Fri, 29 Jul 2011 09:16:29 -0700
7068051: SIGSEGV in PhaseIdealLoop::build_loop_late_post
Summary: Removed predicate cloning from loop peeling optimization and from split fall-in paths.
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 // 'ne' can be replaced with 'lt' only when init < limit.
457 if (stride_con > 0 && init_t->_hi < limit_t->_lo)
458 bt = BoolTest::lt;
459 // 'ne' can be replaced with 'gt' only when init > limit.
460 if (stride_con < 0 && init_t->_lo > limit_t->_hi)
461 bt = BoolTest::gt;
462 }
464 if (incl_limit) {
465 // The limit check guaranties that 'limit <= (max_jint - stride)' so
466 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
467 //
468 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
469 limit = gvn->transform(new (C, 3) AddINode(limit, one));
470 if (bt == BoolTest::le)
471 bt = BoolTest::lt;
472 else if (bt == BoolTest::ge)
473 bt = BoolTest::gt;
474 else
475 ShouldNotReachHere();
476 }
477 set_subtree_ctrl( limit );
479 } else { // LoopLimitCheck
481 // If compare points to incr, we are ok. Otherwise the compare
482 // can directly point to the phi; in this case adjust the compare so that
483 // it points to the incr by adjusting the limit.
484 if (cmp->in(1) == phi || cmp->in(2) == phi)
485 limit = gvn->transform(new (C, 3) AddINode(limit,stride));
487 // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
488 // Final value for iterator should be: trip_count * stride + init_trip.
489 Node *one_p = gvn->intcon( 1);
490 Node *one_m = gvn->intcon(-1);
492 Node *trip_count = NULL;
493 switch( bt ) {
494 case BoolTest::eq:
495 ShouldNotReachHere();
496 case BoolTest::ne: // Ahh, the case we desire
497 if (stride_con == 1)
498 trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip));
499 else if (stride_con == -1)
500 trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit));
501 else
502 ShouldNotReachHere();
503 set_subtree_ctrl(trip_count);
504 //_loop.map(trip_count->_idx,loop(limit));
505 break;
506 case BoolTest::le: // Maybe convert to '<' case
507 limit = gvn->transform(new (C, 3) AddINode(limit,one_p));
508 set_subtree_ctrl( limit );
509 hook->init_req(4, limit);
511 bt = BoolTest::lt;
512 // Make the new limit be in the same loop nest as the old limit
513 //_loop.map(limit->_idx,limit_loop);
514 // Fall into next case
515 case BoolTest::lt: { // Maybe convert to '!=' case
516 if (stride_con < 0) // Count down loop rolls through MAXINT
517 ShouldNotReachHere();
518 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
519 set_subtree_ctrl( range );
520 hook->init_req(0, range);
522 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
523 set_subtree_ctrl( bias );
524 hook->init_req(1, bias);
526 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m));
527 set_subtree_ctrl( bias1 );
528 hook->init_req(2, bias1);
530 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
531 set_subtree_ctrl( trip_count );
532 hook->init_req(3, trip_count);
533 break;
534 }
536 case BoolTest::ge: // Maybe convert to '>' case
537 limit = gvn->transform(new (C, 3) AddINode(limit,one_m));
538 set_subtree_ctrl( limit );
539 hook->init_req(4 ,limit);
541 bt = BoolTest::gt;
542 // Make the new limit be in the same loop nest as the old limit
543 //_loop.map(limit->_idx,limit_loop);
544 // Fall into next case
545 case BoolTest::gt: { // Maybe convert to '!=' case
546 if (stride_con > 0) // count up loop rolls through MININT
547 ShouldNotReachHere();
548 Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip));
549 set_subtree_ctrl( range );
550 hook->init_req(0, range);
552 Node *bias = gvn->transform(new (C, 3) AddINode(range,stride));
553 set_subtree_ctrl( bias );
554 hook->init_req(1, bias);
556 Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p));
557 set_subtree_ctrl( bias1 );
558 hook->init_req(2, bias1);
560 trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride));
561 set_subtree_ctrl( trip_count );
562 hook->init_req(3, trip_count);
563 break;
564 }
565 } // switch( bt )
567 Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride));
568 set_subtree_ctrl( span );
569 hook->init_req(5, span);
571 limit = gvn->transform(new (C, 3) AddINode(span,init_trip));
572 set_subtree_ctrl( limit );
574 } // LoopLimitCheck
576 // Check for SafePoint on backedge and remove
577 Node *sfpt = x->in(LoopNode::LoopBackControl);
578 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
579 lazy_replace( sfpt, iftrue );
580 loop->_tail = iftrue;
581 }
583 // Build a canonical trip test.
584 // Clone code, as old values may be in use.
585 Node* nphi = PhiNode::make(x, init_trip, TypeInt::INT);
586 nphi = _igvn.register_new_node_with_optimizer(nphi);
587 set_ctrl(nphi, get_ctrl(phi));
589 incr = incr->clone();
590 incr->set_req(1,nphi);
591 incr->set_req(2,stride);
592 incr = _igvn.register_new_node_with_optimizer(incr);
593 set_early_ctrl( incr );
595 nphi->set_req(LoopNode::LoopBackControl, incr);
596 _igvn.replace_node(phi, nphi);
597 phi = nphi->as_Phi();
599 cmp = cmp->clone();
600 cmp->set_req(1,incr);
601 cmp->set_req(2,limit);
602 cmp = _igvn.register_new_node_with_optimizer(cmp);
603 set_ctrl(cmp, iff->in(0));
605 test = test->clone()->as_Bool();
606 (*(BoolTest*)&test->_test)._test = bt;
607 test->set_req(1,cmp);
608 _igvn.register_new_node_with_optimizer(test);
609 set_ctrl(test, iff->in(0));
611 // Replace the old IfNode with a new LoopEndNode
612 Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
613 IfNode *le = lex->as_If();
614 uint dd = dom_depth(iff);
615 set_idom(le, le->in(0), dd); // Update dominance for loop exit
616 set_loop(le, loop);
618 // Get the loop-exit control
619 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
621 // Need to swap loop-exit and loop-back control?
