountedLoopNode to replace the existing LoopNode
772 CountedLoopNode *l = new (C) CountedLoopNode(init_control, back_control);
773 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
774 // The following assert is approximately true, and defines the intention
775 // of can_be_counted_loop. It fails, however, because phase->type
776 // is not yet initialized for this loop and its parts.
777 //assert(l->can_be_counted_loop(this), "sanity");
778 _igvn.register_new_node_with_optimizer(l);
779 set_loop(l, loop);
780 loop->_head = l;
781 // Fix all data nodes placed at the old loop head.
782 // Uses the lazy-update mechanism of 'get_ctrl'.
783 lazy_replace( x, l );
784 set_idom(l, init_control, dom_depth(x));
786 if (!UseCountedLoopSafepoints) {
787 // Check for immediately preceding SafePoint and remove
788 Node *sfpt2 = le->in(0);
789 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {
790 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
791 if (loop->_safepts != NULL) {
792 loop->_safepts->yank(sfpt2);
793 }
794 }
795 }
797 // Free up intermediate goo
798 _igvn.remove_dead_node(hook);
800 #ifdef ASSERT
801 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
802 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
803 #endif
804 #ifndef PRODUCT
805 if (TraceLoopOpts) {
806 tty->print("Counted ");
807 loop->dump_head();
808 }
809 #endif
811 C->print_method(PHASE_AFTER_CLOOPS, 3);
813 return true;
814 }
816 //----------------------exact_limit-------------------------------------------
817 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
818 assert(loop->_head->is_CountedLoop(), "");
819 CountedLoopNode *cl = loop->_head->as_CountedLoop();
820 assert(cl->is_valid_counted_loop(), "");
822 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
823 cl->limit()->Opcode() == Op_LoopLimit) {
824 // Old code has exact limit (it could be incorrect in case of int overflow).
825 // Loop limit is exact with stride == 1. And loop may already have exact limit.
826 return cl->limit();
827 }
828 Node *limit = NULL;
829 #ifdef ASSERT
830 BoolTest::mask bt = cl->loopexit()->test_trip();
831 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
832 #endif
833 if (cl->has_exact_trip_count()) {
834 // Simple case: loop has constant boundaries.
835 // Use jlongs to avoid integer overflow.
836 int stride_con = cl->stride_con();
837 jlong init_con = cl->init_trip()->get_int();
838 jlong limit_con = cl->limit()->get_int();
839 julong trip_cnt = cl->trip_count();
840 jlong final_con = init_con + trip_cnt*stride_con;
841 int final_int = (int)final_con;
842 // The final value should be in integer range since the loop
843 // is counted and the limit was checked for overflow.
844 assert(final_con == (jlong)final_int, "final value should be integer");
845 limit = _igvn.intcon(final_int);
846 } else {
847 // Create new LoopLimit node to get exact limit (final iv value).
848 limit = new (C) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
849 register_new_node(limit, cl->in(LoopNode::EntryControl));
850 }
851 assert(limit != NULL, "sanity");
852 return limit;
853 }
855 //------------------------------Ideal------------------------------------------
856 // Return a node which is more "ideal" than the current node.
857 // Attempt to convert into a counted-loop.
858 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
859 if (!can_be_counted_loop(phase)) {
860 phase->C->set_major_progress();
861 }
862 return RegionNode::Ideal(phase, can_reshape);
863 }
866 //=============================================================================
867 //------------------------------Ideal------------------------------------------
868 // Return a node which is more "ideal" than the current node.
869 // Attempt to convert into a counted-loop.
870 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
871 return RegionNode::Ideal(phase, can_reshape);
872 }
874 //------------------------------dump_spec--------------------------------------
875 // Dump special per-node info
876 #ifndef PRODUCT
877 void CountedLoopNode::dump_spec(outputStream *st) const {
878 LoopNode::dump_spec(st);
879 if (stride_is_con()) {
880 st->print("stride: %d ",stride_con());
881 }
882 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
883 if (is_main_loop()) st->print("main of N%d", _idx);
884 if (is_post_loop()) st->print("post of N%d", _main_idx);
885 }
886 #endif
888 //=============================================================================
889 int CountedLoopEndNode::stride_con() const {
890 return stride()->bottom_type()->is_int()->get_con();
891 }
893 //=============================================================================
894 //------------------------------Value-----------------------------------------
895 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
896 const Type* init_t = phase->type(in(Init));
897 const Type* limit_t = phase->type(in(Limit));
898 const Type* stride_t = phase->type(in(Stride));
899 // Either input is TOP ==> the result is TOP
900 if (init_t == Type::TOP) return Type::TOP;
901 if (limit_t == Type::TOP) return Type::TOP;
902 if (stride_t == Type::TOP) return Type::TOP;
904 int stride_con = stride_t->is_int()->get_con();
905 if (stride_con == 1)
906 return NULL; // Identity
908 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
909 // Use jlongs to avoid integer overflow.
910 jlong init_con = init_t->is_int()->get_con();
911 jlong limit_con = limit_t->is_int()->get_con();
912 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
913 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
914 jlong final_con = init_con + stride_con*trip_count;
915 int final_int = (int)final_con;
916 // The final value should be in integer range since the loop
917 // is counted and the limit was checked for overflow.
918 assert(final_con == (jlong)final_int, "final value should be integer");
919 return TypeInt::make(final_int);
920 }
922 return bottom_type(); // TypeInt::INT
923 }
925 //------------------------------Ideal------------------------------------------
926 // Return a node which is more "ideal" than the current node.
927 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
928 if (phase->type(in(Init)) == Type::TOP ||
929 phase->type(in(Limit)) == Type::TOP ||
930 phase->type(in(Stride)) == Type::TOP)
931 return NULL; // Dead
933 int stride_con = phase->type(in(Stride))->is_int()->get_con();
934 if (stride_con == 1)
935 return NULL; // Identity
937 if (in(Init)->is_Con() && in(Limit)->is_Con())
938 return NULL; // Value
940 // Delay following optimizations until all loop optimizations
941 // done to keep Ideal graph simple.
942 if (!can_reshape || phase->C->major_progress())
943 return NULL;
945 const TypeInt* init_t = phase->type(in(Init) )->is_int();
946 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
947 int stride_p;
948 jlong lim, ini;
949 julong max;
950 if (stride_con > 0) {
951 stride_p = stride_con;
952 lim = limit_t->_hi;
953 ini = init_t->_lo;
954 max = (julong)max_jint;
955 } else {
956 stride_p = -stride_con;
957 lim = init_t->_hi;
958 ini = limit_t->_lo;
959 max = (julong)min_jint;
960 }
961 julong range = lim - ini + stride_p;
962 if (range <= max) {
963 // Convert to integer expression if it is not overflow.
964 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
965 Node *range = phase->transform(new (phase->C) SubINode(in(Limit), in(Init)));
966 Node *bias = phase->transform(new (phase->C) AddINode(range, stride_m));
967 Node *trip = phase->transform(new (phase->C) DivINode(0, bias, in(Stride)));
968 Node *span = phase->transform(new (phase->C) MulINode(trip, in(Stride)));
969 return new (phase->C) AddINode(span, in(Init)); // exact limit
970 }
972 if (is_power_of_2(stride_p) || // divisor is 2^n
973 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
974 // Convert to long expression to avoid integer overflow
975 // and let igvn optimizer convert this division.
976 //
977 Node* init = phase->transform( new (phase->C) ConvI2LNode(in(Init)));
978 Node* limit = phase->transform( new (phase->C) ConvI2LNode(in(Limit)));
979 Node* stride = phase->longcon(stride_con);
980 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
982 Node *range = phase->transform(new (phase->C) SubLNode(limit, init));
983 Node *bias = phase->transform(new (phase->C) AddLNode(range, stride_m));
984 Node *span;
985 if (stride_con > 0 && is_power_of_2(stride_p)) {
986 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
987 // and avoid generating rounding for division. Zero trip guard should
988 // guarantee that init < limit but sometimes the guard is missing and
989 // we can get situation when init > limit. Note, for the empty loop
990 // optimization zero trip guard is generated explicitly which leaves
991 // only RCE predicate where exact limit is used and the predicate
992 // will simply fail forcing recompilation.
993 Node* neg_stride = phase->longcon(-stride_con);
994 span = phase->transform(new (phase->C) AndLNode(bias, neg_stride));
995 } else {
996 Node *trip = phase->transform(new (phase->C) DivLNode(0, bias, stride));
997 span = phase->transform(new (phase->C) MulLNode(trip, stride));
998 }
999 // Convert back to int
1000 Node *span_int = phase->transform(new (phase->C) ConvL2INode(span));
1001 return new (phase->C) AddINode(span_int, in(Init)); // exact limit
1002 }
1004 return NULL; // No progress
1005 }
1007 //------------------------------Identity---------------------------------------
1008 // If stride == 1 return limit node.
1009 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
1010 int stride_con = phase->type(in(Stride))->is_int()->get_con();
1011 if (stride_con == 1 || stride_con == -1)
1012 return in(Limit);
1013 return this;
1014 }
1016 //=============================================================================
1017 //----------------------match_incr_with_optional_truncation--------------------
1018 // Match increment with optional truncation:
1019 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1020 // Return NULL for failure. Success returns the increment node.
1021 Node* CountedLoopNode::match_incr_with_optional_truncation(
1022 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1023 // Quick cutouts:
1024 if (expr == NULL || expr->req() != 3) return NULL;
1026 Node *t1 = NULL;
1027 Node *t2 = NULL;
1028 const TypeInt* trunc_t = TypeInt::INT;
1029 Node* n1 = expr;
1030 int n1op = n1->Opcode();
1032 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1033 if (n1op == Op_AndI &&
1034 n1->in(2)->is_Con() &&
1035 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1036 // %%% This check should match any mask of 2**K-1.
1037 t1 = n1;
1038 n1 = t1->in(1);
1039 n1op = n1->Opcode();
1040 trunc_t = TypeInt::CHAR;
1041 } else if (n1op == Op_RShiftI &&
1042 n1->in(1) != NULL &&
1043 n1->in(1)->Opcode() == Op_LShiftI &&
1044 n1->in(2) == n1->in(1)->in(2) &&
1045 n1->in(2)->is_Con()) {
1046 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1047 // %%% This check should match any shift in [1..31].
1048 if (shift == 16 || shift == 8) {
1049 t1 = n1;
1050 t2 = t1->in(1);
1051 n1 = t2->in(1);
1052 n1op = n1->Opcode();
1053 if (shift == 16) {
1054 trunc_t = TypeInt::SHORT;
1055 } else if (shift == 8) {
1056 trunc_t = TypeInt::BYTE;
1057 }
1058 }
1059 }
1061 // If (maybe after stripping) it is an AddI, we won:
1062 if (n1op == Op_AddI) {
1063 *trunc1 = t1;
1064 *trunc2 = t2;
1065 *trunc_type = trunc_t;
1066 return n1;
1067 }
1069 // failed
1070 return NULL;
1071 }
1074 //------------------------------filtered_type--------------------------------
1075 // Return a type based on condition control flow
1076 // A successful return will be a type that is restricted due
1077 // to a series of dominating if-tests, such as:
1078 // if (i < 10) {
1079 // if (i > 0) {
1080 // here: "i" type is [1..10)
1081 // }
1082 // }
1083 // or a control flow merge
1084 // if (i < 10) {
1085 // do {
1086 // phi( , ) -- at top of loop type is [min_int..10)
1087 // i = ?
1088 // } while ( i < 10)
1089 //
1090 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1091 assert(n && n->bottom_type()->is_int(), "must be int");
1092 const TypeInt* filtered_t = NULL;
1093 if (!n->is_Phi()) {
1094 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1095 filtered_t = filtered_type_from_dominators(n, n_ctrl);
1097 } else {
1098 Node* phi = n->as_Phi();
1099 Node* region = phi->in(0);
1100 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1101 if (region && region != C->top()) {
1102 for (uint i = 1; i < phi->req(); i++) {
1103 Node* val = phi->in(i);
1104 Node* use_c = region->in(i);
1105 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1106 if (val_t != NULL) {
1107 if (filtered_t == NULL) {
1108 filtered_t = val_t;
1109 } else {
1110 filtered_t = filtered_t->meet(val_t)->is_int();
1111 }
1112 }
1113 }
1114 }
1115 }
1116 const TypeInt* n_t = _igvn.type(n)->is_int();
1117 if (filtered_t != NULL) {
1118 n_t = n_t->join(filtered_t)->is_int();
1119 }
1120 return n_t;
1121 }
1124 //------------------------------filtered_type_from_dominators--------------------------------
1125 // Return a possibly more restrictive type for val based on condition control flow of dominators
1126 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1127 if (val->is_Con()) {
1128 return val->bottom_type()->is_int();
1129 }
1130 uint if_limit = 10; // Max number of dominating if's visited
1131 const TypeInt* rtn_t = NULL;
1133 if (use_ctrl && use_ctrl != C->top()) {
1134 Node* val_ctrl = get_ctrl(val);
1135 uint val_dom_depth = dom_depth(val_ctrl);
1136 Node* pred = use_ctrl;
1137 uint if_cnt = 0;
1138 while (if_cnt < if_limit) {
1139 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1140 if_cnt++;
1141 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1142 if (if_t != NULL) {
1143 if (rtn_t == NULL) {
1144 rtn_t = if_t;
1145 } else {
1146 rtn_t = rtn_t->join(if_t)->is_int();
1147 }
1148 }
1149 }
1150 pred = idom(pred);
1151 if (pred == NULL || pred == C->top()) {
1152 break;
1153 }
1154 // Stop if going beyond definition block of val
1155 if (dom_depth(pred) < val_dom_depth) {
1156 break;
1157 }
1158 }
1159 }
1160 return rtn_t;
1161 }
1164 //------------------------------dump_spec--------------------------------------
1165 // Dump special per-node info
1166 #ifndef PRODUCT
1167 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1168 if( in(TestValue)->is_Bool() ) {
1169 BoolTest bt( test_trip()); // Added this for g++.
1171 st->print("[");
1172 bt.dump_on(st);
1173 st->print("]");
1174 }
1175 st->print(" ");
1176 IfNode::dump_spec(st);
1177 }
1178 #endif
1180 //=============================================================================
1181 //------------------------------is_member--------------------------------------
1182 // Is 'l' a member of 'this'?
