Wed, 04 May 2011 13:12:42 -0700
5091921: Sign flip issues in loop optimizer
Summary: Fix integer overflow problem in the code generated by loop optimizer.
Reviewed-by: never
1 /*
2 * Copyright (c) 1998, 2011, 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) ) {
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 return early;
136 }
138 //------------------------------set_early_ctrl---------------------------------
139 // Set earliest legal control
140 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
141 Node *early = get_early_ctrl(n);
143 // Record earliest legal location
144 set_ctrl(n, early);
145 }
147 //------------------------------set_subtree_ctrl-------------------------------
148 // set missing _ctrl entries on new nodes
149 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
150 // Already set? Get out.
151 if( _nodes[n->_idx] ) return;
152 // Recursively set _nodes array to indicate where the Node goes
153 uint i;
154 for( i = 0; i < n->req(); ++i ) {
155 Node *m = n->in(i);
156 if( m && m != C->root() )
157 set_subtree_ctrl( m );
158 }
160 // Fixup self
161 set_early_ctrl( n );
162 }
164 //------------------------------is_counted_loop--------------------------------
165 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
166 PhaseGVN *gvn = &_igvn;
168 // Counted loop head must be a good RegionNode with only 3 not NULL
169 // control input edges: Self, Entry, LoopBack.
170 if (x->in(LoopNode::Self) == NULL || x->req() != 3)
171 return false;
173 Node *init_control = x->in(LoopNode::EntryControl);
174 Node *back_control = x->in(LoopNode::LoopBackControl);
175 if (init_control == NULL || back_control == NULL) // Partially dead
176 return false;
177 // Must also check for TOP when looking for a dead loop
178 if (init_control->is_top() || back_control->is_top())
179 return false;
181 // Allow funny placement of Safepoint
182 if (back_control->Opcode() == Op_SafePoint)
183 back_control = back_control->in(TypeFunc::Control);
185 // Controlling test for loop
186 Node *iftrue = back_control;
187 uint iftrue_op = iftrue->Opcode();
188 if (iftrue_op != Op_IfTrue &&
189 iftrue_op != Op_IfFalse)
190 // I have a weird back-control. Probably the loop-exit test is in
191 // the middle of the loop and I am looking at some trailing control-flow
192 // merge point. To fix this I would have to partially peel the loop.
193 return false; // Obscure back-control
195 // Get boolean guarding loop-back test
196 Node *iff = iftrue->in(0);
197 if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
198 return false;
199 BoolNode *test = iff->in(1)->as_Bool();
200 BoolTest::mask bt = test->_test._test;
201 float cl_prob = iff->as_If()->_prob;
202 if (iftrue_op == Op_IfFalse) {
203 bt = BoolTest(bt).negate();
204 cl_prob = 1.0 - cl_prob;
205 }
206 // Get backedge compare
207 Node *cmp = test->in(1);
208 int cmp_op = cmp->Opcode();
209 if (cmp_op != Op_CmpI)
210 return false; // Avoid pointer & float compares
212 // Find the trip-counter increment & limit. Limit must be loop invariant.
213 Node *incr = cmp->in(1);
214 Node *limit = cmp->in(2);
216 // ---------
217 // need 'loop()' test to tell if limit is loop invariant
218 // ---------
220 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
221 Node *tmp = incr; // Then reverse order into the CmpI
222 incr = limit;
223 limit = tmp;
224 bt = BoolTest(bt).commute(); // And commute the exit test
225 }
226 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
227 return false;
228 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
229 return false;
231 Node* phi_incr = NULL;
232 // Trip-counter increment must be commutative & associative.
233 if (incr->is_Phi()) {
234 if (incr->as_Phi()->region() != x || incr->req() != 3)
235 return false; // Not simple trip counter expression
236 phi_incr = incr;
237 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
238 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
239 return false;
240 }
242 Node* trunc1 = NULL;
243 Node* trunc2 = NULL;
244 const TypeInt* iv_trunc_t = NULL;
245 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
246 return false; // Funny increment opcode
247 }
248 assert(incr->Opcode() == Op_AddI, "wrong increment code");
250 // Get merge point
251 Node *xphi = incr->in(1);
252 Node *stride = incr->in(2);
253 if (!stride->is_Con()) { // Oops, swap these
254 if (!xphi->is_Con()) // Is the other guy a constant?
255 return false; // Nope, unknown stride, bail out
256 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
257 xphi = stride;
258 stride = tmp;
259 }
260 // Stride must be constant
261 int stride_con = stride->get_int();
262 if (stride_con == 0)
263 return false; // missed some peephole opt
265 if (!xphi->is_Phi())
266 return false; // Too much math on the trip counter
267 if (phi_incr != NULL && phi_incr != xphi)
268 return false;
269 PhiNode *phi = xphi->as_Phi();
271 // Phi must be of loop header; backedge must wrap to increment
272 if (phi->region() != x)
273 return false;
274 if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
275 trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
276 return false;
277 }
278 Node *init_trip = phi->in(LoopNode::EntryControl);
280 // If iv trunc type is smaller than int, check for possible wrap.
281 if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
282 assert(trunc1 != NULL, "must have found some truncation");
284 // Get a better type for the phi (filtered thru if's)
285 const TypeInt* phi_ft = filtered_type(phi);
287 // Can iv take on a value that will wrap?
288 //
289 // Ensure iv's limit is not within "stride" of the wrap value.
290 //
291 // Example for "short" type
292 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
293 // If the stride is +10, then the last value of the induction
294 // variable before the increment (phi_ft->_hi) must be
295 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
296 // ensure no truncation occurs after the increment.
298 if (stride_con > 0) {
299 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
300 iv_trunc_t->_lo > phi_ft->_lo) {
301 return false; // truncation may occur
302 }
303 } else if (stride_con < 0) {
304 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
305 iv_trunc_t->_hi < phi_ft->_hi) {
306 return false; // truncation may occur
307 }
308 }
309 // No possibility of wrap so truncation can be discarded
310 // Promote iv type to Int
311 } else {
312 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
313 }
315 // If the condition is inverted and we will be rolling
316 // through MININT to MAXINT, then bail out.
317 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
318 // Odd stride
319 bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
320 // Count down loop rolls through MAXINT
321 (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
322 // Count up loop rolls through MININT
323 (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
324 return false; // Bail out
325 }
327 const TypeInt* init_t = gvn->type(init_trip)->is_int();
328 const TypeInt* limit_t = gvn->type(limit)->is_int();
330 if (stride_con > 0) {
331 long init_p = (long)init_t->_lo + stride_con;
332 if (init_p > (long)max_jint || init_p > (long)limit_t->_hi)
333 return false; // cyclic loop or this loop trips only once
334 } else {
335 long init_p = (long)init_t->_hi + stride_con;
336 if (init_p < (long)min_jint || init_p < (long)limit_t->_lo)
337 return false; // cyclic loop or this loop trips only once
338 }
340 // =================================================
341 // ---- SUCCESS! Found A Trip-Counted Loop! -----
342 //
343 assert(x->Opcode() == Op_Loop, "regular loops only");
344 C->print_method("Before CountedLoop", 3);
346 Node *hook = new (C, 6) Node(6);
348 if (LoopLimitCheck) {
350 // ===================================================
351 // Generate loop limit check to avoid integer overflow
352 // in cases like next (cyclic loops):
353 //
354 // for (i=0; i <= max_jint; i++) {}
355 // for (i=0; i < max_jint; i+=2) {}
356 //
357 //
358 // Limit check predicate depends on the loop test:
359 //
360 // for(;i != limit; i++) --> limit <= (max_jint)
361 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1)
362 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride )
363 //
365 // Check if limit is excluded to do more precise int overflow check.
366 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
367 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
369 // If compare points directly to the phi we need to adjust
370 // the compare so that it points to the incr. Limit have
371 // to be adjusted to keep trip count the same and the
372 // adjusted limit should be checked for int overflow.
373 if (phi_incr != NULL) {
374 stride_m += stride_con;
375 }
377 if (limit->is_Con()) {
378 int limit_con = limit->get_int();
379 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
380 (stride_con < 0 && limit_con < (min_jint - stride_m))) {
381 // Bailout: it could be integer overflow.
382 return false;
383 }
384 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
385 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
386 // Limit's type may satisfy the condition, for example,
387 // when it is an array length.
388 } else {
389 // Generate loop's limit check.
390 // Loop limit check predicate should be near the loop.
391 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
392 if (!limit_check_proj) {
393 // The limit check predicate is not generated if this method trapped here before.
394 #ifdef ASSERT
395 if (TraceLoopLimitCheck) {
396 tty->print("missing loop limit check:");
397 loop->dump_head();
398 x->dump(1);
399 }
400 #endif
401 return false;
402 }
404 IfNode* check_iff = limit_check_proj->in(0)->as_If();
405 Node* cmp_limit;
406 Node* bol;
408 if (stride_con > 0) {
409 cmp_limit = new (C, 3) CmpINode(limit, _igvn.intcon(max_jint - stride_m));
410 bol = new (C, 2) BoolNode(cmp_limit, BoolTest::le);
411 } else {
412 cmp_limit = new (C, 3) CmpINode(limit, _igvn.intcon(min_jint - stride_m));
413 bol = new (C, 2) BoolNode(cmp_limit, BoolTest::ge);
414 }
415 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
416 bol = _igvn.register_new_node_with_optimizer(bol);
417 set_subtree_ctrl(bol);
419 // Replace condition in original predicate but preserve Opaque node
420 // so that previous predicates could be found.
421 assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
422 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
423 Node* opq = check_iff->in(1)->in(1);
424 _igvn.hash_delete(opq);
425 opq->set_req(1, bol);
426 // Update ctrl.
427 set_ctrl(opq, check_iff->in(0));
428 set_ctrl(check_iff->in(1), check_iff->in(0));
430 #ifndef PRODUCT
431 // report that the loop predication has been actually performed
432 // for this loop
433 if (TraceLoopLimitCheck) {
434 tty->print_cr("Counted Loop Limit Check generated:");
435 debug_only( bol->dump(2); )
436 }
437 #endif
438 }
440 if (phi_incr != NULL) {
441 // If compare points directly to the phi we need to adjust
442 // the compare so that it points to the incr. Limit have
443 // to be adjusted to keep trip count the same and we
444 // should avoid int overflow.
445 //
446 // i = init; do {} while(i++ < limit);
447 // is converted to
448 // i = init; do {} while(++i < limit+1);
449 //
450 limit = gvn->transform(new (C, 3) AddINode(limit, stride));
451 }
453 // Now we need to canonicalize loop condition.
454 if (bt == BoolTest::ne) {
455 assert(stride_con == 1 || stride_con == -1, "simple increment only");
456 bt = (stride_con > 0) ? BoolTest::lt : BoolTest::gt;
457 }
459 if (incl_limit) {
460 // The limit check guaranties that 'limit <= (max_jint - stride)' so
461 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
462 //
463 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
464 limit = gvn->transform(new (C, 3) AddINode(limit, one));
465 if (bt == BoolTest::le)
466 bt = BoolTest::lt;
467 else if (bt == BoolTest::ge)
468 bt = BoolTest::gt;
469 else
470 ShouldNotReachHere();
471 }
472 set_subtree_ctrl( limit );
474 } else { // LoopLimitCheck
476 // If compare points to incr, we are ok. Otherwise the compare
477 // can directly point to the phi; in this case adjust the compare so that
478 // it points to the incr by adjusting the limit.
479 if (cmp->in(1) == phi || cmp->in(2) == phi)
480 limit = gvn->transform(new (C, 3) AddINode(limit,stride));
482 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
483 // Final value for iterator should be: trip_count * stride + init_trip.
484 Node *one_p = gvn->intcon( 1);
485 Node *one_m = gvn->intcon(-1);
487 Node *trip_count = NULL;
488 switch( bt ) {
489 case BoolTest::eq:
490 ShouldNotReachHere();
491 case BoolTest::ne: // Ahh, the case we desire
492 if (stride_con == 1)
493 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip));
494 else if (stride_con == -1)
495 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit));
496 else
497 ShouldNotReachHere();
498 set_subtree_ctrl(trip_count);
499 //_loop.map(trip_count->_idx,loop(limit));
500 break;
501 case BoolTest::le: // Maybe convert to '<' case
502 limit = gvn->transform(new (C, 3) AddINode(limit,one_p));
503 set_subtree_ctrl( limit );
504 hook->init_req(4, limit);
506 bt = BoolTest::lt;
507 // Make the new limit be in the same loop nest as the old limit
508 //_loop.map(limit->_idx,limit_loop);
509 // Fall into next case
510 case BoolTest::lt: { // Maybe convert to '!=' case
511 if (stride_con < 0) // Count down loop rolls through MAXINT
512 ShouldNotReachHere();
513 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
514 set_subtree_ctrl( range );
515 hook->init_req(0, range);
517 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
518 set_subtree_ctrl( bias );
519 hook->init_req(1, bias);
521 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m));
522 set_subtree_ctrl( bias1 );
523 hook->init_req(2, bias1);
525 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
526 set_subtree_ctrl( trip_count );
527 hook->init_req(3, trip_count);
528 break;
529 }
531 case BoolTest::ge: // Maybe convert to '>' case
532 limit = gvn->transform(new (C, 3) AddINode(limit,one_m));
533 set_subtree_ctrl( limit );
534 hook->init_req(4 ,limit);
536 bt = BoolTest::gt;
537 // Make the new limit be in the same loop nest as the old limit
538 //_loop.map(limit->_idx,limit_loop);
539 // Fall into next case
540 case BoolTest::gt: { // Maybe convert to '!=' case
541 if (stride_con > 0) // count up loop rolls through MININT
542 ShouldNotReachHere();
543 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
544 set_subtree_ctrl( range );
545 hook->init_req(0, range);
547 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
548 set_subtree_ctrl( bias );
549 hook->init_req(1, bias);
551 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p));
552 set_subtree_ctrl( bias1 );
553 hook->init_req(2, bias1);
555 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
556 set_subtree_ctrl( trip_count );
557 hook->init_req(3, trip_count);
558 break;
559 }
560 } // switch( bt )
562 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride));
563 set_subtree_ctrl( span );
564 hook->init_req(5, span);
566 limit = gvn->transform(new (C, 3) AddINode(span,init_trip));
567 set_subtree_ctrl( limit );
569 } // LoopLimitCheck
571 // Check for SafePoint on backedge and remove
572 Node *sfpt = x->in(LoopNode::LoopBackControl);
573 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
574 lazy_replace( sfpt, iftrue );
575 loop->_tail = iftrue;
576 }
578 // Build a canonical trip test.