622 if (iftrue_op == Op_IfFalse) {
623 Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le));
624 Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le));
626 loop->_tail = back_control = ift2;
627 set_loop(ift2, loop);
628 set_loop(iff2, get_loop(iffalse));
630 // Lazy update of 'get_ctrl' mechanism.
631 lazy_replace_proj( iffalse, iff2 );
632 lazy_replace_proj( iftrue, ift2 );
634 // Swap names
635 iffalse = iff2;
636 iftrue = ift2;
637 } else {
638 _igvn.hash_delete(iffalse);
639 _igvn.hash_delete(iftrue);
640 iffalse->set_req_X( 0, le, &_igvn );
641 iftrue ->set_req_X( 0, le, &_igvn );
642 }
644 set_idom(iftrue, le, dd+1);
645 set_idom(iffalse, le, dd+1);
646 assert(iff->outcnt() == 0, "should be dead now");
647 lazy_replace( iff, le ); // fix 'get_ctrl'
649 // Now setup a new CountedLoopNode to replace the existing LoopNode
650 CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control);
651 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
652 // The following assert is approximately true, and defines the intention
653 // of can_be_counted_loop. It fails, however, because phase->type
654 // is not yet initialized for this loop and its parts.
655 //assert(l->can_be_counted_loop(this), "sanity");
656 _igvn.register_new_node_with_optimizer(l);
657 set_loop(l, loop);
658 loop->_head = l;
659 // Fix all data nodes placed at the old loop head.
660 // Uses the lazy-update mechanism of 'get_ctrl'.
661 lazy_replace( x, l );
662 set_idom(l, init_control, dom_depth(x));
664 // Check for immediately preceding SafePoint and remove
665 Node *sfpt2 = le->in(0);
666 if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2))
667 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
669 // Free up intermediate goo
670 _igvn.remove_dead_node(hook);
672 #ifdef ASSERT
673 assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
674 assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
675 #endif
676 #ifndef PRODUCT
677 if (TraceLoopOpts) {
678 tty->print("Counted ");
679 loop->dump_head();
680 }
681 #endif
683 C->print_method("After CountedLoop", 3);
685 return true;
686 }
688 //----------------------exact_limit-------------------------------------------
689 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
690 assert(loop->_head->is_CountedLoop(), "");
691 CountedLoopNode *cl = loop->_head->as_CountedLoop();
693 if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
694 cl->limit()->Opcode() == Op_LoopLimit) {
695 // Old code has exact limit (it could be incorrect in case of int overflow).
696 // Loop limit is exact with stride == 1. And loop may already have exact limit.
697 return cl->limit();
698 }
699 Node *limit = NULL;
700 #ifdef ASSERT
701 BoolTest::mask bt = cl->loopexit()->test_trip();
702 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
703 #endif
704 if (cl->has_exact_trip_count()) {
705 // Simple case: loop has constant boundaries.
706 // Use longs to avoid integer overflow.
707 int stride_con = cl->stride_con();
708 long init_con = cl->init_trip()->get_int();
709 long limit_con = cl->limit()->get_int();
710 julong trip_cnt = cl->trip_count();
711 long final_con = init_con + trip_cnt*stride_con;
712 final_con -= stride_con;
713 int final_int = (int)final_con;
714 // The final value should be in integer range since the loop
715 // is counted and the limit was checked for overflow.
716 assert(final_con == (long)final_int, "final value should be integer");
717 limit = _igvn.intcon(final_int);
718 } else {
719 // Create new LoopLimit node to get exact limit (final iv value).
720 limit = new (C, 4) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
721 register_new_node(limit, cl->in(LoopNode::EntryControl));
722 }
723 assert(limit != NULL, "sanity");
724 return limit;
725 }
727 //------------------------------Ideal------------------------------------------
728 // Return a node which is more "ideal" than the current node.
729 // Attempt to convert into a counted-loop.
730 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
731 if (!can_be_counted_loop(phase)) {
732 phase->C->set_major_progress();
733 }
734 return RegionNode::Ideal(phase, can_reshape);
735 }
738 //=============================================================================
739 //------------------------------Ideal------------------------------------------
740 // Return a node which is more "ideal" than the current node.
741 // Attempt to convert into a counted-loop.
742 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
743 return RegionNode::Ideal(phase, can_reshape);
744 }
746 //------------------------------dump_spec--------------------------------------
747 // Dump special per-node info
748 #ifndef PRODUCT
749 void CountedLoopNode::dump_spec(outputStream *st) const {
750 LoopNode::dump_spec(st);
751 if (stride_is_con()) {
752 st->print("stride: %d ",stride_con());
753 }
754 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
755 if (is_main_loop()) st->print("main of N%d", _idx);
756 if (is_post_loop()) st->print("post of N%d", _main_idx);
757 }
758 #endif
760 //=============================================================================
761 int CountedLoopEndNode::stride_con() const {
762 return stride()->bottom_type()->is_int()->get_con();
763 }
765 //=============================================================================
766 //------------------------------Value-----------------------------------------
767 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {
768 const Type* init_t = phase->type(in(Init));
769 const Type* limit_t = phase->type(in(Limit));
770 const Type* stride_t = phase->type(in(Stride));
771 // Either input is TOP ==> the result is TOP
772 if (init_t == Type::TOP) return Type::TOP;
773 if (limit_t == Type::TOP) return Type::TOP;
774 if (stride_t == Type::TOP) return Type::TOP;
776 int stride_con = stride_t->is_int()->get_con();
777 if (stride_con == 1)
778 return NULL; // Identity
780 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
781 // Use longs to avoid integer overflow.