1183 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
1184 while( l->_nest > _nest ) l = l->_parent;
1185 return l == this;
1186 }
1188 //------------------------------set_nest---------------------------------------
1189 // Set loop tree nesting depth. Accumulate _has_call bits.
1190 int IdealLoopTree::set_nest( uint depth ) {
1191 _nest = depth;
1192 int bits = _has_call;
1193 if( _child ) bits |= _child->set_nest(depth+1);
1194 if( bits ) _has_call = 1;
1195 if( _next ) bits |= _next ->set_nest(depth );
1196 return bits;
1197 }
1199 //------------------------------split_fall_in----------------------------------
1200 // Split out multiple fall-in edges from the loop header. Move them to a
1201 // private RegionNode before the loop. This becomes the loop landing pad.
1202 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1203 PhaseIterGVN &igvn = phase->_igvn;
1204 uint i;
1206 // Make a new RegionNode to be the landing pad.
1207 Node *landing_pad = new (phase->C) RegionNode( fall_in_cnt+1 );
1208 phase->set_loop(landing_pad,_parent);
1209 // Gather all the fall-in control paths into the landing pad
1210 uint icnt = fall_in_cnt;
1211 uint oreq = _head->req();
1212 for( i = oreq-1; i>0; i-- )
1213 if( !phase->is_member( this, _head->in(i) ) )
1214 landing_pad->set_req(icnt--,_head->in(i));
1216 // Peel off PhiNode edges as well
1217 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1218 Node *oj = _head->fast_out(j);
1219 if( oj->is_Phi() ) {
1220 PhiNode* old_phi = oj->as_Phi();
1221 assert( old_phi->region() == _head, "" );
1222 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1223 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1224 uint icnt = fall_in_cnt;
1225 for( i = oreq-1; i>0; i-- ) {
1226 if( !phase->is_member( this, _head->in(i) ) ) {
1227 p->init_req(icnt--, old_phi->in(i));
1228 // Go ahead and clean out old edges from old phi
1229 old_phi->del_req(i);
1230 }
1231 }
1232 // Search for CSE's here, because ZKM.jar does a lot of
1233 // loop hackery and we need to be a little incremental
1234 // with the CSE to avoid O(N^2) node blow-up.
1235 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1236 if( p2 ) { // Found CSE
1237 p->destruct(); // Recover useless new node
1238 p = p2; // Use old node
1239 } else {
1240 igvn.register_new_node_with_optimizer(p, old_phi);
1241 }
1242 // Make old Phi refer to new Phi.
1243 old_phi->add_req(p);
1244 // Check for the special case of making the old phi useless and
1245 // disappear it. In JavaGrande I have a case where this useless
1246 // Phi is the loop limit and prevents recognizing a CountedLoop
1247 // which in turn prevents removing an empty loop.
1248 Node *id_old_phi = old_phi->Identity( &igvn );
1249 if( id_old_phi != old_phi ) { // Found a simple identity?
1250 // Note that I cannot call 'replace_node' here, because
1251 // that will yank the edge from old_phi to the Region and
1252 // I'm mid-iteration over the Region's uses.
1253 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1254 Node* use = old_phi->last_out(i);
1255 igvn.rehash_node_delayed(use);
1256 uint uses_found = 0;
1257 for (uint j = 0; j < use->len(); j++) {
1258 if (use->in(j) == old_phi) {
1259 if (j < use->req()) use->set_req (j, id_old_phi);
1260 else use->set_prec(j, id_old_phi);
1261 uses_found++;
1262 }
1263 }
1264 i -= uses_found; // we deleted 1 or more copies of this edge
1265 }
1266 }
1267 igvn._worklist.push(old_phi);
1268 }
1269 }
1270 // Finally clean out the fall-in edges from the RegionNode
1271 for( i = oreq-1; i>0; i-- ) {
1272 if( !phase->is_member( this, _head->in(i) ) ) {
1273 _head->del_req(i);
1274 }
1275 }
1276 // Transform landing pad
1277 igvn.register_new_node_with_optimizer(landing_pad, _head);
1278 // Insert landing pad into the header
1279 _head->add_req(landing_pad);
1280 }
1282 //------------------------------split_outer_loop-------------------------------
1283 // Split out the outermost loop from this shared header.
1284 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1285 PhaseIterGVN &igvn = phase->_igvn;
1287 // Find index of outermost loop; it should also be my tail.
1288 uint outer_idx = 1;
1289 while( _head->in(outer_idx) != _tail ) outer_idx++;
1291 // Make a LoopNode for the outermost loop.
1292 Node *ctl = _head->in(LoopNode::EntryControl);
1293 Node *outer = new (phase->C) LoopNode( ctl, _head->in(outer_idx) );
1294 outer = igvn.register_new_node_with_optimizer(outer, _head);
1295 phase->set_created_loop_node();
1297 // Outermost loop falls into '_head' loop
1298 _head->set_req(LoopNode::EntryControl, outer);
1299 _head->del_req(outer_idx);
1300 // Split all the Phis up between '_head' loop and 'outer' loop.
1301 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1302 Node *out = _head->fast_out(j);
1303 if( out->is_Phi() ) {
1304 PhiNode *old_phi = out->as_Phi();
1305 assert( old_phi->region() == _head, "" );
1306 Node *phi = PhiNode::make_blank(outer, old_phi);
1307 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1308 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1309 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1310 // Make old Phi point to new Phi on the fall-in path
1311 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1312 old_phi->del_req(outer_idx);
1313 }
1314 }
1316 // Use the new loop head instead of the old shared one
1317 _head = outer;
1318 phase->set_loop(_head, this);
1319 }
1321 //------------------------------fix_parent-------------------------------------
1322 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1323 loop->_parent = parent;
1324 if( loop->_child ) fix_parent( loop->_child, loop );
1325 if( loop->_next ) fix_parent( loop->_next , parent );
1326 }
1328 //------------------------------estimate_path_freq-----------------------------
1329 static float estimate_path_freq( Node *n ) {
1330 // Try to extract some path frequency info
1331 IfNode *iff;
1332 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1333 uint nop = n->Opcode();
1334 if( nop == Op_SafePoint ) { // Skip any safepoint
1335 n = n->in(0);
1336 continue;
1337 }
1338 if( nop == Op_CatchProj ) { // Get count from a prior call
1339 // Assume call does not always throw exceptions: means the call-site
1340 // count is also the frequency of the fall-through path.
1341 assert( n->is_CatchProj(), "" );
1342 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1343 return 0.0f; // Assume call exception path is rare
1344 Node *call = n->in(0)->in(0)->in(0);
1345 assert( call->is_Call(), "expect a call here" );
1346 const JVMState *jvms = ((CallNode*)call)->jvms();
1347 ciMethodData* methodData = jvms->method()->method_data();
1348 if (!methodData->is_mature()) return 0.0f; // No call-site data
1349 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1350 if ((data == NULL) || !data->is_CounterData()) {
1351 // no call profile available, try call's control input
1352 n = n->in(0);
1353 continue;
1354 }
1355 return data->as_CounterData()->count()/FreqCountInvocations;
1356 }
1357 // See if there's a gating IF test
1358 Node *n_c = n->in(0);
1359 if( !n_c->is_If() ) break; // No estimate available
1360 iff = n_c->as_If();
1361 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1362 // Compute how much count comes on this path
1363 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1364 // Have no count info. Skip dull uncommon-trap like branches.
1365 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1366 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1367 break;
1368 // Skip through never-taken branch; look for a real loop exit.
1369 n = iff->in(0);
1370 }
1371 return 0.0f; // No estimate available
1372 }
1374 //------------------------------merge_many_backedges---------------------------
1375 // Merge all the backedges from the shared header into a private Region.
1376 // Feed that region as the one backedge to this loop.
1377 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1378 uint i;
1380 // Scan for the top 2 hottest backedges
1381 float hotcnt = 0.0f;
1382 float warmcnt = 0.0f;
1383 uint hot_idx = 0;
1384 // Loop starts at 2 because slot 1 is the fall-in path
1385 for( i = 2; i < _head->req(); i++ ) {
1386 float cnt = estimate_path_freq(_head->in(i));
1387 if( cnt > hotcnt ) { // Grab hottest path
1388 warmcnt = hotcnt;
1389 hotcnt = cnt;
1390 hot_idx = i;
1391 } else if( cnt > warmcnt ) { // And 2nd hottest path
1392 warmcnt = cnt;
1393 }
1394 }
1396 // See if the hottest backedge is worthy of being an inner loop
1397 // by being much hotter than the next hottest backedge.
1398 if( hotcnt <= 0.0001 ||
1399 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1401 // Peel out the backedges into a private merge point; peel
1402 // them all except optionally hot_idx.
1403 PhaseIterGVN &igvn = phase->_igvn;
1405 Node *hot_tail = NULL;
1406 // Make a Region for the merge point
1407 Node *r = new (phase->C) RegionNode(1);
1408 for( i = 2; i < _head->req(); i++ ) {
1409 if( i != hot_idx )
1410 r->add_req( _head->in(i) );
1411 else hot_tail = _head->in(i);
1412 }
1413 igvn.register_new_node_with_optimizer(r, _head);
1414 // Plug region into end of loop _head, followed by hot_tail
1415 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1416 _head->set_req(2, r);
1417 if( hot_idx ) _head->add_req(hot_tail);
1419 // Split all the Phis up between '_head' loop and the Region 'r'
1420 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1421 Node *out = _head->fast_out(j);
1422 if( out->is_Phi() ) {
1423 PhiNode* n = out->as_Phi();
1424 igvn.hash_delete(n); // Delete from hash before hacking edges
1425 Node *hot_phi = NULL;
1426 Node *phi = new (phase->C) PhiNode(r, n->type(), n->adr_type());
1427 // Check all inputs for the ones to peel out
1428 uint j = 1;
1429 for( uint i = 2; i < n->req(); i++ ) {
1430 if( i != hot_idx )
1431 phi->set_req( j++, n->in(i) );
1432 else hot_phi = n->in(i);
1433 }
1434 // Register the phi but do not transform until whole place transforms
1435 igvn.register_new_node_with_optimizer(phi, n);
1436 // Add the merge phi to the old Phi
1437 while( n->req() > 3 ) n->del_req( n->req()-1 );
1438 n->set_req(2, phi);
1439 if( hot_idx ) n->add_req(hot_phi);
1440 }
1441 }
1444 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1445 // of self loop tree. Turn self into a loop headed by _head and with
1446 // tail being the new merge point.
1447 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1448 phase->set_loop(_tail,ilt); // Adjust tail
1449 _tail = r; // Self's tail is new merge point
1450 phase->set_loop(r,this);
1451 ilt->_child = _child; // New guy has my children
1452 _child = ilt; // Self has new guy as only child
1453 ilt->_parent = this; // new guy has self for parent
1454 ilt->_nest = _nest; // Same nesting depth (for now)
1456 // Starting with 'ilt', look for child loop trees using the same shared
1457 // header. Flatten these out; they will no longer be loops in the end.
1458 IdealLoopTree **pilt = &_child;
1459 while( ilt ) {
1460 if( ilt->_head == _head ) {
1461 uint i;
1462 for( i = 2; i < _head->req(); i++ )
1463 if( _head->in(i) == ilt->_tail )
1464 break; // Still a loop
1465 if( i == _head->req() ) { // No longer a loop
1466 // Flatten ilt. Hang ilt's "_next" list from the end of
1467 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1468 IdealLoopTree **cp = &ilt->_child;
1469 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1470 *cp = ilt->_next; // Hang next list at end of child list
1471 *pilt = ilt->_child; // Move child up to replace ilt
1472 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1473 ilt = ilt->_child; // Repeat using new ilt
1474 continue; // do not advance over ilt->_child
1475 }
1476 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1477 phase->set_loop(_head,ilt);
1478 }
1479 pilt = &ilt->_child; // Advance to next
1480 ilt = *pilt;
1481 }
1483 if( _child ) fix_parent( _child, this );
1484 }
1486 //------------------------------beautify_loops---------------------------------
1487 // Split shared headers and insert loop landing pads.
1488 // Insert a LoopNode to replace the RegionNode.
1489 // Return TRUE if loop tree is structurally changed.
1490 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1491 bool result = false;
1492 // Cache parts in locals for easy
1493 PhaseIterGVN &igvn = phase->_igvn;
1495 igvn.hash_delete(_head); // Yank from hash before hacking edges
1497 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1498 int fall_in_cnt = 0;
1499 for( uint i = 1; i < _head->req(); i++ )
1500 if( !phase->is_member( this, _head->in(i) ) )
1501 fall_in_cnt++;
1502 assert( fall_in_cnt, "at least 1 fall-in path" );
1503 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1504 split_fall_in( phase, fall_in_cnt );
1506 // Swap inputs to the _head and all Phis to move the fall-in edge to
1507 // the left.