579 // Clone code, as old values may be in use.
580 Node* nphi = PhiNode::make(x, init_trip, TypeInt::INT);
581 nphi = _igvn.register_new_node_with_optimizer(nphi);
582 set_ctrl(nphi, get_ctrl(phi));
584 incr = incr->clone();
585 incr->set_req(1,nphi);
586 incr->set_req(2,stride);
587 incr = _igvn.register_new_node_with_optimizer(incr);
588 set_early_ctrl( incr );
590 nphi->set_req(LoopNode::LoopBackControl, incr);
591 _igvn.replace_node(phi, nphi);
592 phi = nphi->as_Phi();
594 cmp = cmp->clone();
595 cmp->set_req(1,incr);
596 cmp->set_req(2,limit);
597 cmp = _igvn.register_new_node_with_optimizer(cmp);
598 set_ctrl(cmp, iff->in(0));
600 test = test->clone()->as_Bool();
601 (*(BoolTest*)&test->_test)._test = bt;
602 test->set_req(1,cmp);
603 _igvn.register_new_node_with_optimizer(test);
604 set_ctrl(test, iff->in(0));
606 // Replace the old IfNode with a new LoopEndNode
607 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
608 IfNode *le = lex->as_If();
609 uint dd = dom_depth(iff);
610 set_idom(le, le->in(0), dd); // Update dominance for loop exit
611 set_loop(le, loop);
613 // Get the loop-exit control
614 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
616 // Need to swap loop-exit and loop-back control?
617 if (iftrue_op == Op_IfFalse) {
618 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le));
619 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le));
621 loop->_tail = back_control = ift2;
622 set_loop(ift2, loop);
623 set_loop(iff2, get_loop(iffalse));
625 // Lazy update of 'get_ctrl' mechanism.
626 lazy_replace_proj( iffalse, iff2 );
627 lazy_replace_proj( iftrue, ift2 );
629 // Swap names
630 iffalse = iff2;
631 iftrue = ift2;
632 } else {
633 _igvn.hash_delete(iffalse);
634 _igvn.hash_delete(iftrue);
635 iffalse->set_req_X( 0, le, &_igvn );
636 iftrue ->set_req_X( 0, le, &_igvn );
637 }
639 set_idom(iftrue, le, dd+1);
640 set_idom(iffalse, le, dd+1);
641 assert(iff->outcnt() == 0, "should be dead now");
642 lazy_replace( iff, le ); // fix 'get_ctrl'
644 // Now setup a new CountedLoopNode to replace the existing LoopNode
645 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control);
646 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
647 // The following assert is approximately true, and defines the intention
648 // of can_be_counted_loop. It fails, however, because phase->type
649 // is not yet initialized for this loop and its parts.
650 //assert(l->can_be_counted_loop(this), "sanity");
651 _igvn.register_new_node_with_optimizer(l);
652 set_loop(l, loop);
653 loop->_head = l;
654 // Fix all data nodes placed at the old loop head.
655 // Uses the lazy-update mechanism of 'get_ctrl'.
656 lazy_replace( x, l );
657 set_idom(l, init_control, dom_depth(x));
659 // Check for immediately preceding SafePoint and remove
660 Node *sfpt2 = le->in(0);
661 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2))
662 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
664 // Free up intermediate goo
665 _igvn.remove_dead_node(hook);
667 #ifdef ASSERT
668 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
669 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
670 #endif
671 #ifndef PRODUCT
672 if (TraceLoopOpts) {
673 tty->print("Counted ");
674 loop->dump_head();
675 }
676 #endif
678 C->print_method("After CountedLoop", 3);
680 return true;
681 }
683 //----------------------exact_limit-------------------------------------------
684 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
685 assert(loop->_head->is_CountedLoop(), "");
686 CountedLoopNode *cl = loop->_head->as_CountedLoop();
688 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
689 cl->limit()->Opcode() == Op_LoopLimit) {
690 // Old code has exact limit (it could be incorrect in case of int overflow).
691 // Loop limit is exact with stride == 1. And loop may already have exact limit.
692 return cl->limit();
693 }
694 Node *limit = NULL;
695 #ifdef ASSERT
696 BoolTest::mask bt = cl->loopexit()->test_trip();
697 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
698 #endif
699 if (cl->has_exact_trip_count()) {
700 // Simple case: loop has constant boundaries.
701 // Use longs to avoid integer overflow.
702 int stride_con = cl->stride_con();
703 long init_con = cl->init_trip()->get_int();
704 long limit_con = cl->limit()->get_int();
705 julong trip_cnt = cl->trip_count();
706 long final_con = init_con + trip_cnt*stride_con;
707 final_con -= stride_con;
708 int final_int = (int)final_con;
709 // The final value should be in integer range since the loop
710 // is counted and the limit was checked for overflow.
711 assert(final_con == (long)final_int, "final value should be integer");
712 limit = _igvn.intcon(final_int);
713 } else {
714 // Create new LoopLimit node to get exact limit (final iv value).
715 limit = new (C, 4) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
716 register_new_node(limit, cl->in(LoopNode::EntryControl));
717 }
718 assert(limit != NULL, "sanity");
719 return limit;
720 }
722 //------------------------------Ideal------------------------------------------
723 // Return a node which is more "ideal" than the current node.
724 // Attempt to convert into a counted-loop.
725 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
726 if (!can_be_counted_loop(phase)) {
727 phase->C->set_major_progress();
728 }
729 return RegionNode::Ideal(phase, can_reshape);
730 }
733 //=============================================================================
734 //------------------------------Ideal------------------------------------------
735 // Return a node which is more "ideal" than the current node.
736 // Attempt to convert into a counted-loop.
737 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
738 return RegionNode::Ideal(phase, can_reshape);
739 }
741 //------------------------------dump_spec--------------------------------------
742 // Dump special per-node info
743 #ifndef PRODUCT
744 void CountedLoopNode::dump_spec(outputStream *st) const {
745 LoopNode::dump_spec(st);
746 if (stride_is_con()) {
747 st->print("stride: %d ",stride_con());
748 }
749 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
750 if (is_main_loop()) st->print("main of N%d", _idx);
751 if (is_post_loop()) st->print("post of N%d", _main_idx);
752 }
753 #endif
755 //=============================================================================
756 int CountedLoopEndNode::stride_con() const {
757 return stride()->bottom_type()->is_int()->get_con();
758 }
760 //=============================================================================
761 //------------------------------Value-----------------------------------------
762 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
763 const Type* init_t = phase->type(in(Init));
764 const Type* limit_t = phase->type(in(Limit));
765 const Type* stride_t = phase->type(in(Stride));
766 // Either input is TOP ==> the result is TOP
767 if (init_t == Type::TOP) return Type::TOP;
768 if (limit_t == Type::TOP) return Type::TOP;
769 if (stride_t == Type::TOP) return Type::TOP;
771 int stride_con = stride_t->is_int()->get_con();
772 if (stride_con == 1)
773 return NULL; // Identity
775 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
776 // Use longs to avoid integer overflow.
777 long init_con = init_t->is_int()->get_con();
778 long limit_con = limit_t->is_int()->get_con();
779 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
780 long trip_count = (limit_con - init_con + stride_m)/stride_con;
781 long final_con = init_con + stride_con*trip_count;
782 int final_int = (int)final_con;
783 // The final value should be in integer range since the loop
784 // is counted and the limit was checked for overflow.
785 assert(final_con == (long)final_int, "final value should be integer");
786 return TypeInt::make(final_int);
787 }
789 return bottom_type(); // TypeInt::INT
790 }
792 //------------------------------Ideal------------------------------------------
793 // Return a node which is more "ideal" than the current node.
794 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
795 if (phase->type(in(Init)) == Type::TOP ||
796 phase->type(in(Limit)) == Type::TOP ||
797 phase->type(in(Stride)) == Type::TOP)
798 return NULL; // Dead
800 int stride_con = phase->type(in(Stride))->is_int()->get_con();
801 if (stride_con == 1)
802 return NULL; // Identity
804 if (in(Init)->is_Con() && in(Limit)->is_Con())
805 return NULL; // Value
807 // Delay following optimizations until all loop optimizations
808 // done to keep Ideal graph simple.
809 if (!can_reshape || phase->C->major_progress())
810 return NULL;
812 const TypeInt* init_t = phase->type(in(Init) )->is_int();
813 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
814 int stride_p;
815 long lim, ini;
816 julong max;
817 if (stride_con > 0) {
818 stride_p = stride_con;
819 lim = limit_t->_hi;
820 ini = init_t->_lo;
821 max = (julong)max_jint;
822 } else {
823 stride_p = -stride_con;
824 lim = init_t->_hi;
825 ini = limit_t->_lo;
826 max = (julong)min_jint;
827 }
828 julong range = lim - ini + stride_p;
829 if (range <= max) {
830 // Convert to integer expression if it is not overflow.
831 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
832 Node *range = phase->transform(new (phase->C, 3) SubINode(in(Limit), in(Init)));
833 Node *bias = phase->transform(new (phase->C, 3) AddINode(range, stride_m));
834 Node *trip = phase->transform(new (phase->C, 3) DivINode(0, bias, in(Stride)));
835 Node *span = phase->transform(new (phase->C, 3) MulINode(trip, in(Stride)));
836 return new (phase->C, 3) AddINode(span, in(Init)); // exact limit
837 }
839 if (is_power_of_2(stride_p) || // divisor is 2^n
840 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
841 // Convert to long expression to avoid integer overflow
842 // and let igvn optimizer convert this division.
843 //
844 Node* init = phase->transform( new (phase->C, 2) ConvI2LNode(in(Init)));
845 Node* limit = phase->transform( new (phase->C, 2) ConvI2LNode(in(Limit)));
846 Node* stride = phase->longcon(stride_con);
847 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
849 Node *range = phase->transform(new (phase->C, 3) SubLNode(limit, init));
850 Node *bias = phase->transform(new (phase->C, 3) AddLNode(range, stride_m));
851 Node *span;
852 if (stride_con > 0 && is_power_of_2(stride_p)) {
853 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
854 // and avoid generating rounding for division. Zero trip guard should
855 // guarantee that init < limit but sometimes the guard is missing and
856 // we can get situation when init > limit. Note, for the empty loop
857 // optimization zero trip guard is generated explicitly which leaves
858 // only RCE predicate where exact limit is used and the predicate
859 // will simply fail forcing recompilation.
860 Node* neg_stride = phase->longcon(-stride_con);
861 span = phase->transform(new (phase->C, 3) AndLNode(bias, neg_stride));
862 } else {
863 Node *trip = phase->transform(new (phase->C, 3) DivLNode(0, bias, stride));
864 span = phase->transform(new (phase->C, 3) MulLNode(trip, stride));
865 }
866 // Convert back to int
867 Node *span_int = phase->transform(new (phase->C, 2) ConvL2INode(span));
868 return new (phase->C, 3) AddINode(span_int, in(Init)); // exact limit
869 }
871 return NULL; // No progress
872 }
874 //------------------------------Identity---------------------------------------
875 // If stride == 1 return limit node.
876 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
877 int stride_con = phase->type(in(Stride))->is_int()->get_con();
878 if (stride_con == 1 || stride_con == -1)
879 return in(Limit);
880 return this;
881 }
883 //=============================================================================
884 //----------------------match_incr_with_optional_truncation--------------------
885 // Match increment with optional truncation:
886 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
887 // Return NULL for failure. Success returns the increment node.
888 Node* CountedLoopNode::match_incr_with_optional_truncation(
889 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
890 // Quick cutouts:
891 if (expr == NULL || expr->req() != 3) return false;
893 Node *t1 = NULL;
894 Node *t2 = NULL;
895 const TypeInt* trunc_t = TypeInt::INT;
896 Node* n1 = expr;
897 int n1op = n1->Opcode();
899 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
900 if (n1op == Op_AndI &&
901 n1->in(2)->is_Con() &&
902 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
903 // %%% This check should match any mask of 2**K-1.
904 t1 = n1;
905 n1 = t1->in(1);
906 n1op = n1->Opcode();
907 trunc_t = TypeInt::CHAR;
908 } else if (n1op == Op_RShiftI &&
909 n1->in(1) != NULL &&
910 n1->in(1)->Opcode() == Op_LShiftI &&
911 n1->in(2) == n1->in(1)->in(2) &&
912 n1->in(2)->is_Con()) {
913 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
914 // %%% This check should match any shift in [1..31].
915 if (shift == 16 || shift == 8) {
916 t1 = n1;
917 t2 = t1->in(1);
918 n1 = t2->in(1);
919 n1op = n1->Opcode();
920 if (shift == 16) {
921 trunc_t = TypeInt::SHORT;
922 } else if (shift == 8) {
923 trunc_t = TypeInt::BYTE;
924 }
925 }
926 }
928 // If (maybe after stripping) it is an AddI, we won:
929 if (n1op == Op_AddI) {
930 *trunc1 = t1;
931 *trunc2 = t2;
932 *trunc_type = trunc_t;
933 return n1;
934 }
936 // failed
937 return NULL;
938 }
941 //------------------------------filtered_type--------------------------------
942 // Return a type based on condition control flow
943 // A successful return will be a type that is restricted due
944 // to a series of dominating if-tests, such as:
945 // if (i < 10) {
946 // if (i > 0) {
947 // here: "i" type is [1..10)
948 // }
949 // }
950 // or a control flow merge
951 // if (i < 10) {
952 // do {
953 // phi( , ) -- at top of loop type is [min_int..10)
954 // i = ?