782 long init_con = init_t->is_int()->get_con();
783 long limit_con = limit_t->is_int()->get_con();
784 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
785 long trip_count = (limit_con - init_con + stride_m)/stride_con;
786 long final_con = init_con + stride_con*trip_count;
787 int final_int = (int)final_con;
788 // The final value should be in integer range since the loop
789 // is counted and the limit was checked for overflow.
790 assert(final_con == (long)final_int, "final value should be integer");
791 return TypeInt::make(final_int);
792 }
794 return bottom_type(); // TypeInt::INT
795 }
797 //------------------------------Ideal------------------------------------------
798 // Return a node which is more "ideal" than the current node.
799 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
800 if (phase->type(in(Init)) == Type::TOP ||
801 phase->type(in(Limit)) == Type::TOP ||
802 phase->type(in(Stride)) == Type::TOP)
803 return NULL; // Dead
805 int stride_con = phase->type(in(Stride))->is_int()->get_con();
806 if (stride_con == 1)
807 return NULL; // Identity
809 if (in(Init)->is_Con() && in(Limit)->is_Con())
810 return NULL; // Value
812 // Delay following optimizations until all loop optimizations
813 // done to keep Ideal graph simple.
814 if (!can_reshape || phase->C->major_progress())
815 return NULL;
817 const TypeInt* init_t = phase->type(in(Init) )->is_int();
818 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
819 int stride_p;
820 long lim, ini;
821 julong max;
822 if (stride_con > 0) {
823 stride_p = stride_con;
824 lim = limit_t->_hi;
825 ini = init_t->_lo;
826 max = (julong)max_jint;
827 } else {
828 stride_p = -stride_con;
829 lim = init_t->_hi;
830 ini = limit_t->_lo;
831 max = (julong)min_jint;
832 }
833 julong range = lim - ini + stride_p;
834 if (range <= max) {
835 // Convert to integer expression if it is not overflow.
836 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
837 Node *range = phase->transform(new (phase->C, 3) SubINode(in(Limit), in(Init)));
838 Node *bias = phase->transform(new (phase->C, 3) AddINode(range, stride_m));
839 Node *trip = phase->transform(new (phase->C, 3) DivINode(0, bias, in(Stride)));
840 Node *span = phase->transform(new (phase->C, 3) MulINode(trip, in(Stride)));
841 return new (phase->C, 3) AddINode(span, in(Init)); // exact limit
842 }
844 if (is_power_of_2(stride_p) || // divisor is 2^n
845 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
846 // Convert to long expression to avoid integer overflow
847 // and let igvn optimizer convert this division.
848 //
849 Node* init = phase->transform( new (phase->C, 2) ConvI2LNode(in(Init)));
850 Node* limit = phase->transform( new (phase->C, 2) ConvI2LNode(in(Limit)));
851 Node* stride = phase->longcon(stride_con);
852 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
854 Node *range = phase->transform(new (phase->C, 3) SubLNode(limit, init));
855 Node *bias = phase->transform(new (phase->C, 3) AddLNode(range, stride_m));
856 Node *span;
857 if (stride_con > 0 && is_power_of_2(stride_p)) {
858 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
859 // and avoid generating rounding for division. Zero trip guard should
860 // guarantee that init < limit but sometimes the guard is missing and
861 // we can get situation when init > limit. Note, for the empty loop
862 // optimization zero trip guard is generated explicitly which leaves
863 // only RCE predicate where exact limit is used and the predicate
864 // will simply fail forcing recompilation.
865 Node* neg_stride = phase->longcon(-stride_con);
866 span = phase->transform(new (phase->C, 3) AndLNode(bias, neg_stride));
867 } else {
868 Node *trip = phase->transform(new (phase->C, 3) DivLNode(0, bias, stride));
869 span = phase->transform(new (phase->C, 3) MulLNode(trip, stride));
870 }
871 // Convert back to int
872 Node *span_int = phase->transform(new (phase->C, 2) ConvL2INode(span));
873 return new (phase->C, 3) AddINode(span_int, in(Init)); // exact limit
874 }
876 return NULL; // No progress
877 }
879 //------------------------------Identity---------------------------------------
880 // If stride == 1 return limit node.
881 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {
882 int stride_con = phase->type(in(Stride))->is_int()->get_con();
883 if (stride_con == 1 || stride_con == -1)
884 return in(Limit);
885 return this;
886 }
888 //=============================================================================
889 //----------------------match_incr_with_optional_truncation--------------------
890 // Match increment with optional truncation:
891 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
892 // Return NULL for failure. Success returns the increment node.
893 Node* CountedLoopNode::match_incr_with_optional_truncation(
894 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
895 // Quick cutouts:
896 if (expr == NULL || expr->req() != 3) return false;
898 Node *t1 = NULL;
899 Node *t2 = NULL;
900 const TypeInt* trunc_t = TypeInt::INT;
901 Node* n1 = expr;
902 int n1op = n1->Opcode();
904 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
905 if (n1op == Op_AndI &&
906 n1->in(2)->is_Con() &&
907 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
908 // %%% This check should match any mask of 2**K-1.
909 t1 = n1;
910 n1 = t1->in(1);
911 n1op = n1->Opcode();
912 trunc_t = TypeInt::CHAR;
913 } else if (n1op == Op_RShiftI &&
914 n1->in(1) != NULL &&
915 n1->in(1)->Opcode() == Op_LShiftI &&
916 n1->in(2) == n1->in(1)->in(2) &&
917 n1->in(2)->is_Con()) {
918 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
919 // %%% This check should match any shift in [1..31].