1508 fall_in_cnt = 1;
1509 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1510 fall_in_cnt++;
1511 if( fall_in_cnt > 1 ) {
1512 // Since I am just swapping inputs I do not need to update def-use info
1513 Node *tmp = _head->in(1);
1514 _head->set_req( 1, _head->in(fall_in_cnt) );
1515 _head->set_req( fall_in_cnt, tmp );
1516 // Swap also all Phis
1517 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1518 Node* phi = _head->fast_out(i);
1519 if( phi->is_Phi() ) {
1520 igvn.hash_delete(phi); // Yank from hash before hacking edges
1521 tmp = phi->in(1);
1522 phi->set_req( 1, phi->in(fall_in_cnt) );
1523 phi->set_req( fall_in_cnt, tmp );
1524 }
1525 }
1526 }
1527 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1528 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1530 // If I am a shared header (multiple backedges), peel off the many
1531 // backedges into a private merge point and use the merge point as
1532 // the one true backedge.
1533 if( _head->req() > 3 ) {
1534 // Merge the many backedges into a single backedge but leave
1535 // the hottest backedge as separate edge for the following peel.
1536 merge_many_backedges( phase );
1537 result = true;
1538 }
1540 // If I have one hot backedge, peel off myself loop.
1541 // I better be the outermost loop.
1542 if (_head->req() > 3 && !_irreducible) {
1543 split_outer_loop( phase );
1544 result = true;
1546 } else if (!_head->is_Loop() && !_irreducible) {
1547 // Make a new LoopNode to replace the old loop head
1548 Node *l = new (phase->C) LoopNode( _head->in(1), _head->in(2) );
1549 l = igvn.register_new_node_with_optimizer(l, _head);
1550 phase->set_created_loop_node();
1551 // Go ahead and replace _head
1552 phase->_igvn.replace_node( _head, l );
1553 _head = l;
1554 phase->set_loop(_head, this);
1555 }
1557 // Now recursively beautify nested loops
1558 if( _child ) result |= _child->beautify_loops( phase );
1559 if( _next ) result |= _next ->beautify_loops( phase );
1560 return result;
1561 }
1563 //------------------------------allpaths_check_safepts----------------------------
1564 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1565 // encountered. Helper for check_safepts.
1566 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1567 assert(stack.size() == 0, "empty stack");
1568 stack.push(_tail);
1569 visited.Clear();
1570 visited.set(_tail->_idx);
1571 while (stack.size() > 0) {
1572 Node* n = stack.pop();
1573 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1574 // Terminate this path
1575 } else if (n->Opcode() == Op_SafePoint) {
1576 if (_phase->get_loop(n) != this) {
1577 if (_required_safept == NULL) _required_safept = new Node_List();
1578 _required_safept->push(n); // save the one closest to the tail
1579 }
1580 // Terminate this path
1581 } else {
1582 uint start = n->is_Region() ? 1 : 0;
1583 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1584 for (uint i = start; i < end; i++) {
1585 Node* in = n->in(i);
1586 assert(in->is_CFG(), "must be");
1587 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1588 stack.push(in);
1589 }
1590 }
1591 }
1592 }
1593 }
1595 //------------------------------check_safepts----------------------------
1596 // Given dominators, try to find loops with calls that must always be
1597 // executed (call dominates loop tail). These loops do not need non-call
1598 // safepoints (ncsfpt).
1599 //
1600 // A complication is that a safepoint in a inner loop may be needed
1601 // by an outer loop. In the following, the inner loop sees it has a
1602 // call (block 3) on every path from the head (block 2) to the
1603 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1604 // in block 2, _but_ this leaves the outer loop without a safepoint.
1605 //
1606 // entry 0
1607 // |
1608 // v
1609 // outer 1,2 +->1
1610 // | |
1611 // | v
1612 // | 2<---+ ncsfpt in 2
1613 // |_/|\ |
1614 // | v |
1615 // inner 2,3 / 3 | call in 3
1616 // / | |
1617 // v +--+
1618 // exit 4
1619 //
1620 //
1621 // This method creates a list (_required_safept) of ncsfpt nodes that must
1622 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1623 // is first looked for in the lists for the outer loops of the current loop.
1624 //
1625 // The insights into the problem:
1626 // A) counted loops are okay
1627 // B) innermost loops are okay (only an inner loop can delete
1628 // a ncsfpt needed by an outer loop)
1629 // C) a loop is immune from an inner loop deleting a safepoint
1630 // if the loop has a call on the idom-path
1631 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1632 // idom-path that is not in a nested loop
1633 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1634 // loop needs to be prevented from deletion by an inner loop
1635 //
1636 // There are two analyses:
1637 // 1) The first, and cheaper one, scans the loop body from
1638 // tail to head following the idom (immediate dominator)
1639 // chain, looking for the cases (C,D,E) above.
1640 // Since inner loops are scanned before outer loops, there is summary
1641 // information about inner loops. Inner loops can be skipped over
1642 // when the tail of an inner loop is encountered.
1643 //
1644 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1645 // the idom path (which is rare), scans all predecessor control paths
1646 // from the tail to the head, terminating a path when a call or sfpt
1647 // is encountered, to find the ncsfpt's that are closest to the tail.
1648 //
1649 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1650 // Bottom up traversal
1651 IdealLoopTree* ch = _child;
1652 if (_child) _child->check_safepts(visited, stack);
1653 if (_next) _next ->check_safepts(visited, stack);
1655 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1656 bool has_call = false; // call on dom-path
1657 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1658 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1659 // Scan the dom-path nodes from tail to head
1660 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1661 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1662 has_call = true;
1663 _has_sfpt = 1; // Then no need for a safept!
1664 break;
1665 } else if (n->Opcode() == Op_SafePoint) {
1666 if (_phase->get_loop(n) == this) {
1667 has_local_ncsfpt = true;
1668 break;
1669 }
1670 if (nonlocal_ncsfpt == NULL) {
1671 nonlocal_ncsfpt = n; // save the one closest to the tail
1672 }
1673 } else {
1674 IdealLoopTree* nlpt = _phase->get_loop(n);
1675 if (this != nlpt) {
1676 // If at an inner loop tail, see if the inner loop has already
1677 // recorded seeing a call on the dom-path (and stop.) If not,
1678 // jump to the head of the inner loop.
1679 assert(is_member(nlpt), "nested loop");
1680 Node* tail = nlpt->_tail;
1681 if (tail->in(0)->is_If()) tail = tail->in(0);
1682 if (n == tail) {
1683 // If inner loop has call on dom-path, so does outer loop
1684 if (nlpt->_has_sfpt) {
1685 has_call = true;
1686 _has_sfpt = 1;
1687 break;
1688 }
1689 // Skip to head of inner loop
1690 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1691 n = nlpt->_head;
1692 }
1693 }
1694 }
1695 }
1696 // Record safept's that this loop needs preserved when an
1697 // inner loop attempts to delete it's safepoints.
1698 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1699 if (nonlocal_ncsfpt != NULL) {
1700 if (_required_safept == NULL) _required_safept = new Node_List();
1701 _required_safept->push(nonlocal_ncsfpt);
1702 } else {
1703 // Failed to find a suitable safept on the dom-path. Now use
1704 // an all paths walk from tail to head, looking for safepoints to preserve.
1705 allpaths_check_safepts(visited, stack);
1706 }
1707 }
1708 }
1709 }
1711 //---------------------------is_deleteable_safept----------------------------
1712 // Is safept not required by an outer loop?
1713 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1714 assert(sfpt->Opcode() == Op_SafePoint, "");
1715 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1716 while (lp != NULL) {
1717 Node_List* sfpts = lp->_required_safept;
1718 if (sfpts != NULL) {
1719 for (uint i = 0; i < sfpts->size(); i++) {
1720 if (sfpt == sfpts->at(i))
1721 return false;
1722 }
1723 }
1724 lp = lp->_parent;
1725 }
1726 return true;
1727 }
1729 //---------------------------replace_parallel_iv-------------------------------
1730 // Replace parallel induction variable (parallel to trip counter)
1731 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1732 assert(loop->_head->is_CountedLoop(), "");
1733 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1734 if (!cl->is_valid_counted_loop())
1735 return; // skip malformed counted loop
1736 Node *incr = cl->incr();
1737 if (incr == NULL)
1738 return; // Dead loop?
1739 Node *init = cl->init_trip();
1740 Node *phi = cl->phi();
1741 int stride_con = cl->stride_con();
1743 // Visit all children, looking for Phis
1744 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1745 Node *out = cl->out(i);
1746 // Look for other phis (secondary IVs). Skip dead ones
1747 if (!out->is_Phi() || out == phi || !has_node(out))
1748 continue;
1749 PhiNode* phi2 = out->as_Phi();
1750 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1751 // Look for induction variables of the form: X += constant
1752 if (phi2->region() != loop->_head ||
1753 incr2->req() != 3 ||
1754 incr2->in(1) != phi2 ||
1755 incr2 == incr ||
1756 incr2->Opcode() != Op_AddI ||
1757 !incr2->in(2)->is_Con())
1758 continue;
1760 // Check for parallel induction variable (parallel to trip counter)
1761 // via an affine function. In particular, count-down loops with
1762 // count-up array indices are common. We only RCE references off
1763 // the trip-counter, so we need to convert all these to trip-counter
1764 // expressions.
1765 Node *init2 = phi2->in( LoopNode::EntryControl );
1766 int stride_con2 = incr2->in(2)->get_int();
1768 // The general case here gets a little tricky. We want to find the
1769 // GCD of all possible parallel IV's and make a new IV using this
1770 // GCD for the loop. Then all possible IVs are simple multiples of
1771 // the GCD. In practice, this will cover very few extra loops.
1772 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1773 // where +/-1 is the common case, but other integer multiples are
1774 // also easy to handle.
1775 int ratio_con = stride_con2/stride_con;
1777 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1778 #ifndef PRODUCT
1779 if (TraceLoopOpts) {
1780 tty->print("Parallel IV: %d ", phi2->_idx);
1781 loop->dump_head();
1782 }
1783 #endif
1784 // Convert to using the trip counter. The parallel induction
1785 // variable differs from the trip counter by a loop-invariant
1786 // amount, the difference between their respective initial values.
1787 // It is scaled by the 'ratio_con'.
1788 Node* ratio = _igvn.intcon(ratio_con);
1789 set_ctrl(ratio, C->root());
1790 Node* ratio_init = new (C) MulINode(init, ratio);
1791 _igvn.register_new_node_with_optimizer(ratio_init, init);
1792 set_early_ctrl(ratio_init);
1793 Node* diff = new (C) SubINode(init2, ratio_init);
1794 _igvn.register_new_node_with_optimizer(diff, init2);
1795 set_early_ctrl(diff);
1796 Node* ratio_idx = new (C) MulINode(phi, ratio);
1797 _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1798 set_ctrl(ratio_idx, cl);
1799 Node* add = new (C) AddINode(ratio_idx, diff);
1800 _igvn.register_new_node_with_optimizer(add);
1801 set_ctrl(add, cl);
1802 _igvn.replace_node( phi2, add );
1803 // Sometimes an induction variable is unused
1804 if (add->outcnt() == 0) {
1805 _igvn.remove_dead_node(add);
1806 }
1807 --i; // deleted this phi; rescan starting with next position
1808 continue;
1809 }
1810 }
1811 }
1813 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
1814 Node* keep = NULL;
1815 if (keep_one) {
1816 // Look for a safepoint on the idom-path.
1817 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
1818 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
1819 keep = i;
1820 break; // Found one
1821 }
1822 }
1823 }
1825 // Don't remove any safepoints if it is requested to keep a single safepoint and
1826 // no safepoint was found on idom-path. It is not safe to remove any safepoint
1827 // in this case since there's no safepoint dominating all paths in the loop body.
1828 bool prune = !keep_one || keep != NULL;
1830 // Delete other safepoints in this loop.
1831 Node_List* sfpts = _safepts;
1832 if (prune && sfpts != NULL) {
1833 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
1834 for (uint i = 0; i < sfpts->size(); i++) {
1835 Node* n = sfpts->at(i);
1836 assert(phase->get_loop(n) == this, "");
1837 if (n != keep && phase->is_deleteable_safept(n)) {
1838 phase->lazy_replace(n, n->in(TypeFunc::Control));
1839 }
1840 }
1841 }
1842 }
1844 //------------------------------counted_loop-----------------------------------
1845 // Convert to counted loops where possible
1846 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1848 // For grins, set the inner-loop flag here
1849 if (!_child) {
1850 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1851 }
1853 if (_head->is_CountedLoop() ||
1854 phase->is_counted_loop(_head, this)) {
1856 if (!UseCountedLoopSafepoints) {
1857 // Indicate we do not need a safepoint here
1858 _has_sfpt = 1;
1859 }
1861 // Remove safepoints
1862 bool keep_one_sfpt = !(_has_call || _has_sfpt);
1863 remove_safepoints(phase, keep_one_sfpt);
1865 // Look for induction variables
1866 phase->replace_parallel_iv(this);
1868 } else if (_parent != NULL && !_irreducible) {
1869 // Not a counted loop. Keep one safepoint.