955 // } while ( i < 10)
956 //
957 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
958 assert(n && n->bottom_type()->is_int(), "must be int");
959 const TypeInt* filtered_t = NULL;
960 if (!n->is_Phi()) {
961 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
962 filtered_t = filtered_type_from_dominators(n, n_ctrl);
964 } else {
965 Node* phi = n->as_Phi();
966 Node* region = phi->in(0);
967 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
968 if (region && region != C->top()) {
969 for (uint i = 1; i < phi->req(); i++) {
970 Node* val = phi->in(i);
971 Node* use_c = region->in(i);
972 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
973 if (val_t != NULL) {
974 if (filtered_t == NULL) {
975 filtered_t = val_t;
976 } else {
977 filtered_t = filtered_t->meet(val_t)->is_int();
978 }
979 }
980 }
981 }
982 }
983 const TypeInt* n_t = _igvn.type(n)->is_int();
984 if (filtered_t != NULL) {
985 n_t = n_t->join(filtered_t)->is_int();
986 }
987 return n_t;
988 }
991 //------------------------------filtered_type_from_dominators--------------------------------
992 // Return a possibly more restrictive type for val based on condition control flow of dominators
993 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
994 if (val->is_Con()) {
995 return val->bottom_type()->is_int();
996 }
997 uint if_limit = 10; // Max number of dominating if's visited
998 const TypeInt* rtn_t = NULL;
1000 if (use_ctrl && use_ctrl != C->top()) {
1001 Node* val_ctrl = get_ctrl(val);
1002 uint val_dom_depth = dom_depth(val_ctrl);
1003 Node* pred = use_ctrl;
1004 uint if_cnt = 0;
1005 while (if_cnt < if_limit) {
1006 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1007 if_cnt++;
1008 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1009 if (if_t != NULL) {
1010 if (rtn_t == NULL) {
1011 rtn_t = if_t;
1012 } else {
1013 rtn_t = rtn_t->join(if_t)->is_int();
1014 }
1015 }
1016 }
1017 pred = idom(pred);
1018 if (pred == NULL || pred == C->top()) {
1019 break;
1020 }
1021 // Stop if going beyond definition block of val
1022 if (dom_depth(pred) < val_dom_depth) {
1023 break;
1024 }
1025 }
1026 }
1027 return rtn_t;
1028 }
1031 //------------------------------dump_spec--------------------------------------
1032 // Dump special per-node info
1033 #ifndef PRODUCT
1034 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1035 if( in(TestValue)->is_Bool() ) {
1036 BoolTest bt( test_trip()); // Added this for g++.
1038 st->print("[");
1039 bt.dump_on(st);
1040 st->print("]");
1041 }
1042 st->print(" ");
1043 IfNode::dump_spec(st);
1044 }
1045 #endif
1047 //=============================================================================
1048 //------------------------------is_member--------------------------------------
1049 // Is 'l' a member of 'this'?
1050 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
1051 while( l->_nest > _nest ) l = l->_parent;
1052 return l == this;
1053 }
1055 //------------------------------set_nest---------------------------------------
1056 // Set loop tree nesting depth. Accumulate _has_call bits.
1057 int IdealLoopTree::set_nest( uint depth ) {
1058 _nest = depth;
1059 int bits = _has_call;
1060 if( _child ) bits |= _child->set_nest(depth+1);
1061 if( bits ) _has_call = 1;
1062 if( _next ) bits |= _next ->set_nest(depth );
1063 return bits;
1064 }
1066 //------------------------------split_fall_in----------------------------------
1067 // Split out multiple fall-in edges from the loop header. Move them to a
1068 // private RegionNode before the loop. This becomes the loop landing pad.
1069 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1070 PhaseIterGVN &igvn = phase->_igvn;
1071 uint i;
1073 // Make a new RegionNode to be the landing pad.
1074 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 );
1075 phase->set_loop(landing_pad,_parent);
1076 // Gather all the fall-in control paths into the landing pad
1077 uint icnt = fall_in_cnt;
1078 uint oreq = _head->req();
1079 for( i = oreq-1; i>0; i-- )
1080 if( !phase->is_member( this, _head->in(i) ) )
1081 landing_pad->set_req(icnt--,_head->in(i));
1083 // Peel off PhiNode edges as well
1084 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1085 Node *oj = _head->fast_out(j);
1086 if( oj->is_Phi() ) {
1087 PhiNode* old_phi = oj->as_Phi();
1088 assert( old_phi->region() == _head, "" );
1089 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1090 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1091 uint icnt = fall_in_cnt;
1092 for( i = oreq-1; i>0; i-- ) {
1093 if( !phase->is_member( this, _head->in(i) ) ) {
1094 p->init_req(icnt--, old_phi->in(i));
1095 // Go ahead and clean out old edges from old phi
1096 old_phi->del_req(i);
1097 }
1098 }
1099 // Search for CSE's here, because ZKM.jar does a lot of
1100 // loop hackery and we need to be a little incremental
1101 // with the CSE to avoid O(N^2) node blow-up.
1102 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1103 if( p2 ) { // Found CSE
1104 p->destruct(); // Recover useless new node
1105 p = p2; // Use old node
1106 } else {
1107 igvn.register_new_node_with_optimizer(p, old_phi);
1108 }
1109 // Make old Phi refer to new Phi.
1110 old_phi->add_req(p);
1111 // Check for the special case of making the old phi useless and
1112 // disappear it. In JavaGrande I have a case where this useless
1113 // Phi is the loop limit and prevents recognizing a CountedLoop
1114 // which in turn prevents removing an empty loop.
1115 Node *id_old_phi = old_phi->Identity( &igvn );
1116 if( id_old_phi != old_phi ) { // Found a simple identity?
1117 // Note that I cannot call 'replace_node' here, because
1118 // that will yank the edge from old_phi to the Region and
1119 // I'm mid-iteration over the Region's uses.
1120 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1121 Node* use = old_phi->last_out(i);
1122 igvn.hash_delete(use);
1123 igvn._worklist.push(use);
1124 uint uses_found = 0;
1125 for (uint j = 0; j < use->len(); j++) {
1126 if (use->in(j) == old_phi) {
1127 if (j < use->req()) use->set_req (j, id_old_phi);
1128 else use->set_prec(j, id_old_phi);
1129 uses_found++;
1130 }
1131 }
1132 i -= uses_found; // we deleted 1 or more copies of this edge
1133 }
1134 }
1135 igvn._worklist.push(old_phi);
1136 }
1137 }
1138 // Finally clean out the fall-in edges from the RegionNode
1139 for( i = oreq-1; i>0; i-- ) {
1140 if( !phase->is_member( this, _head->in(i) ) ) {
1141 _head->del_req(i);
1142 }
1143 }
1144 // Transform landing pad
1145 igvn.register_new_node_with_optimizer(landing_pad, _head);
1146 // Insert landing pad into the header
1147 _head->add_req(landing_pad);
1148 }
1150 //------------------------------split_outer_loop-------------------------------
1151 // Split out the outermost loop from this shared header.
1152 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1153 PhaseIterGVN &igvn = phase->_igvn;
1155 // Find index of outermost loop; it should also be my tail.
1156 uint outer_idx = 1;
1157 while( _head->in(outer_idx) != _tail ) outer_idx++;
1159 // Make a LoopNode for the outermost loop.
1160 Node *ctl = _head->in(LoopNode::EntryControl);
1161 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
1162 outer = igvn.register_new_node_with_optimizer(outer, _head);
1163 phase->set_created_loop_node();
1165 Node* pred = phase->clone_loop_predicates(ctl, outer, true);
1166 // Outermost loop falls into '_head' loop
1167 _head->set_req(LoopNode::EntryControl, pred);
1168 _head->del_req(outer_idx);
1169 // Split all the Phis up between '_head' loop and 'outer' loop.
1170 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1171 Node *out = _head->fast_out(j);
1172 if( out->is_Phi() ) {
1173 PhiNode *old_phi = out->as_Phi();
1174 assert( old_phi->region() == _head, "" );
1175 Node *phi = PhiNode::make_blank(outer, old_phi);
1176 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1177 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1178 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1179 // Make old Phi point to new Phi on the fall-in path
1180 igvn.hash_delete(old_phi);
1181 old_phi->set_req(LoopNode::EntryControl, phi);
1182 old_phi->del_req(outer_idx);
1183 igvn._worklist.push(old_phi);
1184 }
1185 }
1187 // Use the new loop head instead of the old shared one
1188 _head = outer;
1189 phase->set_loop(_head, this);
1190 }
1192 //------------------------------fix_parent-------------------------------------
1193 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1194 loop->_parent = parent;
1195 if( loop->_child ) fix_parent( loop->_child, loop );
1196 if( loop->_next ) fix_parent( loop->_next , parent );
1197 }
1199 //------------------------------estimate_path_freq-----------------------------
1200 static float estimate_path_freq( Node *n ) {
1201 // Try to extract some path frequency info
1202 IfNode *iff;
1203 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1204 uint nop = n->Opcode();
1205 if( nop == Op_SafePoint ) { // Skip any safepoint
1206 n = n->in(0);
1207 continue;
1208 }
1209 if( nop == Op_CatchProj ) { // Get count from a prior call
1210 // Assume call does not always throw exceptions: means the call-site
1211 // count is also the frequency of the fall-through path.
1212 assert( n->is_CatchProj(), "" );
1213 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1214 return 0.0f; // Assume call exception path is rare
1215 Node *call = n->in(0)->in(0)->in(0);
1216 assert( call->is_Call(), "expect a call here" );
1217 const JVMState *jvms = ((CallNode*)call)->jvms();
1218 ciMethodData* methodData = jvms->method()->method_data();
1219 if (!methodData->is_mature()) return 0.0f; // No call-site data
1220 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1221 if ((data == NULL) || !data->is_CounterData()) {
1222 // no call profile available, try call's control input
1223 n = n->in(0);
1224 continue;
1225 }
1226 return data->as_CounterData()->count()/FreqCountInvocations;
1227 }
1228 // See if there's a gating IF test
1229 Node *n_c = n->in(0);
1230 if( !n_c->is_If() ) break; // No estimate available
1231 iff = n_c->as_If();
1232 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1233 // Compute how much count comes on this path
1234 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1235 // Have no count info. Skip dull uncommon-trap like branches.
1236 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1237 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1238 break;
1239 // Skip through never-taken branch; look for a real loop exit.
1240 n = iff->in(0);
1241 }
1242 return 0.0f; // No estimate available
1243 }
1245 //------------------------------merge_many_backedges---------------------------
1246 // Merge all the backedges from the shared header into a private Region.
1247 // Feed that region as the one backedge to this loop.
1248 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1249 uint i;
1251 // Scan for the top 2 hottest backedges
1252 float hotcnt = 0.0f;
1253 float warmcnt = 0.0f;
1254 uint hot_idx = 0;
1255 // Loop starts at 2 because slot 1 is the fall-in path
1256 for( i = 2; i < _head->req(); i++ ) {
1257 float cnt = estimate_path_freq(_head->in(i));
1258 if( cnt > hotcnt ) { // Grab hottest path
1259 warmcnt = hotcnt;
1260 hotcnt = cnt;
1261 hot_idx = i;
1262 } else if( cnt > warmcnt ) { // And 2nd hottest path
1263 warmcnt = cnt;
1264 }
1265 }
1267 // See if the hottest backedge is worthy of being an inner loop
1268 // by being much hotter than the next hottest backedge.
1269 if( hotcnt <= 0.0001 ||
1270 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1272 // Peel out the backedges into a private merge point; peel
1273 // them all except optionally hot_idx.
1274 PhaseIterGVN &igvn = phase->_igvn;
1276 Node *hot_tail = NULL;
1277 // Make a Region for the merge point
1278 Node *r = new (phase->C, 1) RegionNode(1);
1279 for( i = 2; i < _head->req(); i++ ) {
1280 if( i != hot_idx )
1281 r->add_req( _head->in(i) );
1282 else hot_tail = _head->in(i);
1283 }
1284 igvn.register_new_node_with_optimizer(r, _head);
1285 // Plug region into end of loop _head, followed by hot_tail
1286 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1287 _head->set_req(2, r);
1288 if( hot_idx ) _head->add_req(hot_tail);
1290 // Split all the Phis up between '_head' loop and the Region 'r'
1291 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1292 Node *out = _head->fast_out(j);
1293 if( out->is_Phi() ) {
1294 PhiNode* n = out->as_Phi();
1295 igvn.hash_delete(n); // Delete from hash before hacking edges
1296 Node *hot_phi = NULL;
1297 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type());
1298 // Check all inputs for the ones to peel out
1299 uint j = 1;
1300 for( uint i = 2; i < n->req(); i++ ) {
1301 if( i != hot_idx )
1302 phi->set_req( j++, n->in(i) );
1303 else hot_phi = n->in(i);
1304 }
1305 // Register the phi but do not transform until whole place transforms
1306 igvn.register_new_node_with_optimizer(phi, n);
1307 // Add the merge phi to the old Phi
1308 while( n->req() > 3 ) n->del_req( n->req()-1 );
1309 n->set_req(2, phi);
1310 if( hot_idx ) n->add_req(hot_phi);
1311 }
1312 }
1315 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1316 // of self loop tree. Turn self into a loop headed by _head and with
1317 // tail being the new merge point.
1318 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1319 phase->set_loop(_tail,ilt); // Adjust tail
1320 _tail = r; // Self's tail is new merge point
1321 phase->set_loop(r,this);
1322 ilt->_child = _child; // New guy has my children
1323 _child = ilt; // Self has new guy as only child
1324 ilt->_parent = this; // new guy has self for parent
1325 ilt->_nest = _nest; // Same nesting depth (for now)
1327 // Starting with 'ilt', look for child loop trees using the same shared
1328 // header. Flatten these out; they will no longer be loops in the end.
1329 IdealLoopTree **pilt = &_child;
1330 while( ilt ) {
1331 if( ilt->_head == _head ) {
1332 uint i;
1333 for( i = 2; i < _head->req(); i++ )
1334 if( _head->in(i) == ilt->_tail )
1335 break; // Still a loop
1336 if( i == _head->req() ) { // No longer a loop
1337 // Flatten ilt. Hang ilt's "_next" list from the end of
1338 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1339 IdealLoopTree **cp = &ilt->_child;
1340 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1341 *cp = ilt->_next; // Hang next list at end of child list
1342 *pilt = ilt->_child; // Move child up to replace ilt
1343 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1344 ilt = ilt->_child; // Repeat using new ilt
1345 continue; // do not advance over ilt->_child
1346 }
1347 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1348 phase->set_loop(_head,ilt);
1349 }
1350 pilt = &ilt->_child; // Advance to next
1351 ilt = *pilt;
1352 }
1354 if( _child ) fix_parent( _child, this );
1355 }
1357 //------------------------------beautify_loops---------------------------------
1358 // Split shared headers and insert loop landing pads.
1359 // Insert a LoopNode to replace the RegionNode.
1360 // Return TRUE if loop tree is structurally changed.
1361 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1362 bool result = false;
1363 // Cache parts in locals for easy
1364 PhaseIterGVN &igvn = phase->_igvn;
1366 igvn.hash_delete(_head); // Yank from hash before hacking edges
1368 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1369 int fall_in_cnt = 0;
1370 for( uint i = 1; i < _head->req(); i++ )
1371 if( !phase->is_member( this, _head->in(i) ) )
1372 fall_in_cnt++;
1373 assert( fall_in_cnt, "at least 1 fall-in path" );
1374 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1375 split_fall_in( phase, fall_in_cnt );
1377 // Swap inputs to the _head and all Phis to move the fall-in edge to
1378 // the left.