920 if (shift == 16 || shift == 8) {
921 t1 = n1;
922 t2 = t1->in(1);
923 n1 = t2->in(1);
924 n1op = n1->Opcode();
925 if (shift == 16) {
926 trunc_t = TypeInt::SHORT;
927 } else if (shift == 8) {
928 trunc_t = TypeInt::BYTE;
929 }
930 }
931 }
933 // If (maybe after stripping) it is an AddI, we won:
934 if (n1op == Op_AddI) {
935 *trunc1 = t1;
936 *trunc2 = t2;
937 *trunc_type = trunc_t;
938 return n1;
939 }
941 // failed
942 return NULL;
943 }
946 //------------------------------filtered_type--------------------------------
947 // Return a type based on condition control flow
948 // A successful return will be a type that is restricted due
949 // to a series of dominating if-tests, such as:
950 // if (i < 10) {
951 // if (i > 0) {
952 // here: "i" type is [1..10)
953 // }
954 // }
955 // or a control flow merge
956 // if (i < 10) {
957 // do {
958 // phi( , ) -- at top of loop type is [min_int..10)
959 // i = ?
960 // } while ( i < 10)
961 //
962 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
963 assert(n && n->bottom_type()->is_int(), "must be int");
964 const TypeInt* filtered_t = NULL;
965 if (!n->is_Phi()) {
966 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
967 filtered_t = filtered_type_from_dominators(n, n_ctrl);
969 } else {
970 Node* phi = n->as_Phi();
971 Node* region = phi->in(0);
972 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
973 if (region && region != C->top()) {
974 for (uint i = 1; i < phi->req(); i++) {
975 Node* val = phi->in(i);
976 Node* use_c = region->in(i);
977 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
978 if (val_t != NULL) {
979 if (filtered_t == NULL) {
980 filtered_t = val_t;
981 } else {
982 filtered_t = filtered_t->meet(val_t)->is_int();
983 }
984 }
985 }
986 }
987 }
988 const TypeInt* n_t = _igvn.type(n)->is_int();
989 if (filtered_t != NULL) {
990 n_t = n_t->join(filtered_t)->is_int();
991 }
992 return n_t;
993 }
996 //------------------------------filtered_type_from_dominators--------------------------------
997 // Return a possibly more restrictive type for val based on condition control flow of dominators
998 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
999 if (val->is_Con()) {
1000 return val->bottom_type()->is_int();
1001 }
1002 uint if_limit = 10; // Max number of dominating if's visited
1003 const TypeInt* rtn_t = NULL;
1005 if (use_ctrl && use_ctrl != C->top()) {
1006 Node* val_ctrl = get_ctrl(val);
1007 uint val_dom_depth = dom_depth(val_ctrl);
1008 Node* pred = use_ctrl;
1009 uint if_cnt = 0;
1010 while (if_cnt < if_limit) {
1011 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1012 if_cnt++;
1013 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1014 if (if_t != NULL) {
1015 if (rtn_t == NULL) {
1016 rtn_t = if_t;
1017 } else {
1018 rtn_t = rtn_t->join(if_t)->is_int();
1019 }
1020 }
1021 }
1022 pred = idom(pred);
1023 if (pred == NULL || pred == C->top()) {
1024 break;
1025 }
1026 // Stop if going beyond definition block of val
1027 if (dom_depth(pred) < val_dom_depth) {
1028 break;
1029 }
1030 }
1031 }
1032 return rtn_t;
1033 }
1036 //------------------------------dump_spec--------------------------------------
1037 // Dump special per-node info
1038 #ifndef PRODUCT
1039 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1040 if( in(TestValue)->is_Bool() ) {
1041 BoolTest bt( test_trip()); // Added this for g++.
1043 st->print("[");
1044 bt.dump_on(st);
1045 st->print("]");
1046 }
1047 st->print(" ");
1048 IfNode::dump_spec(st);
1049 }
1050 #endif
1052 //=============================================================================
1053 //------------------------------is_member--------------------------------------
1054 // Is 'l' a member of 'this'?
1055 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {
1056 while( l->_nest > _nest ) l = l->_parent;
1057 return l == this;
1058 }
1060 //------------------------------set_nest---------------------------------------
1061 // Set loop tree nesting depth. Accumulate _has_call bits.
1062 int IdealLoopTree::set_nest( uint depth ) {
1063 _nest = depth;
1064 int bits = _has_call;
1065 if( _child ) bits |= _child->set_nest(depth+1);
1066 if( bits ) _has_call = 1;
1067 if( _next ) bits |= _next ->set_nest(depth );
1068 return bits;
1069 }
1071 //------------------------------split_fall_in----------------------------------
1072 // Split out multiple fall-in edges from the loop header. Move them to a
1073 // private RegionNode before the loop. This becomes the loop landing pad.
1074 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1075 PhaseIterGVN &igvn = phase->_igvn;
1076 uint i;
1078 // Make a new RegionNode to be the landing pad.