1870 bool keep_one_sfpt = true;
1871 remove_safepoints(phase, keep_one_sfpt);
1872 }
1874 // Recursively
1875 if (_child) _child->counted_loop( phase );
1876 if (_next) _next ->counted_loop( phase );
1877 }
1879 #ifndef PRODUCT
1880 //------------------------------dump_head--------------------------------------
1881 // Dump 1 liner for loop header info
1882 void IdealLoopTree::dump_head( ) const {
1883 for (uint i=0; i<_nest; i++)
1884 tty->print(" ");
1885 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1886 if (_irreducible) tty->print(" IRREDUCIBLE");
1887 Node* entry = _head->in(LoopNode::EntryControl);
1888 if (LoopLimitCheck) {
1889 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1890 if (predicate != NULL ) {
1891 tty->print(" limit_check");
1892 entry = entry->in(0)->in(0);
1893 }
1894 }
1895 if (UseLoopPredicate) {
1896 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1897 if (entry != NULL) {
1898 tty->print(" predicated");
1899 }
1900 }
1901 if (_head->is_CountedLoop()) {
1902 CountedLoopNode *cl = _head->as_CountedLoop();
1903 tty->print(" counted");
1905 Node* init_n = cl->init_trip();
1906 if (init_n != NULL && init_n->is_Con())
1907 tty->print(" [%d,", cl->init_trip()->get_int());
1908 else
1909 tty->print(" [int,");
1910 Node* limit_n = cl->limit();
1911 if (limit_n != NULL && limit_n->is_Con())
1912 tty->print("%d),", cl->limit()->get_int());
1913 else
1914 tty->print("int),");
1915 int stride_con = cl->stride_con();
1916 if (stride_con > 0) tty->print("+");
1917 tty->print("%d", stride_con);
1919 tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());
1921 if (cl->is_pre_loop ()) tty->print(" pre" );
1922 if (cl->is_main_loop()) tty->print(" main");
1923 if (cl->is_post_loop()) tty->print(" post");
1924 }
1925 if (_has_call) tty->print(" has_call");
1926 if (_has_sfpt) tty->print(" has_sfpt");
1927 if (_rce_candidate) tty->print(" rce");
1928 if (_safepts != NULL && _safepts->size() > 0) {
1929 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
1930 }
1931 if (_required_safept != NULL && _required_safept->size() > 0) {
1932 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
1933 }
1934 tty->cr();
1935 }
1937 //------------------------------dump-------------------------------------------
1938 // Dump loops by loop tree
1939 void IdealLoopTree::dump( ) const {
1940 dump_head();
1941 if (_child) _child->dump();
1942 if (_next) _next ->dump();
1943 }
1945 #endif
1947 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1948 if (loop == root) {
1949 if (loop->_child != NULL) {
1950 log->begin_head("loop_tree");
1951 log->end_head();
1952 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1953 log->tail("loop_tree");
1954 assert(loop->_next == NULL, "what?");
1955 }
1956 } else {
1957 Node* head = loop->_head;
1958 log->begin_head("loop");
1959 log->print(" idx='%d' ", head->_idx);
1960 if (loop->_irreducible) log->print("irreducible='1' ");
1961 if (head->is_Loop()) {
1962 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1963 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1964 }
1965 if (head->is_CountedLoop()) {
1966 CountedLoopNode* cl = head->as_CountedLoop();
1967 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1968 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1969 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1970 }
1971 log->end_head();
1972 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1973 log->tail("loop");
1974 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1975 }
1976 }
1978 //---------------------collect_potentially_useful_predicates-----------------------
1979 // Helper function to collect potentially useful predicates to prevent them from
1980 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1981 void PhaseIdealLoop::collect_potentially_useful_predicates(
1982 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1983 if (loop->_child) { // child
1984 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1985 }
1987 // self (only loops that we can apply loop predication may use their predicates)
1988 if (loop->_head->is_Loop() &&
1989 !loop->_irreducible &&
1990 !loop->tail()->is_top()) {
1991 LoopNode* lpn = loop->_head->as_Loop();
1992 Node* entry = lpn->in(LoopNode::EntryControl);
1993 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
1994 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1995 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1996 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1997 entry = entry->in(0)->in(0);
1998 }
1999 predicate_proj = find_predicate(entry); // Predicate
2000 if (predicate_proj != NULL ) {
2001 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2002 }
2003 }
2005 if (loop->_next) { // sibling
2006 collect_potentially_useful_predicates(loop->_next, useful_predicates);
2007 }
2008 }
2010 //------------------------eliminate_useless_predicates-----------------------------
2011 // Eliminate all inserted predicates if they could not be used by loop predication.
2012 // Note: it will also eliminates loop limits check predicate since it also uses
2013 // Opaque1 node (see Parse::add_predicate()).
2014 void PhaseIdealLoop::eliminate_useless_predicates() {
2015 if (C->predicate_count() == 0)
2016 return; // no predicate left
2018 Unique_Node_List useful_predicates; // to store useful predicates
2019 if (C->has_loops()) {
2020 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2021 }
2023 for (int i = C->predicate_count(); i > 0; i--) {
2024 Node * n = C->predicate_opaque1_node(i-1);
2025 assert(n->Opcode() == Op_Opaque1, "must be");
2026 if (!useful_predicates.member(n)) { // not in the useful list
2027 _igvn.replace_node(n, n->in(1));
2028 }
2029 }
2030 }
2032 //------------------------process_expensive_nodes-----------------------------
2033 // Expensive nodes have their control input set to prevent the GVN
2034 // from commoning them and as a result forcing the resulting node to
2035 // be in a more frequent path. Use CFG information here, to change the
2036 // control inputs so that some expensive nodes can be commoned while
2037 // not executed more frequently.
2038 bool PhaseIdealLoop::process_expensive_nodes() {
2039 assert(OptimizeExpensiveOps, "optimization off?");
2041 // Sort nodes to bring similar nodes together
2042 C->sort_expensive_nodes();
2044 bool progress = false;
2046 for (int i = 0; i < C->expensive_count(); ) {
2047 Node* n = C->expensive_node(i);
2048 int start = i;
2049 // Find nodes similar to n
2050 i++;
2051 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2052 int end = i;
2053 // And compare them two by two
2054 for (int j = start; j < end; j++) {
2055 Node* n1 = C->expensive_node(j);
2056 if (is_node_unreachable(n1)) {
2057 continue;
2058 }
2059 for (int k = j+1; k < end; k++) {
2060 Node* n2 = C->expensive_node(k);
2061 if (is_node_unreachable(n2)) {
2062 continue;
2063 }
2065 assert(n1 != n2, "should be pair of nodes");
2067 Node* c1 = n1->in(0);
2068 Node* c2 = n2->in(0);
2070 Node* parent_c1 = c1;
2071 Node* parent_c2 = c2;
2073 // The call to get_early_ctrl_for_expensive() moves the
2074 // expensive nodes up but stops at loops that are in a if
2075 // branch. See whether we can exit the loop and move above the
2076 // If.
2077 if (c1->is_Loop()) {
2078 parent_c1 = c1->in(1);
2079 }
2080 if (c2->is_Loop()) {
2081 parent_c2 = c2->in(1);
2082 }
2084 if (parent_c1 == parent_c2) {
2085 _igvn._worklist.push(n1);
2086 _igvn._worklist.push(n2);
2087 continue;
2088 }
2090 // Look for identical expensive node up the dominator chain.
2091 if (is_dominator(c1, c2)) {
2092 c2 = c1;
2093 } else if (is_dominator(c2, c1)) {
2094 c1 = c2;
2095 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2096 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2097 // Both branches have the same expensive node so move it up
2098 // before the if.
2099 c1 = c2 = idom(parent_c1->in(0));
2100 }
2101 // Do the actual moves
2102 if (n1->in(0) != c1) {
2103 _igvn.hash_delete(n1);
2104 n1->set_req(0, c1);
2105 _igvn.hash_insert(n1);
2106 _igvn._worklist.push(n1);
2107 progress = true;
2108 }
2109 if (n2->in(0) != c2) {
2110 _igvn.hash_delete(n2);
2111 n2->set_req(0, c2);
2112 _igvn.hash_insert(n2);
2113 _igvn._worklist.push(n2);
2114 progress = true;
2115 }
2116 }
2117 }
2118 }
2120 return progress;
2121 }
2124 //=============================================================================
2125 //----------------------------build_and_optimize-------------------------------
2126 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
2127 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
2128 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
2129 ResourceMark rm;
2131 int old_progress = C->major_progress();
2132 uint orig_worklist_size = _igvn._worklist.size();
2134 // Reset major-progress flag for the driver's heuristics
2135 C->clear_major_progress();
2137 #ifndef PRODUCT
2138 // Capture for later assert
2139 uint unique = C->unique();
2140 _loop_invokes++;
2141 _loop_work += unique;
2142 #endif
2144 // True if the method has at least 1 irreducible loop
2145 _has_irreducible_loops = false;
2147 _created_loop_node = false;
2149 Arena *a = Thread::current()->resource_area();
2150 VectorSet visited(a);
2151 // Pre-grow the mapping from Nodes to IdealLoopTrees.
2152 _nodes.map(C->unique(), NULL);
2153 memset(_nodes.adr(), 0, wordSize * C->unique());
2155 // Pre-build the top-level outermost loop tree entry
2156 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2157 // Do not need a safepoint at the top level
2158 _ltree_root->_has_sfpt = 1;
2160 // Initialize Dominators.
2161 // Checked in clone_loop_predicate() during beautify_loops().
2162 _idom_size = 0;
2163 _idom = NULL;
2164 _dom_depth = NULL;
2165 _dom_stk = NULL;
2167 // Empty pre-order array
2168 allocate_preorders();
2170 // Build a loop tree on the fly. Build a mapping from CFG nodes to
2171 // IdealLoopTree entries. Data nodes are NOT walked.
2172 build_loop_tree();
2173 // Check for bailout, and return
2174 if (C->failing()) {
2175 return;
2176 }
2178 // No loops after all
2179 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2181 // There should always be an outer loop containing the Root and Return nodes.
2182 // If not, we have a degenerate empty program. Bail out in this case.
2183 if (!has_node(C->root())) {
2184 if (!_verify_only) {
2185 C->clear_major_progress();
2186 C->record_method_not_compilable("empty program detected during loop optimization");
2187 }
2188 return;
2189 }
2191 // Nothing to do, so get out
2192 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
2193 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2194 if (stop_early && !do_expensive_nodes) {
2195 _igvn.optimize(); // Cleanup NeverBranches
2196 return;
2197 }
2199 // Set loop nesting depth
2200 _ltree_root->set_nest( 0 );
2202 // Split shared headers and insert loop landing pads.
2203 // Do not bother doing this on the Root loop of course.
2204 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2205 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2206 if( _ltree_root->_child->beautify_loops( this ) ) {
2207 // Re-build loop tree!
2208 _ltree_root->_child = NULL;
2209 _nodes.clear();
2210 reallocate_preorders();
2211 build_loop_tree();
2212 // Check for bailout, and return
2213 if (C->failing()) {
2214 return;
2215 }
2216 // Reset loop nesting depth
2217 _ltree_root->set_nest( 0 );
2219 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2220 }
2221 }
2223 // Build Dominators for elision of NULL checks & loop finding.
2224 // Since nodes do not have a slot for immediate dominator, make
2225 // a persistent side array for that info indexed on node->_idx.
2226 _idom_size = C->unique();
2227 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2228 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
2229 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
2230 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2232 Dominators();
2234 if (!_verify_only) {
2235 // As a side effect, Dominators removed any unreachable CFG paths
2236 // into RegionNodes. It doesn't do this test against Root, so
2237 // we do it here.
2238 for( uint i = 1; i < C->root()->req(); i++ ) {
2239 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
2240 _igvn.delete_input_of(C->root(), i);
2241 i--; // Rerun same iteration on compressed edges
2242 }
2243 }
2245 // Given dominators, try to find inner loops with calls that must
2246 // always be executed (call dominates loop tail). These loops do
2247 // not need a separate safepoint.
2248 Node_List cisstack(a);
2249 _ltree_root->check_safepts(visited, cisstack);
2250 }
2252 // Walk the DATA nodes and place into loops. Find earliest control
2253 // node. For CFG nodes, the _nodes array starts out and remains
2254 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2255 // _nodes array holds the earliest legal controlling CFG node.
2257 // Allocate stack with enough space to avoid frequent realloc
2258 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2259 Node_Stack nstack( a, stack_size );
2261 visited.Clear();
2262 Node_List worklist(a);
2263 // Don't need C->root() on worklist since
2264 // it will be processed among C->top() inputs
2265 worklist.push( C->top() );
2266 visited.set( C->top()->_idx ); // Set C->top() as visited now
2267 build_loop_early( visited, worklist, nstack );
2269 // Given early legal placement, try finding counted loops. This placement
2270 // is good enough to discover most loop invariants.
2271 if( !_verify_me && !_verify_only )
2272 _ltree_root->counted_loop( this );
2274 // Find latest loop placement. Find ideal loop placement.