1379 fall_in_cnt = 1;
1380 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1381 fall_in_cnt++;
1382 if( fall_in_cnt > 1 ) {
1383 // Since I am just swapping inputs I do not need to update def-use info
1384 Node *tmp = _head->in(1);
1385 _head->set_req( 1, _head->in(fall_in_cnt) );
1386 _head->set_req( fall_in_cnt, tmp );
1387 // Swap also all Phis
1388 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1389 Node* phi = _head->fast_out(i);
1390 if( phi->is_Phi() ) {
1391 igvn.hash_delete(phi); // Yank from hash before hacking edges
1392 tmp = phi->in(1);
1393 phi->set_req( 1, phi->in(fall_in_cnt) );
1394 phi->set_req( fall_in_cnt, tmp );
1395 }
1396 }
1397 }
1398 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1399 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1401 // If I am a shared header (multiple backedges), peel off the many
1402 // backedges into a private merge point and use the merge point as
1403 // the one true backedge.
1404 if( _head->req() > 3 ) {
1405 // Merge the many backedges into a single backedge but leave
1406 // the hottest backedge as separate edge for the following peel.
1407 merge_many_backedges( phase );
1408 result = true;
1409 }
1411 // If I have one hot backedge, peel off myself loop.
1412 // I better be the outermost loop.
1413 if( _head->req() > 3 ) {
1414 split_outer_loop( phase );
1415 result = true;
1417 } else if( !_head->is_Loop() && !_irreducible ) {
1418 // Make a new LoopNode to replace the old loop head
1419 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) );
1420 l = igvn.register_new_node_with_optimizer(l, _head);
1421 phase->set_created_loop_node();
1422 // Go ahead and replace _head
1423 phase->_igvn.replace_node( _head, l );
1424 _head = l;
1425 phase->set_loop(_head, this);
1426 }
1428 // Now recursively beautify nested loops
1429 if( _child ) result |= _child->beautify_loops( phase );
1430 if( _next ) result |= _next ->beautify_loops( phase );
1431 return result;
1432 }
1434 //------------------------------allpaths_check_safepts----------------------------
1435 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1436 // encountered. Helper for check_safepts.
1437 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1438 assert(stack.size() == 0, "empty stack");
1439 stack.push(_tail);
1440 visited.Clear();
1441 visited.set(_tail->_idx);
1442 while (stack.size() > 0) {
1443 Node* n = stack.pop();
1444 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1445 // Terminate this path
1446 } else if (n->Opcode() == Op_SafePoint) {
1447 if (_phase->get_loop(n) != this) {
1448 if (_required_safept == NULL) _required_safept = new Node_List();
1449 _required_safept->push(n); // save the one closest to the tail
1450 }
1451 // Terminate this path
1452 } else {
1453 uint start = n->is_Region() ? 1 : 0;
1454 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1455 for (uint i = start; i < end; i++) {
1456 Node* in = n->in(i);
1457 assert(in->is_CFG(), "must be");
1458 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1459 stack.push(in);
1460 }
1461 }
1462 }
1463 }
1464 }
1466 //------------------------------check_safepts----------------------------
1467 // Given dominators, try to find loops with calls that must always be
1468 // executed (call dominates loop tail). These loops do not need non-call
1469 // safepoints (ncsfpt).
1470 //
1471 // A complication is that a safepoint in a inner loop may be needed
1472 // by an outer loop. In the following, the inner loop sees it has a
1473 // call (block 3) on every path from the head (block 2) to the
1474 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1475 // in block 2, _but_ this leaves the outer loop without a safepoint.
1476 //
1477 // entry 0
1478 // |
1479 // v
1480 // outer 1,2 +->1
1481 // | |
1482 // | v
1483 // | 2<---+ ncsfpt in 2
1484 // |_/|\ |
1485 // | v |
1486 // inner 2,3 / 3 | call in 3
1487 // / | |
1488 // v +--+
1489 // exit 4
1490 //
1491 //
1492 // This method creates a list (_required_safept) of ncsfpt nodes that must
1493 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1494 // is first looked for in the lists for the outer loops of the current loop.
1495 //
1496 // The insights into the problem:
1497 // A) counted loops are okay
1498 // B) innermost loops are okay (only an inner loop can delete
1499 // a ncsfpt needed by an outer loop)
1500 // C) a loop is immune from an inner loop deleting a safepoint
1501 // if the loop has a call on the idom-path
1502 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1503 // idom-path that is not in a nested loop
1504 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1505 // loop needs to be prevented from deletion by an inner loop
1506 //
1507 // There are two analyses:
1508 // 1) The first, and cheaper one, scans the loop body from
1509 // tail to head following the idom (immediate dominator)
1510 // chain, looking for the cases (C,D,E) above.
1511 // Since inner loops are scanned before outer loops, there is summary
1512 // information about inner loops. Inner loops can be skipped over
1513 // when the tail of an inner loop is encountered.
1514 //
1515 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1516 // the idom path (which is rare), scans all predecessor control paths
1517 // from the tail to the head, terminating a path when a call or sfpt
1518 // is encountered, to find the ncsfpt's that are closest to the tail.
1519 //
1520 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1521 // Bottom up traversal
1522 IdealLoopTree* ch = _child;
1523 while (ch != NULL) {
1524 ch->check_safepts(visited, stack);
1525 ch = ch->_next;
1526 }
1528 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1529 bool has_call = false; // call on dom-path
1530 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1531 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1532 // Scan the dom-path nodes from tail to head
1533 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1534 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1535 has_call = true;
1536 _has_sfpt = 1; // Then no need for a safept!
1537 break;
1538 } else if (n->Opcode() == Op_SafePoint) {
1539 if (_phase->get_loop(n) == this) {
1540 has_local_ncsfpt = true;
1541 break;
1542 }
1543 if (nonlocal_ncsfpt == NULL) {
1544 nonlocal_ncsfpt = n; // save the one closest to the tail
1545 }
1546 } else {
1547 IdealLoopTree* nlpt = _phase->get_loop(n);
1548 if (this != nlpt) {
1549 // If at an inner loop tail, see if the inner loop has already
1550 // recorded seeing a call on the dom-path (and stop.) If not,
1551 // jump to the head of the inner loop.
1552 assert(is_member(nlpt), "nested loop");
1553 Node* tail = nlpt->_tail;
1554 if (tail->in(0)->is_If()) tail = tail->in(0);
1555 if (n == tail) {
1556 // If inner loop has call on dom-path, so does outer loop
1557 if (nlpt->_has_sfpt) {
1558 has_call = true;
1559 _has_sfpt = 1;
1560 break;
1561 }
1562 // Skip to head of inner loop
1563 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1564 n = nlpt->_head;
1565 }
1566 }
1567 }
1568 }
1569 // Record safept's that this loop needs preserved when an
1570 // inner loop attempts to delete it's safepoints.
1571 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1572 if (nonlocal_ncsfpt != NULL) {
1573 if (_required_safept == NULL) _required_safept = new Node_List();
1574 _required_safept->push(nonlocal_ncsfpt);
1575 } else {
1576 // Failed to find a suitable safept on the dom-path. Now use
1577 // an all paths walk from tail to head, looking for safepoints to preserve.
1578 allpaths_check_safepts(visited, stack);
1579 }
1580 }
1581 }
1582 }
1584 //---------------------------is_deleteable_safept----------------------------
1585 // Is safept not required by an outer loop?
1586 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1587 assert(sfpt->Opcode() == Op_SafePoint, "");
1588 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1589 while (lp != NULL) {
1590 Node_List* sfpts = lp->_required_safept;
1591 if (sfpts != NULL) {
1592 for (uint i = 0; i < sfpts->size(); i++) {
1593 if (sfpt == sfpts->at(i))
1594 return false;
1595 }
1596 }
1597 lp = lp->_parent;
1598 }
1599 return true;
1600 }
1602 //---------------------------replace_parallel_iv-------------------------------
1603 // Replace parallel induction variable (parallel to trip counter)
1604 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1605 assert(loop->_head->is_CountedLoop(), "");
1606 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1607 Node *incr = cl->incr();
1608 if (incr == NULL)
1609 return; // Dead loop?
1610 Node *init = cl->init_trip();
1611 Node *phi = cl->phi();
1612 // protect against stride not being a constant
1613 if (!cl->stride_is_con())
1614 return;
1615 int stride_con = cl->stride_con();
1617 PhaseGVN *gvn = &_igvn;
1619 // Visit all children, looking for Phis
1620 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1621 Node *out = cl->out(i);
1622 // Look for other phis (secondary IVs). Skip dead ones
1623 if (!out->is_Phi() || out == phi || !has_node(out))
1624 continue;
1625 PhiNode* phi2 = out->as_Phi();
1626 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1627 // Look for induction variables of the form: X += constant
1628 if (phi2->region() != loop->_head ||
1629 incr2->req() != 3 ||
1630 incr2->in(1) != phi2 ||
1631 incr2 == incr ||
1632 incr2->Opcode() != Op_AddI ||
1633 !incr2->in(2)->is_Con())
1634 continue;
1636 // Check for parallel induction variable (parallel to trip counter)
1637 // via an affine function. In particular, count-down loops with
1638 // count-up array indices are common. We only RCE references off
1639 // the trip-counter, so we need to convert all these to trip-counter
1640 // expressions.
1641 Node *init2 = phi2->in( LoopNode::EntryControl );
1642 int stride_con2 = incr2->in(2)->get_int();
1644 // The general case here gets a little tricky. We want to find the
1645 // GCD of all possible parallel IV's and make a new IV using this
1646 // GCD for the loop. Then all possible IVs are simple multiples of
1647 // the GCD. In practice, this will cover very few extra loops.
1648 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1649 // where +/-1 is the common case, but other integer multiples are
1650 // also easy to handle.
1651 int ratio_con = stride_con2/stride_con;
1653 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1654 // Convert to using the trip counter. The parallel induction
1655 // variable differs from the trip counter by a loop-invariant
1656 // amount, the difference between their respective initial values.
1657 // It is scaled by the 'ratio_con'.
1658 // Perform local Ideal transformation since in most cases ratio == 1.
1659 Node* ratio = _igvn.intcon(ratio_con);
1660 set_ctrl(ratio, C->root());
1661 Node* hook = new (C, 3) Node(3);
1662 Node* ratio_init = gvn->transform(new (C, 3) MulINode(init, ratio));
1663 hook->init_req(0, ratio_init);
1664 Node* diff = gvn->transform(new (C, 3) SubINode(init2, ratio_init));
1665 hook->init_req(1, diff);
1666 Node* ratio_idx = gvn->transform(new (C, 3) MulINode(phi, ratio));
1667 hook->init_req(2, ratio_idx);
1668 Node* add = gvn->transform(new (C, 3) AddINode(ratio_idx, diff));
1669 set_subtree_ctrl(add);
1670 _igvn.replace_node( phi2, add );
1671 // Free up intermediate goo
1672 _igvn.remove_dead_node(hook);
1673 // Sometimes an induction variable is unused
1674 if (add->outcnt() == 0) {
1675 _igvn.remove_dead_node(add);
1676 }
1677 --i; // deleted this phi; rescan starting with next position
1678 continue;
1679 }
1680 }
1681 }
1683 //------------------------------counted_loop-----------------------------------
1684 // Convert to counted loops where possible
1685 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1687 // For grins, set the inner-loop flag here
1688 if (!_child) {
1689 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1690 }
1692 if (_head->is_CountedLoop() ||
1693 phase->is_counted_loop(_head, this)) {
1694 _has_sfpt = 1; // Indicate we do not need a safepoint here
1696 // Look for a safepoint to remove
1697 for (Node* n = tail(); n != _head; n = phase->idom(n))
1698 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1699 phase->is_deleteable_safept(n))
1700 phase->lazy_replace(n,n->in(TypeFunc::Control));
1702 // Look for induction variables
1703 phase->replace_parallel_iv(this);
1705 } else if (_parent != NULL && !_irreducible) {
1706 // Not a counted loop.