1079 Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 );
1080 phase->set_loop(landing_pad,_parent);
1081 // Gather all the fall-in control paths into the landing pad
1082 uint icnt = fall_in_cnt;
1083 uint oreq = _head->req();
1084 for( i = oreq-1; i>0; i-- )
1085 if( !phase->is_member( this, _head->in(i) ) )
1086 landing_pad->set_req(icnt--,_head->in(i));
1088 // Peel off PhiNode edges as well
1089 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1090 Node *oj = _head->fast_out(j);
1091 if( oj->is_Phi() ) {
1092 PhiNode* old_phi = oj->as_Phi();
1093 assert( old_phi->region() == _head, "" );
1094 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
1095 Node *p = PhiNode::make_blank(landing_pad, old_phi);
1096 uint icnt = fall_in_cnt;
1097 for( i = oreq-1; i>0; i-- ) {
1098 if( !phase->is_member( this, _head->in(i) ) ) {
1099 p->init_req(icnt--, old_phi->in(i));
1100 // Go ahead and clean out old edges from old phi
1101 old_phi->del_req(i);
1102 }
1103 }
1104 // Search for CSE's here, because ZKM.jar does a lot of
1105 // loop hackery and we need to be a little incremental
1106 // with the CSE to avoid O(N^2) node blow-up.
1107 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1108 if( p2 ) { // Found CSE
1109 p->destruct(); // Recover useless new node
1110 p = p2; // Use old node
1111 } else {
1112 igvn.register_new_node_with_optimizer(p, old_phi);
1113 }
1114 // Make old Phi refer to new Phi.
1115 old_phi->add_req(p);
1116 // Check for the special case of making the old phi useless and
1117 // disappear it. In JavaGrande I have a case where this useless
1118 // Phi is the loop limit and prevents recognizing a CountedLoop
1119 // which in turn prevents removing an empty loop.
1120 Node *id_old_phi = old_phi->Identity( &igvn );
1121 if( id_old_phi != old_phi ) { // Found a simple identity?
1122 // Note that I cannot call 'replace_node' here, because
1123 // that will yank the edge from old_phi to the Region and
1124 // I'm mid-iteration over the Region's uses.
1125 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1126 Node* use = old_phi->last_out(i);
1127 igvn.hash_delete(use);
1128 igvn._worklist.push(use);
1129 uint uses_found = 0;
1130 for (uint j = 0; j < use->len(); j++) {
1131 if (use->in(j) == old_phi) {
1132 if (j < use->req()) use->set_req (j, id_old_phi);
1133 else use->set_prec(j, id_old_phi);
1134 uses_found++;
1135 }
1136 }
1137 i -= uses_found; // we deleted 1 or more copies of this edge
1138 }
1139 }
1140 igvn._worklist.push(old_phi);
1141 }
1142 }
1143 // Finally clean out the fall-in edges from the RegionNode
1144 for( i = oreq-1; i>0; i-- ) {
1145 if( !phase->is_member( this, _head->in(i) ) ) {
1146 _head->del_req(i);
1147 }
1148 }
1149 // Transform landing pad
1150 igvn.register_new_node_with_optimizer(landing_pad, _head);
1151 // Insert landing pad into the header
1152 _head->add_req(landing_pad);
1153 }
1155 //------------------------------split_outer_loop-------------------------------
1156 // Split out the outermost loop from this shared header.
1157 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1158 PhaseIterGVN &igvn = phase->_igvn;
1160 // Find index of outermost loop; it should also be my tail.
1161 uint outer_idx = 1;
1162 while( _head->in(outer_idx) != _tail ) outer_idx++;
1164 // Make a LoopNode for the outermost loop.
1165 Node *ctl = _head->in(LoopNode::EntryControl);
1166 Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) );
1167 outer = igvn.register_new_node_with_optimizer(outer, _head);
1168 phase->set_created_loop_node();
1170 // Outermost loop falls into '_head' loop
1171 _head->set_req(LoopNode::EntryControl, outer);
1172 _head->del_req(outer_idx);
1173 // Split all the Phis up between '_head' loop and 'outer' loop.
1174 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1175 Node *out = _head->fast_out(j);
1176 if( out->is_Phi() ) {
1177 PhiNode *old_phi = out->as_Phi();
1178 assert( old_phi->region() == _head, "" );
1179 Node *phi = PhiNode::make_blank(outer, old_phi);
1180 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
1181 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1182 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1183 // Make old Phi point to new Phi on the fall-in path
1184 igvn.hash_delete(old_phi);
1185 old_phi->set_req(LoopNode::EntryControl, phi);
1186 old_phi->del_req(outer_idx);
1187 igvn._worklist.push(old_phi);
1188 }
1189 }
1191 // Use the new loop head instead of the old shared one
1192 _head = outer;
1193 phase->set_loop(_head, this);
1194 }
1196 //------------------------------fix_parent-------------------------------------
1197 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1198 loop->_parent = parent;
1199 if( loop->_child ) fix_parent( loop->_child, loop );
1200 if( loop->_next ) fix_parent( loop->_next , parent );
1201 }
1203 //------------------------------estimate_path_freq-----------------------------
1204 static float estimate_path_freq( Node *n ) {
1205 // Try to extract some path frequency info
1206 IfNode *iff;
1207 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1208 uint nop = n->Opcode();
1209 if( nop == Op_SafePoint ) { // Skip any safepoint
1210 n = n->in(0);
1211 continue;
1212 }
1213 if( nop == Op_CatchProj ) { // Get count from a prior call
1214 // Assume call does not always throw exceptions: means the call-site
1215 // count is also the frequency of the fall-through path.
1216 assert( n->is_CatchProj(), "" );
1217 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1218 return 0.0f; // Assume call exception path is rare
1219 Node *call = n->in(0)->in(0)->in(0);
1220 assert( call->is_Call(), "expect a call here" );
1221 const JVMState *jvms = ((CallNode*)call)->jvms();
1222 ciMethodData* methodData = jvms->method()->method_data();
1223 if (!methodData->is_mature()) return 0.0f; // No call-site data
1224 ciProfileData* data = methodData->bci_to_data(jvms->bci());
1225 if ((data == NULL) || !data->is_CounterData()) {
1226 // no call profile available, try call's control input
1227 n = n->in(0);
1228 continue;
1229 }
1230 return data->as_CounterData()->count()/FreqCountInvocations;
1231 }
1232 // See if there's a gating IF test
1233 Node *n_c = n->in(0);
1234 if( !n_c->is_If() ) break; // No estimate available
1235 iff = n_c->as_If();
1236 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
1237 // Compute how much count comes on this path
1238 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1239 // Have no count info. Skip dull uncommon-trap like branches.