2275 visited.Clear();
2276 init_dom_lca_tags();
2277 // Need C->root() on worklist when processing outs
2278 worklist.push( C->root() );
2279 NOT_PRODUCT( C->verify_graph_edges(); )
2280 worklist.push( C->top() );
2281 build_loop_late( visited, worklist, nstack );
2283 if (_verify_only) {
2284 // restore major progress flag
2285 for (int i = 0; i < old_progress; i++)
2286 C->set_major_progress();
2287 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2288 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2289 return;
2290 }
2292 // clear out the dead code after build_loop_late
2293 while (_deadlist.size()) {
2294 _igvn.remove_globally_dead_node(_deadlist.pop());
2295 }
2297 if (stop_early) {
2298 assert(do_expensive_nodes, "why are we here?");
2299 if (process_expensive_nodes()) {
2300 // If we made some progress when processing expensive nodes then
2301 // the IGVN may modify the graph in a way that will allow us to
2302 // make some more progress: we need to try processing expensive
2303 // nodes again.
2304 C->set_major_progress();
2305 }
2306 _igvn.optimize();
2307 return;
2308 }
2310 // Some parser-inserted loop predicates could never be used by loop
2311 // predication or they were moved away from loop during some optimizations.
2312 // For example, peeling. Eliminate them before next loop optimizations.
2313 if (UseLoopPredicate || LoopLimitCheck) {
2314 eliminate_useless_predicates();
2315 }
2317 #ifndef PRODUCT
2318 C->verify_graph_edges();
2319 if (_verify_me) { // Nested verify pass?
2320 // Check to see if the verify mode is broken
2321 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2322 return;
2323 }
2324 if(VerifyLoopOptimizations) verify();
2325 if(TraceLoopOpts && C->has_loops()) {
2326 _ltree_root->dump();
2327 }
2328 #endif
2330 if (skip_loop_opts) {
2331 // restore major progress flag
2332 for (int i = 0; i < old_progress; i++) {
2333 C->set_major_progress();
2334 }
2336 // Cleanup any modified bits
2337 _igvn.optimize();
2339 if (C->log() != NULL) {
2340 log_loop_tree(_ltree_root, _ltree_root, C->log());
2341 }
2342 return;
2343 }
2345 if (ReassociateInvariants) {
2346 // Reassociate invariants and prep for split_thru_phi
2347 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2348 IdealLoopTree* lpt = iter.current();
2349 if (!lpt->is_counted() || !lpt->is_inner()) continue;
2351 lpt->reassociate_invariants(this);
2353 // Because RCE opportunities can be masked by split_thru_phi,
2354 // look for RCE candidates and inhibit split_thru_phi
2355 // on just their loop-phi's for this pass of loop opts
2356 if (SplitIfBlocks && do_split_ifs) {
2357 if (lpt->policy_range_check(this)) {
2358 lpt->_rce_candidate = 1; // = true
2359 }
2360 }
2361 }
2362 }
2364 // Check for aggressive application of split-if and other transforms
2365 // that require basic-block info (like cloning through Phi's)
2366 if( SplitIfBlocks && do_split_ifs ) {
2367 visited.Clear();
2368 split_if_with_blocks( visited, nstack );
2369 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2370 }
2372 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2373 C->set_major_progress();
2374 }
2376 // Perform loop predication before iteration splitting
2377 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2378 _ltree_root->_child->loop_predication(this);
2379 }
2381 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2382 if (do_intrinsify_fill()) {
2383 C->set_major_progress();
2384 }
2385 }
2387 // Perform iteration-splitting on inner loops. Split iterations to avoid
2388 // range checks or one-shot null checks.
2390 // If split-if's didn't hack the graph too bad (no CFG changes)
2391 // then do loop opts.
2392 if (C->has_loops() && !C->major_progress()) {
2393 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2394 _ltree_root->_child->iteration_split( this, worklist );
2395 // No verify after peeling! GCM has hoisted code out of the loop.
2396 // After peeling, the hoisted code could sink inside the peeled area.
2397 // The peeling code does not try to recompute the best location for
2398 // all the code before the peeled area, so the verify pass will always
2399 // complain about it.
2400 }
2401 // Do verify graph edges in any case
2402 NOT_PRODUCT( C->verify_graph_edges(); );
2404 if (!do_split_ifs) {
2405 // We saw major progress in Split-If to get here. We forced a
2406 // pass with unrolling and not split-if, however more split-if's
2407 // might make progress. If the unrolling didn't make progress
2408 // then the major-progress flag got cleared and we won't try
2409 // another round of Split-If. In particular the ever-common
2410 // instance-of/check-cast pattern requires at least 2 rounds of
2411 // Split-If to clear out.
2412 C->set_major_progress();
2413 }
2415 // Repeat loop optimizations if new loops were seen
2416 if (created_loop_node()) {
2417 C->set_major_progress();
2418 }
2420 // Keep loop predicates and perform optimizations with them
2421 // until no more loop optimizations could be done.
2422 // After that switch predicates off and do more loop optimizations.
2423 if (!C->major_progress() && (C->predicate_count() > 0)) {
2424 C->cleanup_loop_predicates(_igvn);
2425 #ifndef PRODUCT
2426 if (TraceLoopOpts) {
2427 tty->print_cr("PredicatesOff");
2428 }
2429 #endif
2430 C->set_major_progress();
2431 }
2433 // Convert scalar to superword operations at the end of all loop opts.
2434 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2435 // SuperWord transform
2436 SuperWord sw(this);
2437 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2438 IdealLoopTree* lpt = iter.current();
2439 if (lpt->is_counted()) {
2440 sw.transform_loop(lpt);
2441 }
2442 }
2443 }
2445 // Cleanup any modified bits
2446 _igvn.optimize();
2448 // disable assert until issue with split_flow_path is resolved (6742111)
2449 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2450 // "shouldn't introduce irreducible loops");
2452 if (C->log() != NULL) {
2453 log_loop_tree(_ltree_root, _ltree_root, C->log());
2454 }
2455 }
2457 #ifndef PRODUCT
2458 //------------------------------print_statistics-------------------------------
2459 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2460 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2461 void PhaseIdealLoop::print_statistics() {
2462 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2463 }
2465 //------------------------------verify-----------------------------------------
2466 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2467 static int fail; // debug only, so its multi-thread dont care
2468 void PhaseIdealLoop::verify() const {
2469 int old_progress = C->major_progress();
2470 ResourceMark rm;
2471 PhaseIdealLoop loop_verify( _igvn, this );
2472 VectorSet visited(Thread::current()->resource_area());
2474 fail = 0;
2475 verify_compare( C->root(), &loop_verify, visited );
2476 assert( fail == 0, "verify loops failed" );
2477 // Verify loop structure is the same
2478 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2479 // Reset major-progress. It was cleared by creating a verify version of
2480 // PhaseIdealLoop.
2481 for( int i=0; i<old_progress; i++ )
2482 C->set_major_progress();
2483 }
2485 //------------------------------verify_compare---------------------------------
2486 // Make sure me and the given PhaseIdealLoop agree on key data structures
2487 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2488 if( !n ) return;
2489 if( visited.test_set( n->_idx ) ) return;
2490 if( !_nodes[n->_idx] ) { // Unreachable
2491 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2492 return;
2493 }
2495 uint i;
2496 for( i = 0; i < n->req(); i++ )
2497 verify_compare( n->in(i), loop_verify, visited );
2499 // Check the '_nodes' block/loop structure
2500 i = n->_idx;
2501 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2502 if( _nodes[i] != loop_verify->_nodes[i] &&
2503 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2504 tty->print("Mismatched control setting for: ");
2505 n->dump();
2506 if( fail++ > 10 ) return;
2507 Node *c = get_ctrl_no_update(n);
2508 tty->print("We have it as: ");
2509 if( c->in(0) ) c->dump();
2510 else tty->print_cr("N%d",c->_idx);
2511 tty->print("Verify thinks: ");
2512 if( loop_verify->has_ctrl(n) )
2513 loop_verify->get_ctrl_no_update(n)->dump();
2514 else
2515 loop_verify->get_loop_idx(n)->dump();
2516 tty->cr();
2517 }
2518 } else { // We have a loop
2519 IdealLoopTree *us = get_loop_idx(n);
2520 if( loop_verify->has_ctrl(n) ) {
2521 tty->print("Mismatched loop setting for: ");
2522 n->dump();
2523 if( fail++ > 10 ) return;
2524 tty->print("We have it as: ");
2525 us->dump();
2526 tty->print("Verify thinks: ");
2527 loop_verify->get_ctrl_no_update(n)->dump();
2528 tty->cr();
2529 } else if (!C->major_progress()) {
2530 // Loop selection can be messed up if we did a major progress
2531 // operation, like split-if. Do not verify in that case.
2532 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2533 if( us->_head != them->_head || us->_tail != them->_tail ) {
2534 tty->print("Unequals loops for: ");
2535 n->dump();
2536 if( fail++ > 10 ) return;
2537 tty->print("We have it as: ");
2538 us->dump();
2539 tty->print("Verify thinks: ");
2540 them->dump();
2541 tty->cr();
2542 }
2543 }
2544 }
2546 // Check for immediate dominators being equal
2547 if( i >= _idom_size ) {
2548 if( !n->is_CFG() ) return;
2549 tty->print("CFG Node with no idom: ");
2550 n->dump();
2551 return;
2552 }
2553 if( !n->is_CFG() ) return;
2554 if( n == C->root() ) return; // No IDOM here
2556 assert(n->_idx == i, "sanity");
2557 Node *id = idom_no_update(n);
2558 if( id != loop_verify->idom_no_update(n) ) {
2559 tty->print("Unequals idoms for: ");
2560 n->dump();
2561 if( fail++ > 10 ) return;
2562 tty->print("We have it as: ");
2563 id->dump();
2564 tty->print("Verify thinks: ");
2565 loop_verify->idom_no_update(n)->dump();
2566 tty->cr();
2567 }
2569 }
2571 //------------------------------verify_tree------------------------------------
2572 // Verify that tree structures match. Because the CFG can change, siblings
2573 // within the loop tree can be reordered. We attempt to deal with that by
2574 // reordering the verify's loop tree if possible.
2575 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2576 assert( _parent == parent, "Badly formed loop tree" );
2578 // Siblings not in same order? Attempt to re-order.
2579 if( _head != loop->_head ) {
2580 // Find _next pointer to update
2581 IdealLoopTree **pp = &loop->_parent->_child;
2582 while( *pp != loop )
2583 pp = &((*pp)->_next);
2584 // Find proper sibling to be next
2585 IdealLoopTree **nn = &loop->_next;
2586 while( (*nn) && (*nn)->_head != _head )
2587 nn = &((*nn)->_next);
2589 // Check for no match.
2590 if( !(*nn) ) {
2591 // Annoyingly, irreducible loops can pick different headers
2592 // after a major_progress operation, so the rest of the loop
2593 // tree cannot be matched.
2594 if (_irreducible && Compile::current()->major_progress()) return;
2595 assert( 0, "failed to match loop tree" );
2596 }
2598 // Move (*nn) to (*pp)
2599 IdealLoopTree *hit = *nn;
2600 *nn = hit->_next;
2601 hit->_next = loop;
2602 *pp = loop;
2603 loop = hit;
2604 // Now try again to verify
2605 }
2607 assert( _head == loop->_head , "mismatched loop head" );
2608 Node *tail = _tail; // Inline a non-updating version of
2609 while( !tail->in(0) ) // the 'tail()' call.
2610 tail = tail->in(1);
2611 assert( tail == loop->_tail, "mismatched loop tail" );
2613 // Counted loops that are guarded should be able to find their guards
2614 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2615 CountedLoopNode *cl = _head->as_CountedLoop();
2616 Node *init = cl->init_trip();
2617 Node *ctrl = cl->in(LoopNode::EntryControl);
2618 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2619 Node *iff = ctrl->in(0);
2620 assert( iff->Opcode() == Op_If, "" );
2621 Node *bol = iff->in(1);
2622 assert( bol->Opcode() == Op_Bool, "" );
2623 Node *cmp = bol->in(1);
2624 assert( cmp->Opcode() == Op_CmpI, "" );
2625 Node *add = cmp->in(1);
2626 Node *opaq;
2627 if( add->Opcode() == Op_Opaque1 ) {
2628 opaq = add;
2629 } else {
2630 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2631 assert( add == init, "" );
2632 opaq = cmp->in(2);
2633 }
2634 assert( opaq->Opcode() == Op_Opaque1, "" );
2636 }
2638 if (_child != NULL) _child->verify_tree(loop->_child, this);
2639 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2640 // Innermost loops need to verify loop bodies,
2641 // but only if no 'major_progress'
2642 int fail = 0;
2643 if (!Compile::current()->major_progress() && _child == NULL) {
2644 for( uint i = 0; i < _body.size(); i++ ) {
2645 Node *n = _body.at(i);
2646 if (n->outcnt() == 0) continue; // Ignore dead
2647 uint j;
2648 for( j = 0; j < loop->_body.size(); j++ )
2649 if( loop->_body.at(j) == n )
2650 break;
2651 if( j == loop->_body.size() ) { // Not found in loop body
2652 // Last ditch effort to avoid assertion: Its possible that we
2653 // have some users (so outcnt not zero) but are still dead.
2654 // Try to find from root.
2655 if (Compile::current()->root()->find(n->_idx)) {
2656 fail++;
2657 tty->print("We have that verify does not: ");
2658 n->dump();
2659 }
2660 }
2661 }
2662 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2663 Node *n = loop->_body.at(i2);
2664 if (n->outcnt() == 0) continue; // Ignore dead
2665 uint j;
2666 for( j = 0; j < _body.size(); j++ )
2667 if( _body.at(j) == n )
2668 break;
2669 if( j == _body.size() ) { // Not found in loop body
2670 // Last ditch effort to avoid assertion: Its possible that we
2671 // have some users (so outcnt not zero) but are still dead.