1707 // Look for a safepoint on the idom-path to remove, preserving the first one
1708 bool found = false;
1709 Node* n = tail();
1710 for (; n != _head && !found; n = phase->idom(n)) {
1711 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this)
1712 found = true; // Found one
1713 }
1714 // Skip past it and delete the others
1715 for (; n != _head; n = phase->idom(n)) {
1716 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1717 phase->is_deleteable_safept(n))
1718 phase->lazy_replace(n,n->in(TypeFunc::Control));
1719 }
1720 }
1722 // Recursively
1723 if (_child) _child->counted_loop( phase );
1724 if (_next) _next ->counted_loop( phase );
1725 }
1727 #ifndef PRODUCT
1728 //------------------------------dump_head--------------------------------------
1729 // Dump 1 liner for loop header info
1730 void IdealLoopTree::dump_head( ) const {
1731 for (uint i=0; i<_nest; i++)
1732 tty->print(" ");
1733 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1734 if (_irreducible) tty->print(" IRREDUCIBLE");
1735 Node* entry = _head->in(LoopNode::EntryControl);
1736 if (LoopLimitCheck) {
1737 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1738 if (predicate != NULL ) {
1739 tty->print(" limit_check");
1740 entry = entry->in(0)->in(0);
1741 }
1742 }
1743 if (UseLoopPredicate) {
1744 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1745 if (entry != NULL) {
1746 tty->print(" predicated");
1747 }
1748 }
1749 if (_head->is_CountedLoop()) {
1750 CountedLoopNode *cl = _head->as_CountedLoop();
1751 tty->print(" counted");
1753 Node* init_n = cl->init_trip();
1754 if (init_n != NULL && init_n->is_Con())
1755 tty->print(" [%d,", cl->init_trip()->get_int());
1756 else
1757 tty->print(" [int,");
1758 Node* limit_n = cl->limit();
1759 if (limit_n != NULL && limit_n->is_Con())
1760 tty->print("%d),", cl->limit()->get_int());
1761 else
1762 tty->print("int),");
1763 int stride_con = cl->stride_con();
1764 if (stride_con > 0) tty->print("+");
1765 tty->print("%d", stride_con);
1767 if (cl->is_pre_loop ()) tty->print(" pre" );
1768 if (cl->is_main_loop()) tty->print(" main");
1769 if (cl->is_post_loop()) tty->print(" post");
1770 }
1771 tty->cr();
1772 }
1774 //------------------------------dump-------------------------------------------
1775 // Dump loops by loop tree
1776 void IdealLoopTree::dump( ) const {
1777 dump_head();
1778 if (_child) _child->dump();
1779 if (_next) _next ->dump();
1780 }
1782 #endif
1784 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1785 if (loop == root) {
1786 if (loop->_child != NULL) {
1787 log->begin_head("loop_tree");
1788 log->end_head();
1789 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1790 log->tail("loop_tree");
1791 assert(loop->_next == NULL, "what?");
1792 }
1793 } else {
1794 Node* head = loop->_head;
1795 log->begin_head("loop");
1796 log->print(" idx='%d' ", head->_idx);
1797 if (loop->_irreducible) log->print("irreducible='1' ");
1798 if (head->is_Loop()) {
1799 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1800 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1801 }
1802 if (head->is_CountedLoop()) {
1803 CountedLoopNode* cl = head->as_CountedLoop();
1804 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1805 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1806 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1807 }
1808 log->end_head();
1809 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1810 log->tail("loop");
1811 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1812 }
1813 }
1815 //---------------------collect_potentially_useful_predicates-----------------------
1816 // Helper function to collect potentially useful predicates to prevent them from
1817 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1818 void PhaseIdealLoop::collect_potentially_useful_predicates(
1819 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1820 if (loop->_child) { // child
1821 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1822 }
1824 // self (only loops that we can apply loop predication may use their predicates)
1825 if (loop->_head->is_Loop() &&
1826 !loop->_irreducible &&
1827 !loop->tail()->is_top()) {
1828 LoopNode* lpn = loop->_head->as_Loop();
1829 Node* entry = lpn->in(LoopNode::EntryControl);
1830 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
1831 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1832 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1833 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1834 entry = entry->in(0)->in(0);
1835 }
1836 predicate_proj = find_predicate(entry); // Predicate
1837 if (predicate_proj != NULL ) {
1838 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1839 }
1840 }
1842 if (loop->_next) { // sibling
1843 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1844 }
1845 }
1847 //------------------------eliminate_useless_predicates-----------------------------
1848 // Eliminate all inserted predicates if they could not be used by loop predication.
1849 // Note: it will also eliminates loop limits check predicate since it also uses
1850 // Opaque1 node (see Parse::add_predicate()).
1851 void PhaseIdealLoop::eliminate_useless_predicates() {
1852 if (C->predicate_count() == 0)
1853 return; // no predicate left
1855 Unique_Node_List useful_predicates; // to store useful predicates
1856 if (C->has_loops()) {
1857 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
1858 }
1860 for (int i = C->predicate_count(); i > 0; i--) {
1861 Node * n = C->predicate_opaque1_node(i-1);
1862 assert(n->Opcode() == Op_Opaque1, "must be");
1863 if (!useful_predicates.member(n)) { // not in the useful list
1864 _igvn.replace_node(n, n->in(1));
1865 }
1866 }
1867 }
1869 //=============================================================================
1870 //----------------------------build_and_optimize-------------------------------
1871 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1872 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1873 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs) {
1874 ResourceMark rm;
1876 int old_progress = C->major_progress();
1877 uint orig_worklist_size = _igvn._worklist.size();
1879 // Reset major-progress flag for the driver's heuristics
1880 C->clear_major_progress();
1882 #ifndef PRODUCT
1883 // Capture for later assert
1884 uint unique = C->unique();
1885 _loop_invokes++;
1886 _loop_work += unique;
1887 #endif
1889 // True if the method has at least 1 irreducible loop
1890 _has_irreducible_loops = false;
1892 _created_loop_node = false;
1894 Arena *a = Thread::current()->resource_area();
1895 VectorSet visited(a);
1896 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1897 _nodes.map(C->unique(), NULL);
1898 memset(_nodes.adr(), 0, wordSize * C->unique());
1900 // Pre-build the top-level outermost loop tree entry
1901 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1902 // Do not need a safepoint at the top level
1903 _ltree_root->_has_sfpt = 1;
1905 // Initialize Dominators.
1906 // Checked in clone_loop_predicate() during beautify_loops().
1907 _idom_size = 0;
1908 _idom = NULL;
1909 _dom_depth = NULL;
1910 _dom_stk = NULL;
1912 // Empty pre-order array
1913 allocate_preorders();
1915 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1916 // IdealLoopTree entries. Data nodes are NOT walked.
1917 build_loop_tree();
1918 // Check for bailout, and return
1919 if (C->failing()) {
1920 return;
1921 }
1923 // No loops after all
1924 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
1926 // There should always be an outer loop containing the Root and Return nodes.
1927 // If not, we have a degenerate empty program. Bail out in this case.
1928 if (!has_node(C->root())) {
1929 if (!_verify_only) {
1930 C->clear_major_progress();
1931 C->record_method_not_compilable("empty program detected during loop optimization");
1932 }
1933 return;
1934 }
1936 // Nothing to do, so get out
1937 if( !C->has_loops() && !do_split_ifs && !_verify_me && !_verify_only ) {
1938 _igvn.optimize(); // Cleanup NeverBranches
1939 return;
1940 }
1942 // Set loop nesting depth
1943 _ltree_root->set_nest( 0 );
1945 // Split shared headers and insert loop landing pads.
1946 // Do not bother doing this on the Root loop of course.
1947 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
1948 C->print_method("Before beautify loops", 3);
1949 if( _ltree_root->_child->beautify_loops( this ) ) {
1950 // Re-build loop tree!
1951 _ltree_root->_child = NULL;
1952 _nodes.clear();
1953 reallocate_preorders();
1954 build_loop_tree();
1955 // Check for bailout, and return
1956 if (C->failing()) {
1957 return;
1958 }
1959 // Reset loop nesting depth
1960 _ltree_root->set_nest( 0 );
1962 C->print_method("After beautify loops", 3);
1963 }
1964 }
1966 // Build Dominators for elision of NULL checks & loop finding.
1967 // Since nodes do not have a slot for immediate dominator, make
1968 // a persistent side array for that info indexed on node->_idx.
1969 _idom_size = C->unique();
1970 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1971 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1972 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1973 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
1975 Dominators();
1977 if (!_verify_only) {
1978 // As a side effect, Dominators removed any unreachable CFG paths
1979 // into RegionNodes. It doesn't do this test against Root, so
1980 // we do it here.
1981 for( uint i = 1; i < C->root()->req(); i++ ) {
1982 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
1983 _igvn.hash_delete(C->root());
1984 C->root()->del_req(i);
1985 _igvn._worklist.push(C->root());
1986 i--; // Rerun same iteration on compressed edges
1987 }
1988 }
1990 // Given dominators, try to find inner loops with calls that must
1991 // always be executed (call dominates loop tail). These loops do
1992 // not need a separate safepoint.
1993 Node_List cisstack(a);
1994 _ltree_root->check_safepts(visited, cisstack);
1995 }
1997 // Walk the DATA nodes and place into loops. Find earliest control
1998 // node. For CFG nodes, the _nodes array starts out and remains
1999 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2000 // _nodes array holds the earliest legal controlling CFG node.
2002 // Allocate stack with enough space to avoid frequent realloc
2003 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
2004 Node_Stack nstack( a, stack_size );
2006 visited.Clear();
2007 Node_List worklist(a);
2008 // Don't need C->root() on worklist since
2009 // it will be processed among C->top() inputs
2010 worklist.push( C->top() );
2011 visited.set( C->top()->_idx ); // Set C->top() as visited now
2012 build_loop_early( visited, worklist, nstack );
2014 // Given early legal placement, try finding counted loops. This placement
2015 // is good enough to discover most loop invariants.
2016 if( !_verify_me && !_verify_only )
2017 _ltree_root->counted_loop( this );
2019 // Find latest loop placement. Find ideal loop placement.
2020 visited.Clear();
2021 init_dom_lca_tags();
2022 // Need C->root() on worklist when processing outs
2023 worklist.push( C->root() );
2024 NOT_PRODUCT( C->verify_graph_edges(); )
2025 worklist.push( C->top() );
2026 build_loop_late( visited, worklist, nstack );
2028 if (_verify_only) {
2029 // restore major progress flag
2030 for (int i = 0; i < old_progress; i++)
2031 C->set_major_progress();
2032 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2033 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2034 return;
2035 }
2037 // Some parser-inserted loop predicates could never be used by loop
2038 // predication or they were moved away from loop during some optimizations.
2039 // For example, peeling. Eliminate them before next loop optimizations.
2040 if (UseLoopPredicate || LoopLimitCheck) {
2041 eliminate_useless_predicates();
2042 }
2044 // clear out the dead code
2045 while(_deadlist.size()) {
2046 _igvn.remove_globally_dead_node(_deadlist.pop());
2047 }
2049 #ifndef PRODUCT
2050 C->verify_graph_edges();
2051 if (_verify_me) { // Nested verify pass?
2052 // Check to see if the verify mode is broken
2053 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2054 return;
2055 }
2056 if(VerifyLoopOptimizations) verify();
2057 if(TraceLoopOpts && C->has_loops()) {
2058 _ltree_root->dump();
2059 }
2060 #endif
2062 if (ReassociateInvariants) {
2063 // Reassociate invariants and prep for split_thru_phi
2064 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2065 IdealLoopTree* lpt = iter.current();
2066 if (!lpt->is_counted() || !lpt->is_inner()) continue;
2068 lpt->reassociate_invariants(this);
2070 // Because RCE opportunities can be masked by split_thru_phi,
2071 // look for RCE candidates and inhibit split_thru_phi
2072 // on just their loop-phi's for this pass of loop opts
2073 if (SplitIfBlocks && do_split_ifs) {
2074 if (lpt->policy_range_check(this)) {
2075 lpt->_rce_candidate = 1; // = true
2076 }
2077 }
2078 }
2079 }
2081 // Check for aggressive application of split-if and other transforms
2082 // that require basic-block info (like cloning through Phi's)
2083 if( SplitIfBlocks && do_split_ifs ) {
2084 visited.Clear();
2085 split_if_with_blocks( visited, nstack );
2086 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2087 }
2089 // Perform loop predication before iteration splitting
2090 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2091 _ltree_root->_child->loop_predication(this);
2092 }
2094 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2095 if (do_intrinsify_fill()) {
2096 C->set_major_progress();
2097 }
2098 }
2100 // Perform iteration-splitting on inner loops. Split iterations to avoid
2101 // range checks or one-shot null checks.
2103 // If split-if's didn't hack the graph too bad (no CFG changes)
2104 // then do loop opts.
2105 if (C->has_loops() && !C->major_progress()) {
2106 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2107 _ltree_root->_child->iteration_split( this, worklist );
2108 // No verify after peeling! GCM has hoisted code out of the loop.
2109 // After peeling, the hoisted code could sink inside the peeled area.
2110 // The peeling code does not try to recompute the best location for
2111 // all the code before the peeled area, so the verify pass will always
2112 // complain about it.
2113 }
2114 // Do verify graph edges in any case
2115 NOT_PRODUCT( C->verify_graph_edges(); );
2117 if (!do_split_ifs) {
2118 // We saw major progress in Split-If to get here. We forced a
2119 // pass with unrolling and not split-if, however more split-if's
2120 // might make progress. If the unrolling didn't make progress
2121 // then the major-progress flag got cleared and we won't try
2122 // another round of Split-If. In particular the ever-common
2123 // instance-of/check-cast pattern requires at least 2 rounds of
2124 // Split-If to clear out.
2125 C->set_major_progress();
2126 }
2128 // Repeat loop optimizations if new loops were seen
2129 if (created_loop_node()) {
2130 C->set_major_progress();
2131 }
2133 // Keep loop predicates and perform optimizations with them
2134 // until no more loop optimizations could be done.
2135 // After that switch predicates off and do more loop optimizations.
2136 if (!C->major_progress() && (C->predicate_count() > 0)) {
2137 C->cleanup_loop_predicates(_igvn);
2138 #ifndef PRODUCT
2139 if (TraceLoopOpts) {
2140 tty->print_cr("PredicatesOff");
2141 }
2142 #endif
2143 C->set_major_progress();
2144 }
2146 // Convert scalar to superword operations at the end of all loop opts.
2147 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2148 // SuperWord transform
2149 SuperWord sw(this);
2150 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2151 IdealLoopTree* lpt = iter.current();
2152 if (lpt->is_counted()) {
2153 sw.transform_loop(lpt);
2154 }
2155 }
2156 }
2158 // Cleanup any modified bits
2159 _igvn.optimize();
2161 // disable assert until issue with split_flow_path is resolved (6742111)
2162 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2163 // "shouldn't introduce irreducible loops");
2165 if (C->log() != NULL) {
2166 log_loop_tree(_ltree_root, _ltree_root, C->log());
2167 }
2168 }
2170 #ifndef PRODUCT
2171 //------------------------------print_statistics-------------------------------
2172 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2173 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2174 void PhaseIdealLoop::print_statistics() {
2175 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2176 }
2178 //------------------------------verify-----------------------------------------
2179 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2180 static int fail; // debug only, so its multi-thread dont care
2181 void PhaseIdealLoop::verify() const {
2182 int old_progress = C->major_progress();
2183 ResourceMark rm;
2184 PhaseIdealLoop loop_verify( _igvn, this );
2185 VectorSet visited(Thread::current()->resource_area());
2187 fail = 0;
2188 verify_compare( C->root(), &loop_verify, visited );
2189 assert( fail == 0, "verify loops failed" );
2190 // Verify loop structure is the same
2191 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2192 // Reset major-progress. It was cleared by creating a verify version of
2193 // PhaseIdealLoop.