1240 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
1241 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1242 break;
1243 // Skip through never-taken branch; look for a real loop exit.
1244 n = iff->in(0);
1245 }
1246 return 0.0f; // No estimate available
1247 }
1249 //------------------------------merge_many_backedges---------------------------
1250 // Merge all the backedges from the shared header into a private Region.
1251 // Feed that region as the one backedge to this loop.
1252 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1253 uint i;
1255 // Scan for the top 2 hottest backedges
1256 float hotcnt = 0.0f;
1257 float warmcnt = 0.0f;
1258 uint hot_idx = 0;
1259 // Loop starts at 2 because slot 1 is the fall-in path
1260 for( i = 2; i < _head->req(); i++ ) {
1261 float cnt = estimate_path_freq(_head->in(i));
1262 if( cnt > hotcnt ) { // Grab hottest path
1263 warmcnt = hotcnt;
1264 hotcnt = cnt;
1265 hot_idx = i;
1266 } else if( cnt > warmcnt ) { // And 2nd hottest path
1267 warmcnt = cnt;
1268 }
1269 }
1271 // See if the hottest backedge is worthy of being an inner loop
1272 // by being much hotter than the next hottest backedge.
1273 if( hotcnt <= 0.0001 ||
1274 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1276 // Peel out the backedges into a private merge point; peel
1277 // them all except optionally hot_idx.
1278 PhaseIterGVN &igvn = phase->_igvn;
1280 Node *hot_tail = NULL;
1281 // Make a Region for the merge point
1282 Node *r = new (phase->C, 1) RegionNode(1);
1283 for( i = 2; i < _head->req(); i++ ) {
1284 if( i != hot_idx )
1285 r->add_req( _head->in(i) );
1286 else hot_tail = _head->in(i);
1287 }
1288 igvn.register_new_node_with_optimizer(r, _head);
1289 // Plug region into end of loop _head, followed by hot_tail
1290 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1291 _head->set_req(2, r);
1292 if( hot_idx ) _head->add_req(hot_tail);
1294 // Split all the Phis up between '_head' loop and the Region 'r'
1295 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1296 Node *out = _head->fast_out(j);
1297 if( out->is_Phi() ) {
1298 PhiNode* n = out->as_Phi();
1299 igvn.hash_delete(n); // Delete from hash before hacking edges
1300 Node *hot_phi = NULL;
1301 Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type());
1302 // Check all inputs for the ones to peel out
1303 uint j = 1;
1304 for( uint i = 2; i < n->req(); i++ ) {
1305 if( i != hot_idx )
1306 phi->set_req( j++, n->in(i) );
1307 else hot_phi = n->in(i);
1308 }
1309 // Register the phi but do not transform until whole place transforms
1310 igvn.register_new_node_with_optimizer(phi, n);
1311 // Add the merge phi to the old Phi
1312 while( n->req() > 3 ) n->del_req( n->req()-1 );
1313 n->set_req(2, phi);
1314 if( hot_idx ) n->add_req(hot_phi);
1315 }
1316 }
1319 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
1320 // of self loop tree. Turn self into a loop headed by _head and with
1321 // tail being the new merge point.
1322 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1323 phase->set_loop(_tail,ilt); // Adjust tail
1324 _tail = r; // Self's tail is new merge point
1325 phase->set_loop(r,this);
1326 ilt->_child = _child; // New guy has my children
1327 _child = ilt; // Self has new guy as only child
1328 ilt->_parent = this; // new guy has self for parent
1329 ilt->_nest = _nest; // Same nesting depth (for now)
1331 // Starting with 'ilt', look for child loop trees using the same shared
1332 // header. Flatten these out; they will no longer be loops in the end.
1333 IdealLoopTree **pilt = &_child;
1334 while( ilt ) {
1335 if( ilt->_head == _head ) {
1336 uint i;
1337 for( i = 2; i < _head->req(); i++ )
1338 if( _head->in(i) == ilt->_tail )
1339 break; // Still a loop
1340 if( i == _head->req() ) { // No longer a loop
1341 // Flatten ilt. Hang ilt's "_next" list from the end of
1342 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
1343 IdealLoopTree **cp = &ilt->_child;
1344 while( *cp ) cp = &(*cp)->_next; // Find end of child list
1345 *cp = ilt->_next; // Hang next list at end of child list
1346 *pilt = ilt->_child; // Move child up to replace ilt
1347 ilt->_head = NULL; // Flag as a loop UNIONED into parent
1348 ilt = ilt->_child; // Repeat using new ilt
1349 continue; // do not advance over ilt->_child
1350 }
1351 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1352 phase->set_loop(_head,ilt);
1353 }
1354 pilt = &ilt->_child; // Advance to next
1355 ilt = *pilt;
1356 }
1358 if( _child ) fix_parent( _child, this );
1359 }
1361 //------------------------------beautify_loops---------------------------------
1362 // Split shared headers and insert loop landing pads.
1363 // Insert a LoopNode to replace the RegionNode.
1364 // Return TRUE if loop tree is structurally changed.
1365 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1366 bool result = false;
1367 // Cache parts in locals for easy
1368 PhaseIterGVN &igvn = phase->_igvn;
1370 igvn.hash_delete(_head); // Yank from hash before hacking edges
1372 // Check for multiple fall-in paths. Peel off a landing pad if need be.