2672 // Try to find from root.
2673 if (Compile::current()->root()->find(n->_idx)) {
2674 fail++;
2675 tty->print("Verify has that we do not: ");
2676 n->dump();
2677 }
2678 }
2679 }
2680 assert( !fail, "loop body mismatch" );
2681 }
2682 }
2684 #endif
2686 //------------------------------set_idom---------------------------------------
2687 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2688 uint idx = d->_idx;
2689 if (idx >= _idom_size) {
2690 uint newsize = _idom_size<<1;
2691 while( idx >= newsize ) {
2692 newsize <<= 1;
2693 }
2694 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2695 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2696 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2697 _idom_size = newsize;
2698 }
2699 _idom[idx] = n;
2700 _dom_depth[idx] = dom_depth;
2701 }
2703 //------------------------------recompute_dom_depth---------------------------------------
2704 // The dominator tree is constructed with only parent pointers.
2705 // This recomputes the depth in the tree by first tagging all
2706 // nodes as "no depth yet" marker. The next pass then runs up
2707 // the dom tree from each node marked "no depth yet", and computes
2708 // the depth on the way back down.
2709 void PhaseIdealLoop::recompute_dom_depth() {
2710 uint no_depth_marker = C->unique();
2711 uint i;
2712 // Initialize depth to "no depth yet"
2713 for (i = 0; i < _idom_size; i++) {
2714 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2715 _dom_depth[i] = no_depth_marker;
2716 }
2717 }
2718 if (_dom_stk == NULL) {
2719 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
2720 if (init_size < 10) init_size = 10;
2721 _dom_stk = new GrowableArray<uint>(init_size);
2722 }
2723 // Compute new depth for each node.
2724 for (i = 0; i < _idom_size; i++) {
2725 uint j = i;
2726 // Run up the dom tree to find a node with a depth
2727 while (_dom_depth[j] == no_depth_marker) {
2728 _dom_stk->push(j);
2729 j = _idom[j]->_idx;
2730 }
2731 // Compute the depth on the way back down this tree branch
2732 uint dd = _dom_depth[j] + 1;
2733 while (_dom_stk->length() > 0) {
2734 uint j = _dom_stk->pop();
2735 _dom_depth[j] = dd;
2736 dd++;
2737 }
2738 }
2739 }
2741 //------------------------------sort-------------------------------------------
2742 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2743 // loop tree, not the root.
2744 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2745 if( !innermost ) return loop; // New innermost loop
2747 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2748 assert( loop_preorder, "not yet post-walked loop" );
2749 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2750 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2752 // Insert at start of list
2753 while( l ) { // Insertion sort based on pre-order
2754 if( l == loop ) return innermost; // Already on list!
2755 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2756 assert( l_preorder, "not yet post-walked l" );
2757 // Check header pre-order number to figure proper nesting
2758 if( loop_preorder > l_preorder )
2759 break; // End of insertion
2760 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2761 // Since I split shared headers, you'd think this could not happen.
2762 // BUT: I must first do the preorder numbering before I can discover I
2763 // have shared headers, so the split headers all get the same preorder
2764 // number as the RegionNode they split from.
2765 if( loop_preorder == l_preorder &&
2766 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2767 break; // Also check for shared headers (same pre#)
2768 pp = &l->_parent; // Chain up list
2769 l = *pp;
2770 }
2771 // Link into list
2772 // Point predecessor to me
2773 *pp = loop;
2774 // Point me to successor
2775 IdealLoopTree *p = loop->_parent;
2776 loop->_parent = l; // Point me to successor
2777 if( p ) sort( p, innermost ); // Insert my parents into list as well
2778 return innermost;
2779 }
2781 //------------------------------build_loop_tree--------------------------------
2782 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2783 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2784 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2785 // tightest enclosing IdealLoopTree for post-walked.
2786 //
2787 // During my forward walk I do a short 1-layer lookahead to see if I can find
2788 // a loop backedge with that doesn't have any work on the backedge. This
2789 // helps me construct nested loops with shared headers better.
2790 //
2791 // Once I've done the forward recursion, I do the post-work. For each child
2792 // I check to see if there is a backedge. Backedges define a loop! I
2793 // insert an IdealLoopTree at the target of the backedge.
2794 //
2795 // During the post-work I also check to see if I have several children
2796 // belonging to different loops. If so, then this Node is a decision point
2797 // where control flow can choose to change loop nests. It is at this
2798 // decision point where I can figure out how loops are nested. At this
2799 // time I can properly order the different loop nests from my children.
2800 // Note that there may not be any backedges at the decision point!
2801 //
2802 // Since the decision point can be far removed from the backedges, I can't
2803 // order my loops at the time I discover them. Thus at the decision point
2804 // I need to inspect loop header pre-order numbers to properly nest my
2805 // loops. This means I need to sort my childrens' loops by pre-order.
2806 // The sort is of size number-of-control-children, which generally limits
2807 // it to size 2 (i.e., I just choose between my 2 target loops).
2808 void PhaseIdealLoop::build_loop_tree() {
2809 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
2810 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
2811 Node *n = C->root();
2812 bltstack.push(n);
2813 int pre_order = 1;
2814 int stack_size;
2816 while ( ( stack_size = bltstack.length() ) != 0 ) {
2817 n = bltstack.top(); // Leave node on stack
2818 if ( !is_visited(n) ) {
2819 // ---- Pre-pass Work ----
2820 // Pre-walked but not post-walked nodes need a pre_order number.
2822 set_preorder_visited( n, pre_order ); // set as visited
2824 // ---- Scan over children ----
2825 // Scan first over control projections that lead to loop headers.
2826 // This helps us find inner-to-outer loops with shared headers better.
2828 // Scan children's children for loop headers.
2829 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2830 Node* m = n->raw_out(i); // Child
2831 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2832 // Scan over children's children to find loop
2833 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2834 Node* l = m->fast_out(j);
2835 if( is_visited(l) && // Been visited?
2836 !is_postvisited(l) && // But not post-visited
2837 get_preorder(l) < pre_order ) { // And smaller pre-order
2838 // Found! Scan the DFS down this path before doing other paths
2839 bltstack.push(m);
2840 break;
2841 }
2842 }
2843 }
2844 }
2845 pre_order++;
2846 }
2847 else if ( !is_postvisited(n) ) {
2848 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2849 // such as com.sun.rsasign.am::a.
2850 // For non-recursive version, first, process current children.
2851 // On next iteration, check if additional children were added.
2852 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2853 Node* u = n->raw_out(k);
2854 if ( u->is_CFG() && !is_visited(u) ) {
2855 bltstack.push(u);
2856 }
2857 }
2858 if ( bltstack.length() == stack_size ) {
2859 // There were no additional children, post visit node now
2860 (void)bltstack.pop(); // Remove node from stack
2861 pre_order = build_loop_tree_impl( n, pre_order );
2862 // Check for bailout
2863 if (C->failing()) {
2864 return;
2865 }
2866 // Check to grow _preorders[] array for the case when
2867 // build_loop_tree_impl() adds new nodes.
2868 check_grow_preorders();
2869 }
2870 }
2871 else {
2872 (void)bltstack.pop(); // Remove post-visited node from stack
2873 }
2874 }
2875 }
2877 //------------------------------build_loop_tree_impl---------------------------
2878 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2879 // ---- Post-pass Work ----
2880 // Pre-walked but not post-walked nodes need a pre_order number.
2882 // Tightest enclosing loop for this Node
2883 IdealLoopTree *innermost = NULL;
2885 // For all children, see if any edge is a backedge. If so, make a loop
2886 // for it. Then find the tightest enclosing loop for the self Node.
2887 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2888 Node* m = n->fast_out(i); // Child
2889 if( n == m ) continue; // Ignore control self-cycles
2890 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2892 IdealLoopTree *l; // Child's loop
2893 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2894 // Found a backedge
2895 assert( get_preorder(m) < pre_order, "should be backedge" );
2896 // Check for the RootNode, which is already a LoopNode and is allowed
2897 // to have multiple "backedges".
2898 if( m == C->root()) { // Found the root?
2899 l = _ltree_root; // Root is the outermost LoopNode
2900 } else { // Else found a nested loop
2901 // Insert a LoopNode to mark this loop.
2902 l = new IdealLoopTree(this, m, n);
2903 } // End of Else found a nested loop
2904 if( !has_loop(m) ) // If 'm' does not already have a loop set
2905 set_loop(m, l); // Set loop header to loop now
2907 } else { // Else not a nested loop
2908 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2909 l = get_loop(m); // Get previously determined loop
2910 // If successor is header of a loop (nest), move up-loop till it
2911 // is a member of some outer enclosing loop. Since there are no
2912 // shared headers (I've split them already) I only need to go up
2913 // at most 1 level.
2914 while( l && l->_head == m ) // Successor heads loop?
2915 l = l->_parent; // Move up 1 for me
2916 // If this loop is not properly parented, then this loop
2917 // has no exit path out, i.e. its an infinite loop.
2918 if( !l ) {
2919 // Make loop "reachable" from root so the CFG is reachable. Basically
2920 // insert a bogus loop exit that is never taken. 'm', the loop head,
2921 // points to 'n', one (of possibly many) fall-in paths. There may be
2922 // many backedges as well.
2924 // Here I set the loop to be the root loop. I could have, after
2925 // inserting a bogus loop exit, restarted the recursion and found my
2926 // new loop exit. This would make the infinite loop a first-class
2927 // loop and it would then get properly optimized. What's the use of
2928 // optimizing an infinite loop?
2929 l = _ltree_root; // Oops, found infinite loop
2931 if (!_verify_only) {
2932 // Insert the NeverBranch between 'm' and it's control user.
2933 NeverBranchNode *iff = new (C) NeverBranchNode( m );
2934 _igvn.register_new_node_with_optimizer(iff);
2935 set_loop(iff, l);
2936 Node *if_t = new (C) CProjNode( iff, 0 );
2937 _igvn.register_new_node_with_optimizer(if_t);
2938 set_loop(if_t, l);
2940 Node* cfg = NULL; // Find the One True Control User of m
2941 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2942 Node* x = m->fast_out(j);
2943 if (x->is_CFG() && x != m && x != iff)
2944 { cfg = x; break; }
2945 }
2946 assert(cfg != NULL, "must find the control user of m");
2947 uint k = 0; // Probably cfg->in(0)
2948 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2949 cfg->set_req( k, if_t ); // Now point to NeverBranch
2951 // Now create the never-taken loop exit
2952 Node *if_f = new (C) CProjNode( iff, 1 );
2953 _igvn.register_new_node_with_optimizer(if_f);
2954 set_loop(if_f, l);
2955 // Find frame ptr for Halt. Relies on the optimizer
2956 // V-N'ing. Easier and quicker than searching through
2957 // the program structure.
2958 Node *frame = new (C) ParmNode( C->start(), TypeFunc::FramePtr );
2959 _igvn.register_new_node_with_optimizer(frame);
2960 // Halt & Catch Fire
2961 Node *halt = new (C) HaltNode( if_f, frame );
2962 _igvn.register_new_node_with_optimizer(halt);
2963 set_loop(halt, l);
2964 C->root()->add_req(halt);
2965 }
2966 set_loop(C->root(), _ltree_root);
2967 }
2968 }
2969 // Weeny check for irreducible. This child was already visited (this
2970 // IS the post-work phase). Is this child's loop header post-visited
2971 // as well? If so, then I found another entry into the loop.
2972 if (!_verify_only) {
2973 while( is_postvisited(l->_head) ) {
2974 // found irreducible
2975 l->_irreducible = 1; // = true
2976 l = l->_parent;
2977 _has_irreducible_loops = true;
2978 // Check for bad CFG here to prevent crash, and bailout of compile
2979 if (l == NULL) {
2980 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2981 return pre_order;
2982 }
2983 }
2984 C->set_has_irreducible_loop(_has_irreducible_loops);
2985 }
2987 // This Node might be a decision point for loops. It is only if
2988 // it's children belong to several different loops. The sort call
2989 // does a trivial amount of work if there is only 1 child or all
2990 // children belong to the same loop. If however, the children
2991 // belong to different loops, the sort call will properly set the
2992 // _parent pointers to show how the loops nest.
2993 //
2994 // In any case, it returns the tightest enclosing loop.
2995 innermost = sort( l, innermost );
2996 }
2998 // Def-use info will have some dead stuff; dead stuff will have no
2999 // loop decided on.
3001 // Am I a loop header? If so fix up my parent's child and next ptrs.
3002 if( innermost && innermost->_head == n ) {
3003 assert( get_loop(n) == innermost, "" );
3004 IdealLoopTree *p = innermost->_parent;
3005 IdealLoopTree *l = innermost;
3006 while( p && l->_head == n ) {
3007 l->_next = p->_child; // Put self on parents 'next child'
3008 p->_child = l; // Make self as first child of parent
3009 l = p; // Now walk up the parent chain
3010 p = l->_parent;
3011 }
3012 } else {
3013 // Note that it is possible for a LoopNode to reach here, if the
3014 // backedge has been made unreachable (hence the LoopNode no longer
3015 // denotes a Loop, and will eventually be removed).
3017 // Record tightest enclosing loop for self. Mark as post-visited.