2194 for( int i=0; i<old_progress; i++ )
2195 C->set_major_progress();
2196 }
2198 //------------------------------verify_compare---------------------------------
2199 // Make sure me and the given PhaseIdealLoop agree on key data structures
2200 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2201 if( !n ) return;
2202 if( visited.test_set( n->_idx ) ) return;
2203 if( !_nodes[n->_idx] ) { // Unreachable
2204 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2205 return;
2206 }
2208 uint i;
2209 for( i = 0; i < n->req(); i++ )
2210 verify_compare( n->in(i), loop_verify, visited );
2212 // Check the '_nodes' block/loop structure
2213 i = n->_idx;
2214 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2215 if( _nodes[i] != loop_verify->_nodes[i] &&
2216 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2217 tty->print("Mismatched control setting for: ");
2218 n->dump();
2219 if( fail++ > 10 ) return;
2220 Node *c = get_ctrl_no_update(n);
2221 tty->print("We have it as: ");
2222 if( c->in(0) ) c->dump();
2223 else tty->print_cr("N%d",c->_idx);
2224 tty->print("Verify thinks: ");
2225 if( loop_verify->has_ctrl(n) )
2226 loop_verify->get_ctrl_no_update(n)->dump();
2227 else
2228 loop_verify->get_loop_idx(n)->dump();
2229 tty->cr();
2230 }
2231 } else { // We have a loop
2232 IdealLoopTree *us = get_loop_idx(n);
2233 if( loop_verify->has_ctrl(n) ) {
2234 tty->print("Mismatched loop setting for: ");
2235 n->dump();
2236 if( fail++ > 10 ) return;
2237 tty->print("We have it as: ");
2238 us->dump();
2239 tty->print("Verify thinks: ");
2240 loop_verify->get_ctrl_no_update(n)->dump();
2241 tty->cr();
2242 } else if (!C->major_progress()) {
2243 // Loop selection can be messed up if we did a major progress
2244 // operation, like split-if. Do not verify in that case.
2245 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2246 if( us->_head != them->_head || us->_tail != them->_tail ) {
2247 tty->print("Unequals loops for: ");
2248 n->dump();
2249 if( fail++ > 10 ) return;
2250 tty->print("We have it as: ");
2251 us->dump();
2252 tty->print("Verify thinks: ");
2253 them->dump();
2254 tty->cr();
2255 }
2256 }
2257 }
2259 // Check for immediate dominators being equal
2260 if( i >= _idom_size ) {
2261 if( !n->is_CFG() ) return;
2262 tty->print("CFG Node with no idom: ");
2263 n->dump();
2264 return;
2265 }
2266 if( !n->is_CFG() ) return;
2267 if( n == C->root() ) return; // No IDOM here
2269 assert(n->_idx == i, "sanity");
2270 Node *id = idom_no_update(n);
2271 if( id != loop_verify->idom_no_update(n) ) {
2272 tty->print("Unequals idoms for: ");
2273 n->dump();
2274 if( fail++ > 10 ) return;
2275 tty->print("We have it as: ");
2276 id->dump();
2277 tty->print("Verify thinks: ");
2278 loop_verify->idom_no_update(n)->dump();
2279 tty->cr();
2280 }
2282 }
2284 //------------------------------verify_tree------------------------------------
2285 // Verify that tree structures match. Because the CFG can change, siblings
2286 // within the loop tree can be reordered. We attempt to deal with that by
2287 // reordering the verify's loop tree if possible.
2288 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2289 assert( _parent == parent, "Badly formed loop tree" );
2291 // Siblings not in same order? Attempt to re-order.
2292 if( _head != loop->_head ) {
2293 // Find _next pointer to update
2294 IdealLoopTree **pp = &loop->_parent->_child;
2295 while( *pp != loop )
2296 pp = &((*pp)->_next);
2297 // Find proper sibling to be next
2298 IdealLoopTree **nn = &loop->_next;
2299 while( (*nn) && (*nn)->_head != _head )
2300 nn = &((*nn)->_next);
2302 // Check for no match.
2303 if( !(*nn) ) {
2304 // Annoyingly, irreducible loops can pick different headers
2305 // after a major_progress operation, so the rest of the loop
2306 // tree cannot be matched.
2307 if (_irreducible && Compile::current()->major_progress()) return;
2308 assert( 0, "failed to match loop tree" );
2309 }
2311 // Move (*nn) to (*pp)
2312 IdealLoopTree *hit = *nn;
2313 *nn = hit->_next;
2314 hit->_next = loop;
2315 *pp = loop;
2316 loop = hit;
2317 // Now try again to verify
2318 }
2320 assert( _head == loop->_head , "mismatched loop head" );
2321 Node *tail = _tail; // Inline a non-updating version of
2322 while( !tail->in(0) ) // the 'tail()' call.
2323 tail = tail->in(1);
2324 assert( tail == loop->_tail, "mismatched loop tail" );
2326 // Counted loops that are guarded should be able to find their guards
2327 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2328 CountedLoopNode *cl = _head->as_CountedLoop();
2329 Node *init = cl->init_trip();
2330 Node *ctrl = cl->in(LoopNode::EntryControl);
2331 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2332 Node *iff = ctrl->in(0);
2333 assert( iff->Opcode() == Op_If, "" );
2334 Node *bol = iff->in(1);
2335 assert( bol->Opcode() == Op_Bool, "" );
2336 Node *cmp = bol->in(1);
2337 assert( cmp->Opcode() == Op_CmpI, "" );
2338 Node *add = cmp->in(1);
2339 Node *opaq;
2340 if( add->Opcode() == Op_Opaque1 ) {
2341 opaq = add;
2342 } else {
2343 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2344 assert( add == init, "" );
2345 opaq = cmp->in(2);
2346 }
2347 assert( opaq->Opcode() == Op_Opaque1, "" );
2349 }
2351 if (_child != NULL) _child->verify_tree(loop->_child, this);
2352 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2353 // Innermost loops need to verify loop bodies,
2354 // but only if no 'major_progress'
2355 int fail = 0;
2356 if (!Compile::current()->major_progress() && _child == NULL) {
2357 for( uint i = 0; i < _body.size(); i++ ) {
2358 Node *n = _body.at(i);
2359 if (n->outcnt() == 0) continue; // Ignore dead
2360 uint j;
2361 for( j = 0; j < loop->_body.size(); j++ )
2362 if( loop->_body.at(j) == n )
2363 break;
2364 if( j == loop->_body.size() ) { // Not found in loop body
2365 // Last ditch effort to avoid assertion: Its possible that we
2366 // have some users (so outcnt not zero) but are still dead.
2367 // Try to find from root.
2368 if (Compile::current()->root()->find(n->_idx)) {
2369 fail++;
2370 tty->print("We have that verify does not: ");
2371 n->dump();
2372 }
2373 }
2374 }
2375 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2376 Node *n = loop->_body.at(i2);
2377 if (n->outcnt() == 0) continue; // Ignore dead
2378 uint j;
2379 for( j = 0; j < _body.size(); j++ )
2380 if( _body.at(j) == n )
2381 break;
2382 if( j == _body.size() ) { // Not found in loop body
2383 // Last ditch effort to avoid assertion: Its possible that we
2384 // have some users (so outcnt not zero) but are still dead.
2385 // Try to find from root.
2386 if (Compile::current()->root()->find(n->_idx)) {
2387 fail++;
2388 tty->print("Verify has that we do not: ");
2389 n->dump();
2390 }
2391 }
2392 }
2393 assert( !fail, "loop body mismatch" );
2394 }
2395 }
2397 #endif
2399 //------------------------------set_idom---------------------------------------
2400 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2401 uint idx = d->_idx;
2402 if (idx >= _idom_size) {
2403 uint newsize = _idom_size<<1;
2404 while( idx >= newsize ) {
2405 newsize <<= 1;
2406 }
2407 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2408 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2409 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2410 _idom_size = newsize;
2411 }
2412 _idom[idx] = n;
2413 _dom_depth[idx] = dom_depth;
2414 }
2416 //------------------------------recompute_dom_depth---------------------------------------
2417 // The dominator tree is constructed with only parent pointers.
2418 // This recomputes the depth in the tree by first tagging all
2419 // nodes as "no depth yet" marker. The next pass then runs up
2420 // the dom tree from each node marked "no depth yet", and computes
2421 // the depth on the way back down.
2422 void PhaseIdealLoop::recompute_dom_depth() {
2423 uint no_depth_marker = C->unique();
2424 uint i;
2425 // Initialize depth to "no depth yet"
2426 for (i = 0; i < _idom_size; i++) {
2427 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2428 _dom_depth[i] = no_depth_marker;
2429 }
2430 }
2431 if (_dom_stk == NULL) {
2432 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
2433 if (init_size < 10) init_size = 10;
2434 _dom_stk = new GrowableArray<uint>(init_size);
2435 }
2436 // Compute new depth for each node.
2437 for (i = 0; i < _idom_size; i++) {
2438 uint j = i;
2439 // Run up the dom tree to find a node with a depth
2440 while (_dom_depth[j] == no_depth_marker) {
2441 _dom_stk->push(j);
2442 j = _idom[j]->_idx;
2443 }
2444 // Compute the depth on the way back down this tree branch
2445 uint dd = _dom_depth[j] + 1;
2446 while (_dom_stk->length() > 0) {
2447 uint j = _dom_stk->pop();
2448 _dom_depth[j] = dd;
2449 dd++;
2450 }
2451 }
2452 }
2454 //------------------------------sort-------------------------------------------
2455 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2456 // loop tree, not the root.
2457 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2458 if( !innermost ) return loop; // New innermost loop
2460 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2461 assert( loop_preorder, "not yet post-walked loop" );
2462 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2463 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2465 // Insert at start of list
2466 while( l ) { // Insertion sort based on pre-order
2467 if( l == loop ) return innermost; // Already on list!
2468 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2469 assert( l_preorder, "not yet post-walked l" );
2470 // Check header pre-order number to figure proper nesting
2471 if( loop_preorder > l_preorder )
2472 break; // End of insertion
2473 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2474 // Since I split shared headers, you'd think this could not happen.
2475 // BUT: I must first do the preorder numbering before I can discover I
2476 // have shared headers, so the split headers all get the same preorder
2477 // number as the RegionNode they split from.
2478 if( loop_preorder == l_preorder &&
2479 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2480 break; // Also check for shared headers (same pre#)
2481 pp = &l->_parent; // Chain up list
2482 l = *pp;
2483 }
2484 // Link into list
2485 // Point predecessor to me
2486 *pp = loop;
2487 // Point me to successor
2488 IdealLoopTree *p = loop->_parent;
2489 loop->_parent = l; // Point me to successor
2490 if( p ) sort( p, innermost ); // Insert my parents into list as well
2491 return innermost;
2492 }
2494 //------------------------------build_loop_tree--------------------------------
2495 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2496 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2497 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2498 // tightest enclosing IdealLoopTree for post-walked.
2499 //
2500 // During my forward walk I do a short 1-layer lookahead to see if I can find
2501 // a loop backedge with that doesn't have any work on the backedge. This
2502 // helps me construct nested loops with shared headers better.
2503 //
2504 // Once I've done the forward recursion, I do the post-work. For each child
2505 // I check to see if there is a backedge. Backedges define a loop! I
2506 // insert an IdealLoopTree at the target of the backedge.
2507 //
2508 // During the post-work I also check to see if I have several children
2509 // belonging to different loops. If so, then this Node is a decision point
2510 // where control flow can choose to change loop nests. It is at this
2511 // decision point where I can figure out how loops are nested. At this
2512 // time I can properly order the different loop nests from my children.
2513 // Note that there may not be any backedges at the decision point!
2514 //
2515 // Since the decision point can be far removed from the backedges, I can't
2516 // order my loops at the time I discover them. Thus at the decision point
2517 // I need to inspect loop header pre-order numbers to properly nest my
2518 // loops. This means I need to sort my childrens' loops by pre-order.
2519 // The sort is of size number-of-control-children, which generally limits
2520 // it to size 2 (i.e., I just choose between my 2 target loops).
2521 void PhaseIdealLoop::build_loop_tree() {
2522 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2523 GrowableArray <Node *> bltstack(C->unique() >> 1);
2524 Node *n = C->root();
2525 bltstack.push(n);
2526 int pre_order = 1;
2527 int stack_size;
2529 while ( ( stack_size = bltstack.length() ) != 0 ) {
2530 n = bltstack.top(); // Leave node on stack
2531 if ( !is_visited(n) ) {
2532 // ---- Pre-pass Work ----
2533 // Pre-walked but not post-walked nodes need a pre_order number.
2535 set_preorder_visited( n, pre_order ); // set as visited
2537 // ---- Scan over children ----
2538 // Scan first over control projections that lead to loop headers.
2539 // This helps us find inner-to-outer loops with shared headers better.
2541 // Scan children's children for loop headers.
2542 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2543 Node* m = n->raw_out(i); // Child
2544 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2545 // Scan over children's children to find loop
2546 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2547 Node* l = m->fast_out(j);
2548 if( is_visited(l) && // Been visited?
2549 !is_postvisited(l) && // But not post-visited
2550 get_preorder(l) < pre_order ) { // And smaller pre-order
2551 // Found! Scan the DFS down this path before doing other paths
2552 bltstack.push(m);
2553 break;
2554 }
2555 }
2556 }
2557 }
2558 pre_order++;
2559 }
2560 else if ( !is_postvisited(n) ) {
2561 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2562 // such as com.sun.rsasign.am::a.
2563 // For non-recursive version, first, process current children.
2564 // On next iteration, check if additional children were added.
2565 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2566 Node* u = n->raw_out(k);
2567 if ( u->is_CFG() && !is_visited(u) ) {
2568 bltstack.push(u);
2569 }
2570 }
2571 if ( bltstack.length() == stack_size ) {
2572 // There were no additional children, post visit node now
2573 (void)bltstack.pop(); // Remove node from stack
2574 pre_order = build_loop_tree_impl( n, pre_order );
2575 // Check for bailout
2576 if (C->failing()) {
2577 return;
2578 }
2579 // Check to grow _preorders[] array for the case when
2580 // build_loop_tree_impl() adds new nodes.
2581 check_grow_preorders();
2582 }
2583 }
2584 else {
2585 (void)bltstack.pop(); // Remove post-visited node from stack
2586 }
2587 }
2588 }
2590 //------------------------------build_loop_tree_impl---------------------------
2591 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2592 // ---- Post-pass Work ----
2593 // Pre-walked but not post-walked nodes need a pre_order number.
2595 // Tightest enclosing loop for this Node
2596 IdealLoopTree *innermost = NULL;
2598 // For all children, see if any edge is a backedge. If so, make a loop
2599 // for it. Then find the tightest enclosing loop for the self Node.
2600 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2601 Node* m = n->fast_out(i); // Child
2602 if( n == m ) continue; // Ignore control self-cycles
2603 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2605 IdealLoopTree *l; // Child's loop
2606 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2607 // Found a backedge
2608 assert( get_preorder(m) < pre_order, "should be backedge" );
2609 // Check for the RootNode, which is already a LoopNode and is allowed
2610 // to have multiple "backedges".
2611 if( m == C->root()) { // Found the root?