1373 int fall_in_cnt = 0;
1374 for( uint i = 1; i < _head->req(); i++ )
1375 if( !phase->is_member( this, _head->in(i) ) )
1376 fall_in_cnt++;
1377 assert( fall_in_cnt, "at least 1 fall-in path" );
1378 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
1379 split_fall_in( phase, fall_in_cnt );
1381 // Swap inputs to the _head and all Phis to move the fall-in edge to
1382 // the left.
1383 fall_in_cnt = 1;
1384 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1385 fall_in_cnt++;
1386 if( fall_in_cnt > 1 ) {
1387 // Since I am just swapping inputs I do not need to update def-use info
1388 Node *tmp = _head->in(1);
1389 _head->set_req( 1, _head->in(fall_in_cnt) );
1390 _head->set_req( fall_in_cnt, tmp );
1391 // Swap also all Phis
1392 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1393 Node* phi = _head->fast_out(i);
1394 if( phi->is_Phi() ) {
1395 igvn.hash_delete(phi); // Yank from hash before hacking edges
1396 tmp = phi->in(1);
1397 phi->set_req( 1, phi->in(fall_in_cnt) );
1398 phi->set_req( fall_in_cnt, tmp );
1399 }
1400 }
1401 }
1402 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1403 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
1405 // If I am a shared header (multiple backedges), peel off the many
1406 // backedges into a private merge point and use the merge point as
1407 // the one true backedge.
1408 if( _head->req() > 3 ) {
1409 // Merge the many backedges into a single backedge but leave
1410 // the hottest backedge as separate edge for the following peel.
1411 merge_many_backedges( phase );
1412 result = true;
1413 }
1415 // If I have one hot backedge, peel off myself loop.
1416 // I better be the outermost loop.
1417 if( _head->req() > 3 ) {
1418 split_outer_loop( phase );
1419 result = true;
1421 } else if( !_head->is_Loop() && !_irreducible ) {
1422 // Make a new LoopNode to replace the old loop head
1423 Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) );
1424 l = igvn.register_new_node_with_optimizer(l, _head);
1425 phase->set_created_loop_node();
1426 // Go ahead and replace _head
1427 phase->_igvn.replace_node( _head, l );
1428 _head = l;
1429 phase->set_loop(_head, this);
1430 }
1432 // Now recursively beautify nested loops
1433 if( _child ) result |= _child->beautify_loops( phase );
1434 if( _next ) result |= _next ->beautify_loops( phase );
1435 return result;
1436 }
1438 //------------------------------allpaths_check_safepts----------------------------
1439 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1440 // encountered. Helper for check_safepts.
1441 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1442 assert(stack.size() == 0, "empty stack");
1443 stack.push(_tail);
1444 visited.Clear();
1445 visited.set(_tail->_idx);
1446 while (stack.size() > 0) {
1447 Node* n = stack.pop();
1448 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1449 // Terminate this path
1450 } else if (n->Opcode() == Op_SafePoint) {
1451 if (_phase->get_loop(n) != this) {
1452 if (_required_safept == NULL) _required_safept = new Node_List();
1453 _required_safept->push(n); // save the one closest to the tail
1454 }
1455 // Terminate this path
1456 } else {
1457 uint start = n->is_Region() ? 1 : 0;
1458 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1459 for (uint i = start; i < end; i++) {
1460 Node* in = n->in(i);
1461 assert(in->is_CFG(), "must be");
1462 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1463 stack.push(in);
1464 }
1465 }
1466 }
1467 }
1468 }
1470 //------------------------------check_safepts----------------------------
1471 // Given dominators, try to find loops with calls that must always be
1472 // executed (call dominates loop tail). These loops do not need non-call
1473 // safepoints (ncsfpt).
1474 //
1475 // A complication is that a safepoint in a inner loop may be needed
1476 // by an outer loop. In the following, the inner loop sees it has a
1477 // call (block 3) on every path from the head (block 2) to the
1478 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
1479 // in block 2, _but_ this leaves the outer loop without a safepoint.
1480 //
1481 // entry 0
1482 // |
1483 // v
1484 // outer 1,2 +->1
1485 // | |
1486 // | v
1487 // | 2<---+ ncsfpt in 2
1488 // |_/|\ |
1489 // | v |
1490 // inner 2,3 / 3 | call in 3
1491 // / | |
1492 // v +--+
1493 // exit 4
1494 //
1495 //
1496 // This method creates a list (_required_safept) of ncsfpt nodes that must
1497 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1498 // is first looked for in the lists for the outer loops of the current loop.
1499 //
1500 // The insights into the problem:
1501 // A) counted loops are okay
1502 // B) innermost loops are okay (only an inner loop can delete
1503 // a ncsfpt needed by an outer loop)
1504 // C) a loop is immune from an inner loop deleting a safepoint
1505 // if the loop has a call on the idom-path
1506 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1507 // idom-path that is not in a nested loop
1508 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1509 // loop needs to be prevented from deletion by an inner loop
1510 //
1511 // There are two analyses:
1512 // 1) The first, and cheaper one, scans the loop body from
1513 // tail to head following the idom (immediate dominator)
1514 // chain, looking for the cases (C,D,E) above.
1515 // Since inner loops are scanned before outer loops, there is summary
1516 // information about inner loops. Inner loops can be skipped over
1517 // when the tail of an inner loop is encountered.
1518 //
1519 // 2) The second, invoked if the first fails to find a call or ncsfpt on
1520 // the idom path (which is rare), scans all predecessor control paths
1521 // from the tail to the head, terminating a path when a call or sfpt
1522 // is encountered, to find the ncsfpt's that are closest to the tail.