3018 set_loop(n, innermost);
3019 // Also record has_call flag early on
3020 if( innermost ) {
3021 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3022 // Do not count uncommon calls
3023 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3024 Node *iff = n->in(0)->in(0);
3025 // No any calls for vectorized loops.
3026 if( UseSuperWord || !iff->is_If() ||
3027 (n->in(0)->Opcode() == Op_IfFalse &&
3028 (1.0 - iff->as_If()->_prob) >= 0.01) ||
3029 (iff->as_If()->_prob >= 0.01) )
3030 innermost->_has_call = 1;
3031 }
3032 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3033 // Disable loop optimizations if the loop has a scalar replaceable
3034 // allocation. This disabling may cause a potential performance lost
3035 // if the allocation is not eliminated for some reason.
3036 innermost->_allow_optimizations = false;
3037 innermost->_has_call = 1; // = true
3038 } else if (n->Opcode() == Op_SafePoint) {
3039 // Record all safepoints in this loop.
3040 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3041 innermost->_safepts->push(n);
3042 }
3043 }
3044 }
3046 // Flag as post-visited now
3047 set_postvisited(n);
3048 return pre_order;
3049 }
3052 //------------------------------build_loop_early-------------------------------
3053 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3054 // First pass computes the earliest controlling node possible. This is the
3055 // controlling input with the deepest dominating depth.
3056 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3057 while (worklist.size() != 0) {
3058 // Use local variables nstack_top_n & nstack_top_i to cache values
3059 // on nstack's top.
3060 Node *nstack_top_n = worklist.pop();
3061 uint nstack_top_i = 0;
3062 //while_nstack_nonempty:
3063 while (true) {
3064 // Get parent node and next input's index from stack's top.
3065 Node *n = nstack_top_n;
3066 uint i = nstack_top_i;
3067 uint cnt = n->req(); // Count of inputs
3068 if (i == 0) { // Pre-process the node.
3069 if( has_node(n) && // Have either loop or control already?
3070 !has_ctrl(n) ) { // Have loop picked out already?
3071 // During "merge_many_backedges" we fold up several nested loops
3072 // into a single loop. This makes the members of the original
3073 // loop bodies pointing to dead loops; they need to move up
3074 // to the new UNION'd larger loop. I set the _head field of these
3075 // dead loops to NULL and the _parent field points to the owning
3076 // loop. Shades of UNION-FIND algorithm.
3077 IdealLoopTree *ilt;
3078 while( !(ilt = get_loop(n))->_head ) {
3079 // Normally I would use a set_loop here. But in this one special
3080 // case, it is legal (and expected) to change what loop a Node
3081 // belongs to.
3082 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3083 }
3084 // Remove safepoints ONLY if I've already seen I don't need one.
3085 // (the old code here would yank a 2nd safepoint after seeing a
3086 // first one, even though the 1st did not dominate in the loop body
3087 // and thus could be avoided indefinitely)
3088 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3089 is_deleteable_safept(n)) {
3090 Node *in = n->in(TypeFunc::Control);
3091 lazy_replace(n,in); // Pull safepoint now
3092 if (ilt->_safepts != NULL) {
3093 ilt->_safepts->yank(n);
3094 }
3095 // Carry on with the recursion "as if" we are walking
3096 // only the control input
3097 if( !visited.test_set( in->_idx ) ) {
3098 worklist.push(in); // Visit this guy later, using worklist
3099 }
3100 // Get next node from nstack:
3101 // - skip n's inputs processing by setting i > cnt;
3102 // - we also will not call set_early_ctrl(n) since
3103 // has_node(n) == true (see the condition above).
3104 i = cnt + 1;
3105 }
3106 }
3107 } // if (i == 0)
3109 // Visit all inputs
3110 bool done = true; // Assume all n's inputs will be processed
3111 while (i < cnt) {
3112 Node *in = n->in(i);
3113 ++i;
3114 if (in == NULL) continue;
3115 if (in->pinned() && !in->is_CFG())
3116 set_ctrl(in, in->in(0));
3117 int is_visited = visited.test_set( in->_idx );
3118 if (!has_node(in)) { // No controlling input yet?
3119 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3120 assert( !is_visited, "visit only once" );
3121 nstack.push(n, i); // Save parent node and next input's index.
3122 nstack_top_n = in; // Process current input now.
3123 nstack_top_i = 0;
3124 done = false; // Not all n's inputs processed.
3125 break; // continue while_nstack_nonempty;
3126 } else if (!is_visited) {
3127 // This guy has a location picked out for him, but has not yet
3128 // been visited. Happens to all CFG nodes, for instance.
3129 // Visit him using the worklist instead of recursion, to break
3130 // cycles. Since he has a location already we do not need to
3131 // find his location before proceeding with the current Node.
3132 worklist.push(in); // Visit this guy later, using worklist
3133 }
3134 }
3135 if (done) {
3136 // All of n's inputs have been processed, complete post-processing.
3138 // Compute earliest point this Node can go.
3139 // CFG, Phi, pinned nodes already know their controlling input.
3140 if (!has_node(n)) {
3141 // Record earliest legal location
3142 set_early_ctrl( n );
3143 }
3144 if (nstack.is_empty()) {
3145 // Finished all nodes on stack.
3146 // Process next node on the worklist.
3147 break;
3148 }
3149 // Get saved parent node and next input's index.
3150 nstack_top_n = nstack.node();
3151 nstack_top_i = nstack.index();
3152 nstack.pop();
3153 }
3154 } // while (true)
3155 }
3156 }
3158 //------------------------------dom_lca_internal--------------------------------
3159 // Pair-wise LCA
3160 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3161 if( !n1 ) return n2; // Handle NULL original LCA
3162 assert( n1->is_CFG(), "" );
3163 assert( n2->is_CFG(), "" );
3164 // find LCA of all uses
3165 uint d1 = dom_depth(n1);
3166 uint d2 = dom_depth(n2);
3167 while (n1 != n2) {
3168 if (d1 > d2) {
3169 n1 = idom(n1);
3170 d1 = dom_depth(n1);
3171 } else if (d1 < d2) {
3172 n2 = idom(n2);
3173 d2 = dom_depth(n2);
3174 } else {
3175 // Here d1 == d2. Due to edits of the dominator-tree, sections
3176 // of the tree might have the same depth. These sections have
3177 // to be searched more carefully.
3179 // Scan up all the n1's with equal depth, looking for n2.
3180 Node *t1 = idom(n1);
3181 while (dom_depth(t1) == d1) {
3182 if (t1 == n2) return n2;
3183 t1 = idom(t1);
3184 }
3185 // Scan up all the n2's with equal depth, looking for n1.
3186 Node *t2 = idom(n2);
3187 while (dom_depth(t2) == d2) {
3188 if (t2 == n1) return n1;
3189 t2 = idom(t2);
3190 }
3191 // Move up to a new dominator-depth value as well as up the dom-tree.
3192 n1 = t1;
3193 n2 = t2;
3194 d1 = dom_depth(n1);
3195 d2 = dom_depth(n2);
3196 }
3197 }
3198 return n1;
3199 }
3201 //------------------------------compute_idom-----------------------------------
3202 // Locally compute IDOM using dom_lca call. Correct only if the incoming
3203 // IDOMs are correct.
3204 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3205 assert( region->is_Region(), "" );
3206 Node *LCA = NULL;
3207 for( uint i = 1; i < region->req(); i++ ) {
3208 if( region->in(i) != C->top() )
3209 LCA = dom_lca( LCA, region->in(i) );
3210 }
3211 return LCA;
3212 }
3214 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3215 bool had_error = false;
3216 #ifdef ASSERT
3217 if (early != C->root()) {
3218 // Make sure that there's a dominance path from LCA to early
3219 Node* d = LCA;
3220 while (d != early) {
3221 if (d == C->root()) {
3222 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3223 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3224 had_error = true;
3225 break;
3226 }
3227 d = idom(d);
3228 }
3229 }
3230 #endif
3231 return had_error;
3232 }
3235 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3236 // Compute LCA over list of uses
3237 bool had_error = false;
3238 Node *LCA = NULL;
3239 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3240 Node* c = n->fast_out(i);
3241 if (_nodes[c->_idx] == NULL)
3242 continue; // Skip the occasional dead node
3243 if( c->is_Phi() ) { // For Phis, we must land above on the path
3244 for( uint j=1; j<c->req(); j++ ) {// For all inputs
3245 if( c->in(j) == n ) { // Found matching input?
3246 Node *use = c->in(0)->in(j);
3247 if (_verify_only && use->is_top()) continue;
3248 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3249 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3250 }
3251 }
3252 } else {
3253 // For CFG data-users, use is in the block just prior
3254 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3255 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3256 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3257 }
3258 }
3259 assert(!had_error, "bad dominance");
3260 return LCA;
3261 }
3263 //------------------------------get_late_ctrl----------------------------------
3264 // Compute latest legal control.
3265 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3266 assert(early != NULL, "early control should not be NULL");
3268 Node* LCA = compute_lca_of_uses(n, early);
3269 #ifdef ASSERT
3270 if (LCA == C->root() && LCA != early) {
3271 // def doesn't dominate uses so print some useful debugging output
3272 compute_lca_of_uses(n, early, true);
3273 }
3274 #endif
3276 // if this is a load, check for anti-dependent stores
3277 // We use a conservative algorithm to identify potential interfering
3278 // instructions and for rescheduling the load. The users of the memory
3279 // input of this load are examined. Any use which is not a load and is
3280 // dominated by early is considered a potentially interfering store.
3281 // This can produce false positives.
3282 if (n->is_Load() && LCA != early) {
3283 Node_List worklist;
3285 Node *mem = n->in(MemNode::Memory);
3286 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3287 Node* s = mem->fast_out(i);
3288 worklist.push(s);
3289 }
3290 while(worklist.size() != 0 && LCA != early) {
3291 Node* s = worklist.pop();
3292 if (s->is_Load()) {
3293 continue;
3294 } else if (s->is_MergeMem()) {
3295 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3296 Node* s1 = s->fast_out(i);
3297 worklist.push(s1);
3298 }
3299 } else {
3300 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3301 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3302 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3303 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3304 }
3305 }
3306 }
3307 }
3309 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3310 return LCA;
3311 }
3313 // true if CFG node d dominates CFG node n
3314 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3315 if (d == n)
3316 return true;
3317 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3318 uint dd = dom_depth(d);
3319 while (dom_depth(n) >= dd) {
3320 if (n == d)
3321 return true;
3322 n = idom(n);
3323 }
3324 return false;
3325 }
3327 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3328 // Pair-wise LCA with tags.
3329 // Tag each index with the node 'tag' currently being processed
3330 // before advancing up the dominator chain using idom().
3331 // Later calls that find a match to 'tag' know that this path has already
3332 // been considered in the current LCA (which is input 'n1' by convention).
3333 // Since get_late_ctrl() is only called once for each node, the tag array
3334 // does not need to be cleared between calls to get_late_ctrl().
3335 // Algorithm trades a larger constant factor for better asymptotic behavior
3336 //
3337 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3338 uint d1 = dom_depth(n1);
3339 uint d2 = dom_depth(n2);
3341 do {
3342 if (d1 > d2) {
3343 // current lca is deeper than n2
3344 _dom_lca_tags.map(n1->_idx, tag);
3345 n1 = idom(n1);
3346 d1 = dom_depth(n1);
3347 } else if (d1 < d2) {
3348 // n2 is deeper than current lca
3349 Node *memo = _dom_lca_tags[n2->_idx];
3350 if( memo == tag ) {
3351 return n1; // Return the current LCA
3352 }
3353 _dom_lca_tags.map(n2->_idx, tag);
3354 n2 = idom(n2);
3355 d2 = dom_depth(n2);
3356 } else {
3357 // Here d1 == d2. Due to edits of the dominator-tree, sections
3358 // of the tree might have the same depth. These sections have
3359 // to be searched more carefully.
3361 // Scan up all the n1's with equal depth, looking for n2.
3362 _dom_lca_tags.map(n1->_idx, tag);
3363 Node *t1 = idom(n1);
3364 while (dom_depth(t1) == d1) {
3365 if (t1 == n2) return n2;
3366 _dom_lca_tags.map(t1->_idx, tag);
3367 t1 = idom(t1);
3368 }
3369 // Scan up all the n2's with equal depth, looking for n1.
3370 _dom_lca_tags.map(n2->_idx, tag);
3371 Node *t2 = idom(n2);
3372 while (dom_depth(t2) == d2) {
3373 if (t2 == n1) return n1;
3374 _dom_lca_tags.map(t2->_idx, tag);
3375 t2 = idom(t2);
3376 }
3377 // Move up to a new dominator-depth value as well as up the dom-tree.
3378 n1 = t1;
3379 n2 = t2;
3380 d1 = dom_depth(n1);
3381 d2 = dom_depth(n2);
3382 }
3383 } while (n1 != n2);
3384 return n1;
3385 }
3387 //------------------------------init_dom_lca_tags------------------------------
3388 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3389 // Intended use does not involve any growth for the array, so it could
3390 // be of fixed size.
3391 void PhaseIdealLoop::init_dom_lca_tags() {
3392 uint limit = C->unique() + 1;
3393 _dom_lca_tags.map( limit, NULL );
3394 #ifdef ASSERT
3395 for( uint i = 0; i < limit; ++i ) {
3396 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3397 }
3398 #endif // ASSERT
3399 }
3401 //------------------------------clear_dom_lca_tags------------------------------
3402 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3403 // Intended use does not involve any growth for the array, so it could
3404 // be of fixed size.