2612 l = _ltree_root; // Root is the outermost LoopNode
2613 } else { // Else found a nested loop
2614 // Insert a LoopNode to mark this loop.
2615 l = new IdealLoopTree(this, m, n);
2616 } // End of Else found a nested loop
2617 if( !has_loop(m) ) // If 'm' does not already have a loop set
2618 set_loop(m, l); // Set loop header to loop now
2620 } else { // Else not a nested loop
2621 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2622 l = get_loop(m); // Get previously determined loop
2623 // If successor is header of a loop (nest), move up-loop till it
2624 // is a member of some outer enclosing loop. Since there are no
2625 // shared headers (I've split them already) I only need to go up
2626 // at most 1 level.
2627 while( l && l->_head == m ) // Successor heads loop?
2628 l = l->_parent; // Move up 1 for me
2629 // If this loop is not properly parented, then this loop
2630 // has no exit path out, i.e. its an infinite loop.
2631 if( !l ) {
2632 // Make loop "reachable" from root so the CFG is reachable. Basically
2633 // insert a bogus loop exit that is never taken. 'm', the loop head,
2634 // points to 'n', one (of possibly many) fall-in paths. There may be
2635 // many backedges as well.
2637 // Here I set the loop to be the root loop. I could have, after
2638 // inserting a bogus loop exit, restarted the recursion and found my
2639 // new loop exit. This would make the infinite loop a first-class
2640 // loop and it would then get properly optimized. What's the use of
2641 // optimizing an infinite loop?
2642 l = _ltree_root; // Oops, found infinite loop
2644 if (!_verify_only) {
2645 // Insert the NeverBranch between 'm' and it's control user.
2646 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m );
2647 _igvn.register_new_node_with_optimizer(iff);
2648 set_loop(iff, l);
2649 Node *if_t = new (C, 1) CProjNode( iff, 0 );
2650 _igvn.register_new_node_with_optimizer(if_t);
2651 set_loop(if_t, l);
2653 Node* cfg = NULL; // Find the One True Control User of m
2654 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2655 Node* x = m->fast_out(j);
2656 if (x->is_CFG() && x != m && x != iff)
2657 { cfg = x; break; }
2658 }
2659 assert(cfg != NULL, "must find the control user of m");
2660 uint k = 0; // Probably cfg->in(0)
2661 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2662 cfg->set_req( k, if_t ); // Now point to NeverBranch
2664 // Now create the never-taken loop exit
2665 Node *if_f = new (C, 1) CProjNode( iff, 1 );
2666 _igvn.register_new_node_with_optimizer(if_f);
2667 set_loop(if_f, l);
2668 // Find frame ptr for Halt. Relies on the optimizer
2669 // V-N'ing. Easier and quicker than searching through
2670 // the program structure.
2671 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr );
2672 _igvn.register_new_node_with_optimizer(frame);
2673 // Halt & Catch Fire
2674 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame );
2675 _igvn.register_new_node_with_optimizer(halt);
2676 set_loop(halt, l);
2677 C->root()->add_req(halt);
2678 }
2679 set_loop(C->root(), _ltree_root);
2680 }
2681 }
2682 // Weeny check for irreducible. This child was already visited (this
2683 // IS the post-work phase). Is this child's loop header post-visited
2684 // as well? If so, then I found another entry into the loop.
2685 if (!_verify_only) {
2686 while( is_postvisited(l->_head) ) {
2687 // found irreducible
2688 l->_irreducible = 1; // = true
2689 l = l->_parent;
2690 _has_irreducible_loops = true;
2691 // Check for bad CFG here to prevent crash, and bailout of compile
2692 if (l == NULL) {
2693 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2694 return pre_order;
2695 }
2696 }
2697 }
2699 // This Node might be a decision point for loops. It is only if
2700 // it's children belong to several different loops. The sort call
2701 // does a trivial amount of work if there is only 1 child or all
2702 // children belong to the same loop. If however, the children
2703 // belong to different loops, the sort call will properly set the
2704 // _parent pointers to show how the loops nest.
2705 //
2706 // In any case, it returns the tightest enclosing loop.
2707 innermost = sort( l, innermost );
2708 }
2710 // Def-use info will have some dead stuff; dead stuff will have no
2711 // loop decided on.
2713 // Am I a loop header? If so fix up my parent's child and next ptrs.
2714 if( innermost && innermost->_head == n ) {
2715 assert( get_loop(n) == innermost, "" );
2716 IdealLoopTree *p = innermost->_parent;
2717 IdealLoopTree *l = innermost;
2718 while( p && l->_head == n ) {
2719 l->_next = p->_child; // Put self on parents 'next child'
2720 p->_child = l; // Make self as first child of parent
2721 l = p; // Now walk up the parent chain
2722 p = l->_parent;
2723 }
2724 } else {
2725 // Note that it is possible for a LoopNode to reach here, if the
2726 // backedge has been made unreachable (hence the LoopNode no longer
2727 // denotes a Loop, and will eventually be removed).
2729 // Record tightest enclosing loop for self. Mark as post-visited.
2730 set_loop(n, innermost);
2731 // Also record has_call flag early on
2732 if( innermost ) {
2733 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2734 // Do not count uncommon calls
2735 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2736 Node *iff = n->in(0)->in(0);
2737 if( !iff->is_If() ||
2738 (n->in(0)->Opcode() == Op_IfFalse &&
2739 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2740 (iff->as_If()->_prob >= 0.01) )
2741 innermost->_has_call = 1;
2742 }
2743 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2744 // Disable loop optimizations if the loop has a scalar replaceable
2745 // allocation. This disabling may cause a potential performance lost
2746 // if the allocation is not eliminated for some reason.
2747 innermost->_allow_optimizations = false;
2748 innermost->_has_call = 1; // = true
2749 }
2750 }
2751 }
2753 // Flag as post-visited now
2754 set_postvisited(n);
2755 return pre_order;
2756 }
2759 //------------------------------build_loop_early-------------------------------
2760 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2761 // First pass computes the earliest controlling node possible. This is the
2762 // controlling input with the deepest dominating depth.
2763 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2764 while (worklist.size() != 0) {
2765 // Use local variables nstack_top_n & nstack_top_i to cache values
2766 // on nstack's top.
2767 Node *nstack_top_n = worklist.pop();
2768 uint nstack_top_i = 0;
2769 //while_nstack_nonempty:
2770 while (true) {
2771 // Get parent node and next input's index from stack's top.
2772 Node *n = nstack_top_n;
2773 uint i = nstack_top_i;
2774 uint cnt = n->req(); // Count of inputs
2775 if (i == 0) { // Pre-process the node.
2776 if( has_node(n) && // Have either loop or control already?
2777 !has_ctrl(n) ) { // Have loop picked out already?
2778 // During "merge_many_backedges" we fold up several nested loops
2779 // into a single loop. This makes the members of the original
2780 // loop bodies pointing to dead loops; they need to move up
2781 // to the new UNION'd larger loop. I set the _head field of these
2782 // dead loops to NULL and the _parent field points to the owning
2783 // loop. Shades of UNION-FIND algorithm.
2784 IdealLoopTree *ilt;
2785 while( !(ilt = get_loop(n))->_head ) {
2786 // Normally I would use a set_loop here. But in this one special
2787 // case, it is legal (and expected) to change what loop a Node
2788 // belongs to.
2789 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2790 }
2791 // Remove safepoints ONLY if I've already seen I don't need one.
2792 // (the old code here would yank a 2nd safepoint after seeing a
2793 // first one, even though the 1st did not dominate in the loop body
2794 // and thus could be avoided indefinitely)
2795 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2796 is_deleteable_safept(n)) {
2797 Node *in = n->in(TypeFunc::Control);
2798 lazy_replace(n,in); // Pull safepoint now
2799 // Carry on with the recursion "as if" we are walking
2800 // only the control input
2801 if( !visited.test_set( in->_idx ) ) {
2802 worklist.push(in); // Visit this guy later, using worklist
2803 }
2804 // Get next node from nstack:
2805 // - skip n's inputs processing by setting i > cnt;
2806 // - we also will not call set_early_ctrl(n) since
2807 // has_node(n) == true (see the condition above).
2808 i = cnt + 1;
2809 }
2810 }
2811 } // if (i == 0)
2813 // Visit all inputs
2814 bool done = true; // Assume all n's inputs will be processed
2815 while (i < cnt) {
2816 Node *in = n->in(i);
2817 ++i;
2818 if (in == NULL) continue;
2819 if (in->pinned() && !in->is_CFG())
2820 set_ctrl(in, in->in(0));
2821 int is_visited = visited.test_set( in->_idx );
2822 if (!has_node(in)) { // No controlling input yet?
2823 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2824 assert( !is_visited, "visit only once" );
2825 nstack.push(n, i); // Save parent node and next input's index.
2826 nstack_top_n = in; // Process current input now.
2827 nstack_top_i = 0;
2828 done = false; // Not all n's inputs processed.
2829 break; // continue while_nstack_nonempty;
2830 } else if (!is_visited) {
2831 // This guy has a location picked out for him, but has not yet
2832 // been visited. Happens to all CFG nodes, for instance.
2833 // Visit him using the worklist instead of recursion, to break
2834 // cycles. Since he has a location already we do not need to
2835 // find his location before proceeding with the current Node.
2836 worklist.push(in); // Visit this guy later, using worklist
2837 }
2838 }
2839 if (done) {
2840 // All of n's inputs have been processed, complete post-processing.
2842 // Compute earliest point this Node can go.
2843 // CFG, Phi, pinned nodes already know their controlling input.
2844 if (!has_node(n)) {
2845 // Record earliest legal location
2846 set_early_ctrl( n );
2847 }
2848 if (nstack.is_empty()) {
2849 // Finished all nodes on stack.
2850 // Process next node on the worklist.
2851 break;
2852 }
2853 // Get saved parent node and next input's index.
2854 nstack_top_n = nstack.node();
2855 nstack_top_i = nstack.index();
2856 nstack.pop();
2857 }
2858 } // while (true)
2859 }
2860 }
2862 //------------------------------dom_lca_internal--------------------------------
2863 // Pair-wise LCA
2864 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2865 if( !n1 ) return n2; // Handle NULL original LCA
2866 assert( n1->is_CFG(), "" );
2867 assert( n2->is_CFG(), "" );
2868 // find LCA of all uses
2869 uint d1 = dom_depth(n1);
2870 uint d2 = dom_depth(n2);
2871 while (n1 != n2) {
2872 if (d1 > d2) {
2873 n1 = idom(n1);
2874 d1 = dom_depth(n1);
2875 } else if (d1 < d2) {
2876 n2 = idom(n2);
2877 d2 = dom_depth(n2);
2878 } else {
2879 // Here d1 == d2. Due to edits of the dominator-tree, sections
2880 // of the tree might have the same depth. These sections have
2881 // to be searched more carefully.
2883 // Scan up all the n1's with equal depth, looking for n2.
2884 Node *t1 = idom(n1);
2885 while (dom_depth(t1) == d1) {
2886 if (t1 == n2) return n2;
2887 t1 = idom(t1);
2888 }
2889 // Scan up all the n2's with equal depth, looking for n1.
2890 Node *t2 = idom(n2);
2891 while (dom_depth(t2) == d2) {
2892 if (t2 == n1) return n1;
2893 t2 = idom(t2);
2894 }
2895 // Move up to a new dominator-depth value as well as up the dom-tree.
2896 n1 = t1;
2897 n2 = t2;
2898 d1 = dom_depth(n1);
2899 d2 = dom_depth(n2);
2900 }
2901 }
2902 return n1;
2903 }
2905 //------------------------------compute_idom-----------------------------------
2906 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2907 // IDOMs are correct.
2908 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2909 assert( region->is_Region(), "" );
2910 Node *LCA = NULL;
2911 for( uint i = 1; i < region->req(); i++ ) {
2912 if( region->in(i) != C->top() )
2913 LCA = dom_lca( LCA, region->in(i) );
2914 }
2915 return LCA;
2916 }
2918 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
2919 bool had_error = false;
2920 #ifdef ASSERT
2921 if (early != C->root()) {
2922 // Make sure that there's a dominance path from use to LCA
2923 Node* d = use;
2924 while (d != LCA) {
2925 d = idom(d);
2926 if (d == C->root()) {
2927 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx);
2928 n->dump();
2929 use->dump();
2930 had_error = true;
2931 break;
2932 }
2933 }
2934 }
2935 #endif
2936 return had_error;
2937 }
2940 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
2941 // Compute LCA over list of uses
2942 bool had_error = false;
2943 Node *LCA = NULL;
2944 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2945 Node* c = n->fast_out(i);
2946 if (_nodes[c->_idx] == NULL)
2947 continue; // Skip the occasional dead node
2948 if( c->is_Phi() ) { // For Phis, we must land above on the path
2949 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2950 if( c->in(j) == n ) { // Found matching input?
2951 Node *use = c->in(0)->in(j);
2952 if (_verify_only && use->is_top()) continue;
2953 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2954 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2955 }
2956 }
2957 } else {
2958 // For CFG data-users, use is in the block just prior
2959 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
2960 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2961 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2962 }
2963 }
2964 assert(!had_error, "bad dominance");
2965 return LCA;
2966 }
2968 //------------------------------get_late_ctrl----------------------------------
2969 // Compute latest legal control.
2970 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
2971 assert(early != NULL, "early control should not be NULL");
2973 Node* LCA = compute_lca_of_uses(n, early);
2974 #ifdef ASSERT
2975 if (LCA == C->root() && LCA != early) {
2976 // def doesn't dominate uses so print some useful debugging output
2977 compute_lca_of_uses(n, early, true);
2978 }
2979 #endif
2981 // if this is a load, check for anti-dependent stores
2982 // We use a conservative algorithm to identify potential interfering
2983 // instructions and for rescheduling the load. The users of the memory
2984 // input of this load are examined. Any use which is not a load and is
2985 // dominated by early is considered a potentially interfering store.
2986 // This can produce false positives.
2987 if (n->is_Load() && LCA != early) {
2988 Node_List worklist;
2990 Node *mem = n->in(MemNode::Memory);
2991 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2992 Node* s = mem->fast_out(i);
2993 worklist.push(s);
2994 }
2995 while(worklist.size() != 0 && LCA != early) {
2996 Node* s = worklist.pop();
2997 if (s->is_Load()) {
2998 continue;
2999 } else if (s->is_MergeMem()) {
3000 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3001 Node* s1 = s->fast_out(i);
3002 worklist.push(s1);
3003 }
3004 } else {
3005 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3006 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3007 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3008 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3009 }
3010 }
3011 }
3012 }
3014 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3015 return LCA;
3016 }
3018 // true if CFG node d dominates CFG node n
3019 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3020 if (d == n)
3021 return true;
3022 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3023 uint dd = dom_depth(d);
3024 while (dom_depth(n) >= dd) {
3025 if (n == d)
3026 return true;
3027 n = idom(n);
3028 }
3029 return false;
3030 }
3032 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3033 // Pair-wise LCA with tags.
3034 // Tag each index with the node 'tag' currently being processed
3035 // before advancing up the dominator chain using idom().
3036 // Later calls that find a match to 'tag' know that this path has already
3037 // been considered in the current LCA (which is input 'n1' by convention).
3038 // Since get_late_ctrl() is only called once for each node, the tag array
3039 // does not need to be cleared between calls to get_late_ctrl().
3040 // Algorithm trades a larger constant factor for better asymptotic behavior
3041 //
3042 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3043 uint d1 = dom_depth(n1);
3044 uint d2 = dom_depth(n2);
3046 do {
3047 if (d1 > d2) {
3048 // current lca is deeper than n2
3049 _dom_lca_tags.map(n1->_idx, tag);
3050 n1 = idom(n1);
3051 d1 = dom_depth(n1);
3052 } else if (d1 < d2) {
3053 // n2 is deeper than current lca
3054 Node *memo = _dom_lca_tags[n2->_idx];
3055 if( memo == tag ) {
3056 return n1; // Return the current LCA
3057 }
3058 _dom_lca_tags.map(n2->_idx, tag);
3059 n2 = idom(n2);
3060 d2 = dom_depth(n2);
3061 } else {
3062 // Here d1 == d2. Due to edits of the dominator-tree, sections
3063 // of the tree might have the same depth. These sections have
3064 // to be searched more carefully.
3066 // Scan up all the n1's with equal depth, looking for n2.
3067 _dom_lca_tags.map(n1->_idx, tag);
3068 Node *t1 = idom(n1);
3069 while (dom_depth(t1) == d1) {
3070 if (t1 == n2) return n2;
3071 _dom_lca_tags.map(t1->_idx, tag);
3072 t1 = idom(t1);
3073 }
3074 // Scan up all the n2's with equal depth, looking for n1.
3075 _dom_lca_tags.map(n2->_idx, tag);
3076 Node *t2 = idom(n2);
3077 while (dom_depth(t2) == d2) {
3078 if (t2 == n1) return n1;
3079 _dom_lca_tags.map(t2->_idx, tag);
3080 t2 = idom(t2);
3081 }
3082 // Move up to a new dominator-depth value as well as up the dom-tree.
3083 n1 = t1;
3084 n2 = t2;
3085 d1 = dom_depth(n1);
3086 d2 = dom_depth(n2);
3087 }
3088 } while (n1 != n2);
3089 return n1;
3090 }
3092 //------------------------------init_dom_lca_tags------------------------------
3093 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3094 // Intended use does not involve any growth for the array, so it could
3095 // be of fixed size.
3096 void PhaseIdealLoop::init_dom_lca_tags() {
3097 uint limit = C->unique() + 1;
3098 _dom_lca_tags.map( limit, NULL );
3099 #ifdef ASSERT
3100 for( uint i = 0; i < limit; ++i ) {
3101 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3102 }
3103 #endif // ASSERT
3104 }
3106 //------------------------------clear_dom_lca_tags------------------------------
3107 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3108 // Intended use does not involve any growth for the array, so it could
3109 // be of fixed size.
3110 void PhaseIdealLoop::clear_dom_lca_tags() {
3111 uint limit = C->unique() + 1;
3112 _dom_lca_tags.map( limit, NULL );
3113 _dom_lca_tags.clear();
3114 #ifdef ASSERT
3115 for( uint i = 0; i < limit; ++i ) {
3116 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3117 }
3118 #endif // ASSERT
3119 }
3121 //------------------------------build_loop_late--------------------------------
3122 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3123 // Second pass finds latest legal placement, and ideal loop placement.
3124 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3125 while (worklist.size() != 0) {
3126 Node *n = worklist.pop();
3127 // Only visit once
3128 if (visited.test_set(n->_idx)) continue;
3129 uint cnt = n->outcnt();
3130 uint i = 0;
3131 while (true) {
3132 assert( _nodes[n->_idx], "no dead nodes" );
3133 // Visit all children
3134 if (i < cnt) {
3135 Node* use = n->raw_out(i);
3136 ++i;
3137 // Check for dead uses. Aggressively prune such junk. It might be
3138 // dead in the global sense, but still have local uses so I cannot
3139 // easily call 'remove_dead_node'.
3140 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3141 // Due to cycles, we might not hit the same fixed point in the verify
3142 // pass as we do in the regular pass. Instead, visit such phis as
3143 // simple uses of the loop head.
3144 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3145 if( !visited.test(use->_idx) )
3146 worklist.push(use);
3147 } else if( !visited.test_set(use->_idx) ) {
3148 nstack.push(n, i); // Save parent and next use's index.
3149 n = use; // Process all children of current use.
3150 cnt = use->outcnt();
3151 i = 0;
3152 }
3153 } else {
3154 // Do not visit around the backedge of loops via data edges.
3155 // push dead code onto a worklist
3156 _deadlist.push(use);
3157 }
3158 } else {
3159 // All of n's children have been processed, complete post-processing.
3160 build_loop_late_post(n);
3161 if (nstack.is_empty()) {
3162 // Finished all nodes on stack.
3163 // Process next node on the worklist.
3164 break;
3165 }
3166 // Get saved parent node and next use's index. Visit the rest of uses.
3167 n = nstack.node();
3168 cnt = n->outcnt();
3169 i = nstack.index();
3170 nstack.pop();
3171 }
3172 }
3173 }
3174 }
3176 //------------------------------build_loop_late_post---------------------------
3177 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3178 // Second pass finds latest legal placement, and ideal loop placement.
3179 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3181 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
3182 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3183 }
3185 // CFG and pinned nodes already handled
3186 if( n->in(0) ) {
3187 if( n->in(0)->is_top() ) return; // Dead?
3189 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3190 // _must_ be pinned (they have to observe their control edge of course).
3191 // Unlike Stores (which modify an unallocable resource, the memory
3192 // state), Mods/Loads can float around. So free them up.
3193 bool pinned = true;
3194 switch( n->Opcode() ) {
3195 case Op_DivI:
3196 case Op_DivF:
3197 case Op_DivD:
3198 case Op_ModI:
3199 case Op_ModF:
3200 case Op_ModD:
3201 case Op_LoadB: // Same with Loads; they can sink
3202 case Op_LoadUS: // during loop optimizations.
3203 case Op_LoadD:
3204 case Op_LoadF:
3205 case Op_LoadI:
3206 case Op_LoadKlass:
3207 case Op_LoadNKlass:
3208 case Op_LoadL:
3209 case Op_LoadS:
3210 case Op_LoadP:
3211 case Op_LoadN:
3212 case Op_LoadRange:
3213 case Op_LoadD_unaligned:
3214 case Op_LoadL_unaligned:
3215 case Op_StrComp: // Does a bunch of load-like effects
3216 case Op_StrEquals:
3217 case Op_StrIndexOf:
3218 case Op_AryEq:
3219 pinned = false;
3220 }
3221 if( pinned ) {
3222 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3223 if( !chosen_loop->_child ) // Inner loop?
3224 chosen_loop->_body.push(n); // Collect inner loops
3225 return;
3226 }
3227 } else { // No slot zero
3228 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3229 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3230 return;
3231 }
3232 assert(!n->is_CFG() || n->outcnt() == 0, "");
3233 }
3235 // Do I have a "safe range" I can select over?
3236 Node *early = get_ctrl(n);// Early location already computed
3238 // Compute latest point this Node can go
3239 Node *LCA = get_late_ctrl( n, early );
3240 // LCA is NULL due to uses being dead
3241 if( LCA == NULL ) {
3242 #ifdef ASSERT
3243 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3244 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3245 }
3246 #endif
3247 _nodes.map(n->_idx, 0); // This node is useless
3248 _deadlist.push(n);
3249 return;
3250 }
3251 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3253 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3254 Node *least = legal; // Best legal position so far
3255 while( early != legal ) { // While not at earliest legal
3256 #ifdef ASSERT
3257 if (legal->is_Start() && !early->is_Root()) {
3258 // Bad graph. Print idom path and fail.
3259 tty->print_cr( "Bad graph detected in build_loop_late");
3260 tty->print("n: ");n->dump(); tty->cr();
3261 tty->print("early: ");early->dump(); tty->cr();
3262 int ct = 0;
3263 Node *dbg_legal = LCA;
3264 while(!dbg_legal->is_Start() && ct < 100) {
3265 tty->print("idom[%d] ",ct); dbg_legal->dump(); tty->cr();
3266 ct++;
3267 dbg_legal = idom(dbg_legal);
3268 }
3269 assert(false, "Bad graph detected in build_loop_late");
3270 }
3271 #endif
3272 // Find least loop nesting depth
3273 legal = idom(legal); // Bump up the IDOM tree
3274 // Check for lower nesting depth
3275 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3276 least = legal;
3277 }
3278 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3280 // Try not to place code on a loop entry projection
3281 // which can inhibit range check elimination.
3282 if (least != early) {
3283 Node* ctrl_out = least->unique_ctrl_out();
3284 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3285 least == ctrl_out->in(LoopNode::EntryControl)) {
3286 Node* least_dom = idom(least);
3287 if (get_loop(least_dom)->is_member(get_loop(least))) {
3288 least = least_dom;
3289 }
3290 }
3291 }
3293 #ifdef ASSERT
3294 // If verifying, verify that 'verify_me' has a legal location
3295 // and choose it as our location.
3296 if( _verify_me ) {
3297 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3298 Node *legal = LCA;
3299 while( early != legal ) { // While not at earliest legal
3300 if( legal == v_ctrl ) break; // Check for prior good location
3301 legal = idom(legal) ;// Bump up the IDOM tree
3302 }
3303 // Check for prior good location
3304 if( legal == v_ctrl ) least = legal; // Keep prior if found
3305 }
3306 #endif
3308 // Assign discovered "here or above" point
3309 least = find_non_split_ctrl(least);
3310 set_ctrl(n, least);
3312 // Collect inner loop bodies
3313 IdealLoopTree *chosen_loop = get_loop(least);
3314 if( !chosen_loop->_child ) // Inner loop?
3315 chosen_loop->_body.push(n);// Collect inner loops
3316 }
3318 #ifndef PRODUCT
3319 //------------------------------dump-------------------------------------------
3320 void PhaseIdealLoop::dump( ) const {
3321 ResourceMark rm;
3322 Arena* arena = Thread::current()->resource_area();
3323 Node_Stack stack(arena, C->unique() >> 2);
3324 Node_List rpo_list;
3325 VectorSet visited(arena);
3326 visited.set(C->top()->_idx);
3327 rpo( C->root(), stack, visited, rpo_list );
3328 // Dump root loop indexed by last element in PO order
3329 dump( _ltree_root, rpo_list.size(), rpo_list );
3330 }
3332 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3333 loop->dump_head();
3335 // Now scan for CFG nodes in the same loop
3336 for( uint j=idx; j > 0; j-- ) {
3337 Node *n = rpo_list[j-1];
3338 if( !_nodes[n->_idx] ) // Skip dead nodes
3339 continue;
3340 if( get_loop(n) != loop ) { // Wrong loop nest
3341 if( get_loop(n)->_head == n && // Found nested loop?
3342 get_loop(n)->_parent == loop )
3343 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3344 continue;
3345 }
3347 // Dump controlling node
3348 for( uint x = 0; x < loop->_nest; x++ )
3349 tty->print(" ");
3350 tty->print("C");
3351 if( n == C->root() ) {
3352 n->dump();
3353 } else {
3354 Node* cached_idom = idom_no_update(n);
3355 Node *computed_idom = n->in(0);
3356 if( n->is_Region() ) {
3357 computed_idom = compute_idom(n);
3358 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3359 // any MultiBranch ctrl node), so apply a similar transform to
3360 // the cached idom returned from idom_no_update.
3361 cached_idom = find_non_split_ctrl(cached_idom);
3362 }
3363 tty->print(" ID:%d",computed_idom->_idx);
3364 n->dump();
3365 if( cached_idom != computed_idom ) {
3366 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3367 computed_idom->_idx, cached_idom->_idx);
3368 }
3369 }
3370 // Dump nodes it controls
3371 for( uint k = 0; k < _nodes.Size(); k++ ) {
3372 // (k < C->unique() && get_ctrl(find(k)) == n)
3373 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3374 Node *m = C->root()->find(k);
3375 if( m && m->outcnt() > 0 ) {
3376 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3377 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3378 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3379 }
3380 for( uint j = 0; j < loop->_nest; j++ )
3381 tty->print(" ");
3382 tty->print(" ");
3383 m->dump();
3384 }
3385 }
3386 }
3387 }
3388 }
3390 // Collect a R-P-O for the whole CFG.
3391 // Result list is in post-order (scan backwards for RPO)
3392 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3393 stk.push(start, 0);
3394 visited.set(start->_idx);
3396 while (stk.is_nonempty()) {
3397 Node* m = stk.node();
3398 uint idx = stk.index();
3399 if (idx < m->outcnt()) {
3400 stk.set_index(idx + 1);
3401 Node* n = m->raw_out(idx);
3402 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3403 stk.push(n, 0);
3404 }
3405 } else {
3406 rpo_list.push(m);
3407 stk.pop();
3408 }
3409 }
3410 }
3411 #endif
3414 //=============================================================================
3415 //------------------------------LoopTreeIterator-----------------------------------
3417 // Advance to next loop tree using a preorder, left-to-right traversal.
3418 void LoopTreeIterator::next() {
3419 assert(!done(), "must not be done.");
3420 if (_curnt->_child != NULL) {
3421 _curnt = _curnt->_child;
3422 } else if (_curnt->_next != NULL) {
3423 _curnt = _curnt->_next;
3424 } else {
3425 while (_curnt != _root && _curnt->_next == NULL) {
3426 _curnt = _curnt->_parent;
3427 }
3428 if (_curnt == _root) {
3429 _curnt = NULL;
3430 assert(done(), "must be done.");
3431 } else {
3432 assert(_curnt->_next != NULL, "must be more to do");
3433 _curnt = _curnt->_next;
3434 }
3435 }
3436 }