3405 void PhaseIdealLoop::clear_dom_lca_tags() {
3406 uint limit = C->unique() + 1;
3407 _dom_lca_tags.map( limit, NULL );
3408 _dom_lca_tags.clear();
3409 #ifdef ASSERT
3410 for( uint i = 0; i < limit; ++i ) {
3411 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3412 }
3413 #endif // ASSERT
3414 }
3416 //------------------------------build_loop_late--------------------------------
3417 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3418 // Second pass finds latest legal placement, and ideal loop placement.
3419 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3420 while (worklist.size() != 0) {
3421 Node *n = worklist.pop();
3422 // Only visit once
3423 if (visited.test_set(n->_idx)) continue;
3424 uint cnt = n->outcnt();
3425 uint i = 0;
3426 while (true) {
3427 assert( _nodes[n->_idx], "no dead nodes" );
3428 // Visit all children
3429 if (i < cnt) {
3430 Node* use = n->raw_out(i);
3431 ++i;
3432 // Check for dead uses. Aggressively prune such junk. It might be
3433 // dead in the global sense, but still have local uses so I cannot
3434 // easily call 'remove_dead_node'.
3435 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3436 // Due to cycles, we might not hit the same fixed point in the verify
3437 // pass as we do in the regular pass. Instead, visit such phis as
3438 // simple uses of the loop head.
3439 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3440 if( !visited.test(use->_idx) )
3441 worklist.push(use);
3442 } else if( !visited.test_set(use->_idx) ) {
3443 nstack.push(n, i); // Save parent and next use's index.
3444 n = use; // Process all children of current use.
3445 cnt = use->outcnt();
3446 i = 0;
3447 }
3448 } else {
3449 // Do not visit around the backedge of loops via data edges.
3450 // push dead code onto a worklist
3451 _deadlist.push(use);
3452 }
3453 } else {
3454 // All of n's children have been processed, complete post-processing.
3455 build_loop_late_post(n);
3456 if (nstack.is_empty()) {
3457 // Finished all nodes on stack.
3458 // Process next node on the worklist.
3459 break;
3460 }
3461 // Get saved parent node and next use's index. Visit the rest of uses.
3462 n = nstack.node();
3463 cnt = n->outcnt();
3464 i = nstack.index();
3465 nstack.pop();
3466 }
3467 }
3468 }
3469 }
3471 //------------------------------build_loop_late_post---------------------------
3472 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3473 // Second pass finds latest legal placement, and ideal loop placement.
3474 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3476 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
3477 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3478 }
3480 #ifdef ASSERT
3481 if (_verify_only && !n->is_CFG()) {
3482 // Check def-use domination.
3483 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
3484 }
3485 #endif
3487 // CFG and pinned nodes already handled
3488 if( n->in(0) ) {
3489 if( n->in(0)->is_top() ) return; // Dead?
3491 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3492 // _must_ be pinned (they have to observe their control edge of course).
3493 // Unlike Stores (which modify an unallocable resource, the memory
3494 // state), Mods/Loads can float around. So free them up.
3495 bool pinned = true;
3496 switch( n->Opcode() ) {
3497 case Op_DivI:
3498 case Op_DivF:
3499 case Op_DivD:
3500 case Op_ModI:
3501 case Op_ModF:
3502 case Op_ModD:
3503 case Op_LoadB: // Same with Loads; they can sink
3504 case Op_LoadUB: // during loop optimizations.
3505 case Op_LoadUS:
3506 case Op_LoadD:
3507 case Op_LoadF:
3508 case Op_LoadI:
3509 case Op_LoadKlass:
3510 case Op_LoadNKlass:
3511 case Op_LoadL:
3512 case Op_LoadS:
3513 case Op_LoadP:
3514 case Op_LoadN:
3515 case Op_LoadRange:
3516 case Op_LoadD_unaligned:
3517 case Op_LoadL_unaligned:
3518 case Op_StrComp: // Does a bunch of load-like effects
3519 case Op_StrEquals:
3520 case Op_StrIndexOf:
3521 case Op_AryEq:
3522 pinned = false;
3523 }
3524 if( pinned ) {
3525 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3526 if( !chosen_loop->_child ) // Inner loop?
3527 chosen_loop->_body.push(n); // Collect inner loops
3528 return;
3529 }
3530 } else { // No slot zero
3531 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3532 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3533 return;
3534 }
3535 assert(!n->is_CFG() || n->outcnt() == 0, "");
3536 }
3538 // Do I have a "safe range" I can select over?
3539 Node *early = get_ctrl(n);// Early location already computed
3541 // Compute latest point this Node can go
3542 Node *LCA = get_late_ctrl( n, early );
3543 // LCA is NULL due to uses being dead
3544 if( LCA == NULL ) {
3545 #ifdef ASSERT
3546 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3547 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3548 }
3549 #endif
3550 _nodes.map(n->_idx, 0); // This node is useless
3551 _deadlist.push(n);
3552 return;
3553 }
3554 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3556 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3557 Node *least = legal; // Best legal position so far
3558 while( early != legal ) { // While not at earliest legal
3559 #ifdef ASSERT
3560 if (legal->is_Start() && !early->is_Root()) {
3561 // Bad graph. Print idom path and fail.
3562 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
3563 assert(false, "Bad graph detected in build_loop_late");
3564 }
3565 #endif
3566 // Find least loop nesting depth
3567 legal = idom(legal); // Bump up the IDOM tree
3568 // Check for lower nesting depth
3569 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3570 least = legal;
3571 }
3572 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3574 // Try not to place code on a loop entry projection
3575 // which can inhibit range check elimination.
3576 if (least != early) {
3577 Node* ctrl_out = least->unique_ctrl_out();
3578 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3579 least == ctrl_out->in(LoopNode::EntryControl)) {
3580 Node* least_dom = idom(least);
3581 if (get_loop(least_dom)->is_member(get_loop(least))) {
3582 least = least_dom;
3583 }
3584 }
3585 }
3587 #ifdef ASSERT
3588 // If verifying, verify that 'verify_me' has a legal location
3589 // and choose it as our location.
3590 if( _verify_me ) {
3591 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3592 Node *legal = LCA;
3593 while( early != legal ) { // While not at earliest legal
3594 if( legal == v_ctrl ) break; // Check for prior good location
3595 legal = idom(legal) ;// Bump up the IDOM tree
3596 }
3597 // Check for prior good location
3598 if( legal == v_ctrl ) least = legal; // Keep prior if found
3599 }
3600 #endif
3602 // Assign discovered "here or above" point
3603 least = find_non_split_ctrl(least);
3604 set_ctrl(n, least);
3606 // Collect inner loop bodies
3607 IdealLoopTree *chosen_loop = get_loop(least);
3608 if( !chosen_loop->_child ) // Inner loop?
3609 chosen_loop->_body.push(n);// Collect inner loops
3610 }
3612 #ifdef ASSERT
3613 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
3614 tty->print_cr("%s", msg);
3615 tty->print("n: "); n->dump();
3616 tty->print("early(n): "); early->dump();
3617 if (n->in(0) != NULL && !n->in(0)->is_top() &&
3618 n->in(0) != early && !n->in(0)->is_Root()) {
3619 tty->print("n->in(0): "); n->in(0)->dump();
3620 }
3621 for (uint i = 1; i < n->req(); i++) {
3622 Node* in1 = n->in(i);
3623 if (in1 != NULL && in1 != n && !in1->is_top()) {
3624 tty->print("n->in(%d): ", i); in1->dump();
3625 Node* in1_early = get_ctrl(in1);
3626 tty->print("early(n->in(%d)): ", i); in1_early->dump();
3627 if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
3628 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
3629 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
3630 }
3631 for (uint j = 1; j < in1->req(); j++) {
3632 Node* in2 = in1->in(j);
3633 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
3634 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
3635 Node* in2_early = get_ctrl(in2);
3636 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
3637 if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
3638 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
3639 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
3640 }
3641 }
3642 }
3643 }
3644 }
3645 tty->cr();
3646 tty->print("LCA(n): "); LCA->dump();
3647 for (uint i = 0; i < n->outcnt(); i++) {
3648 Node* u1 = n->raw_out(i);
3649 if (u1 == n)
3650 continue;
3651 tty->print("n->out(%d): ", i); u1->dump();
3652 if (u1->is_CFG()) {
3653 for (uint j = 0; j < u1->outcnt(); j++) {
3654 Node* u2 = u1->raw_out(j);
3655 if (u2 != u1 && u2 != n && u2->is_CFG()) {
3656 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3657 }
3658 }
3659 } else {
3660 Node* u1_later = get_ctrl(u1);
3661 tty->print("later(n->out(%d)): ", i); u1_later->dump();
3662 if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
3663 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
3664 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
3665 }
3666 for (uint j = 0; j < u1->outcnt(); j++) {
3667 Node* u2 = u1->raw_out(j);
3668 if (u2 == n || u2 == u1)
3669 continue;
3670 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3671 if (!u2->is_CFG()) {
3672 Node* u2_later = get_ctrl(u2);
3673 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
3674 if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
3675 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
3676 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
3677 }
3678 }
3679 }
3680 }
3681 }
3682 tty->cr();
3683 int ct = 0;
3684 Node *dbg_legal = LCA;
3685 while(!dbg_legal->is_Start() && ct < 100) {
3686 tty->print("idom[%d] ",ct); dbg_legal->dump();
3687 ct++;
3688 dbg_legal = idom(dbg_legal);
3689 }
3690 tty->cr();
3691 }
3692 #endif
3694 #ifndef PRODUCT
3695 //------------------------------dump-------------------------------------------
3696 void PhaseIdealLoop::dump( ) const {
3697 ResourceMark rm;
3698 Arena* arena = Thread::current()->resource_area();
3699 Node_Stack stack(arena, C->live_nodes() >> 2);
3700 Node_List rpo_list;
3701 VectorSet visited(arena);
3702 visited.set(C->top()->_idx);
3703 rpo( C->root(), stack, visited, rpo_list );
3704 // Dump root loop indexed by last element in PO order
3705 dump( _ltree_root, rpo_list.size(), rpo_list );
3706 }
3708 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3709 loop->dump_head();
3711 // Now scan for CFG nodes in the same loop
3712 for( uint j=idx; j > 0; j-- ) {
3713 Node *n = rpo_list[j-1];
3714 if( !_nodes[n->_idx] ) // Skip dead nodes
3715 continue;
3716 if( get_loop(n) != loop ) { // Wrong loop nest
3717 if( get_loop(n)->_head == n && // Found nested loop?
3718 get_loop(n)->_parent == loop )
3719 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3720 continue;
3721 }
3723 // Dump controlling node
3724 for( uint x = 0; x < loop->_nest; x++ )
3725 tty->print(" ");
3726 tty->print("C");
3727 if( n == C->root() ) {
3728 n->dump();
3729 } else {
3730 Node* cached_idom = idom_no_update(n);
3731 Node *computed_idom = n->in(0);
3732 if( n->is_Region() ) {
3733 computed_idom = compute_idom(n);
3734 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3735 // any MultiBranch ctrl node), so apply a similar transform to
3736 // the cached idom returned from idom_no_update.
3737 cached_idom = find_non_split_ctrl(cached_idom);
3738 }
3739 tty->print(" ID:%d",computed_idom->_idx);
3740 n->dump();
3741 if( cached_idom != computed_idom ) {
3742 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3743 computed_idom->_idx, cached_idom->_idx);
3744 }
3745 }
3746 // Dump nodes it controls
3747 for( uint k = 0; k < _nodes.Size(); k++ ) {
3748 // (k < C->unique() && get_ctrl(find(k)) == n)
3749 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3750 Node *m = C->root()->find(k);
3751 if( m && m->outcnt() > 0 ) {
3752 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3753 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3754 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3755 }
3756 for( uint j = 0; j < loop->_nest; j++ )
3757 tty->print(" ");
3758 tty->print(" ");
3759 m->dump();
3760 }
3761 }
3762 }
3763 }
3764 }
3766 // Collect a R-P-O for the whole CFG.
3767 // Result list is in post-order (scan backwards for RPO)
3768 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3769 stk.push(start, 0);
3770 visited.set(start->_idx);
3772 while (stk.is_nonempty()) {
3773 Node* m = stk.node();
3774 uint idx = stk.index();
3775 if (idx < m->outcnt()) {
3776 stk.set_index(idx + 1);
3777 Node* n = m->raw_out(idx);
3778 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3779 stk.push(n, 0);
3780 }
3781 } else {
3782 rpo_list.push(m);
3783 stk.pop();
3784 }
3785 }
3786 }
3787 #endif
3790 //=============================================================================
3791 //------------------------------LoopTreeIterator-----------------------------------
3793 // Advance to next loop tree using a preorder, left-to-right traversal.
3794 void LoopTreeIterator::next() {
3795 assert(!done(), "must not be done.");
3796 if (_curnt->_child != NULL) {
3797 _curnt = _curnt->_child;
3798 } else if (_curnt->_next != NULL) {
3799 _curnt = _curnt->_next;
3800 } else {
3801 while (_curnt != _root && _curnt->_next == NULL) {
3802 _curnt = _curnt->_parent;
3803 }
3804 if (_curnt == _root) {
3805 _curnt = NULL;
3806 assert(done(), "must be done.");
3807 } else {
3808 assert(_curnt->_next != NULL, "must be more to do");
3809 _curnt = _curnt->_next;
3810 }
3811 }
3812 }