1523 //
1524 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1525 // Bottom up traversal
1526 IdealLoopTree* ch = _child;
1527 while (ch != NULL) {
1528 ch->check_safepts(visited, stack);
1529 ch = ch->_next;
1530 }
1532 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1533 bool has_call = false; // call on dom-path
1534 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1535 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth
1536 // Scan the dom-path nodes from tail to head
1537 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1538 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1539 has_call = true;
1540 _has_sfpt = 1; // Then no need for a safept!
1541 break;
1542 } else if (n->Opcode() == Op_SafePoint) {
1543 if (_phase->get_loop(n) == this) {
1544 has_local_ncsfpt = true;
1545 break;
1546 }
1547 if (nonlocal_ncsfpt == NULL) {
1548 nonlocal_ncsfpt = n; // save the one closest to the tail
1549 }
1550 } else {
1551 IdealLoopTree* nlpt = _phase->get_loop(n);
1552 if (this != nlpt) {
1553 // If at an inner loop tail, see if the inner loop has already
1554 // recorded seeing a call on the dom-path (and stop.) If not,
1555 // jump to the head of the inner loop.
1556 assert(is_member(nlpt), "nested loop");
1557 Node* tail = nlpt->_tail;
1558 if (tail->in(0)->is_If()) tail = tail->in(0);
1559 if (n == tail) {
1560 // If inner loop has call on dom-path, so does outer loop
1561 if (nlpt->_has_sfpt) {
1562 has_call = true;
1563 _has_sfpt = 1;
1564 break;
1565 }
1566 // Skip to head of inner loop
1567 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1568 n = nlpt->_head;
1569 }
1570 }
1571 }
1572 }
1573 // Record safept's that this loop needs preserved when an
1574 // inner loop attempts to delete it's safepoints.
1575 if (_child != NULL && !has_call && !has_local_ncsfpt) {
1576 if (nonlocal_ncsfpt != NULL) {
1577 if (_required_safept == NULL) _required_safept = new Node_List();
1578 _required_safept->push(nonlocal_ncsfpt);
1579 } else {
1580 // Failed to find a suitable safept on the dom-path. Now use
1581 // an all paths walk from tail to head, looking for safepoints to preserve.
1582 allpaths_check_safepts(visited, stack);
1583 }
1584 }
1585 }
1586 }
1588 //---------------------------is_deleteable_safept----------------------------
1589 // Is safept not required by an outer loop?
1590 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1591 assert(sfpt->Opcode() == Op_SafePoint, "");
1592 IdealLoopTree* lp = get_loop(sfpt)->_parent;
1593 while (lp != NULL) {
1594 Node_List* sfpts = lp->_required_safept;
1595 if (sfpts != NULL) {
1596 for (uint i = 0; i < sfpts->size(); i++) {
1597 if (sfpt == sfpts->at(i))
1598 return false;
1599 }
1600 }
1601 lp = lp->_parent;
1602 }
1603 return true;
1604 }
1606 //---------------------------replace_parallel_iv-------------------------------
1607 // Replace parallel induction variable (parallel to trip counter)
1608 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1609 assert(loop->_head->is_CountedLoop(), "");
1610 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1611 Node *incr = cl->incr();
1612 if (incr == NULL)
1613 return; // Dead loop?
1614 Node *init = cl->init_trip();
1615 Node *phi = cl->phi();
1616 // protect against stride not being a constant
1617 if (!cl->stride_is_con())
1618 return;
1619 int stride_con = cl->stride_con();
1621 PhaseGVN *gvn = &_igvn;
1623 // Visit all children, looking for Phis
1624 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1625 Node *out = cl->out(i);
1626 // Look for other phis (secondary IVs). Skip dead ones
1627 if (!out->is_Phi() || out == phi || !has_node(out))
1628 continue;
1629 PhiNode* phi2 = out->as_Phi();
1630 Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1631 // Look for induction variables of the form: X += constant
1632 if (phi2->region() != loop->_head ||
1633 incr2->req() != 3 ||
1634 incr2->in(1) != phi2 ||
1635 incr2 == incr ||
1636 incr2->Opcode() != Op_AddI ||
1637 !incr2->in(2)->is_Con())
1638 continue;
1640 // Check for parallel induction variable (parallel to trip counter)
1641 // via an affine function. In particular, count-down loops with
1642 // count-up array indices are common. We only RCE references off
1643 // the trip-counter, so we need to convert all these to trip-counter
1644 // expressions.
1645 Node *init2 = phi2->in( LoopNode::EntryControl );
1646 int stride_con2 = incr2->in(2)->get_int();
1648 // The general case here gets a little tricky. We want to find the
1649 // GCD of all possible parallel IV's and make a new IV using this
1650 // GCD for the loop. Then all possible IVs are simple multiples of
1651 // the GCD. In practice, this will cover very few extra loops.
1652 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1653 // where +/-1 is the common case, but other integer multiples are
1654 // also easy to handle.
1655 int ratio_con = stride_con2/stride_con;
1657 if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1658 // Convert to using the trip counter. The parallel induction
1659 // variable differs from the trip counter by a loop-invariant
1660 // amount, the difference between their respective initial values.
1661 // It is scaled by the 'ratio_con'.
1662 // Perform local Ideal transformation since in most cases ratio == 1.
1663 Node* ratio = _igvn.intcon(ratio_con);
1664 set_ctrl(ratio, C->root());
1665 Node* hook = new (C, 3) Node(3);
1666 Node* ratio_init = gvn->transform(new (C, 3) MulINode(init, ratio));
1667 hook->init_req(0, ratio_init);
1668 Node* diff = gvn->transform(new (C, 3) SubINode(init2, ratio_init));
1669 hook->init_req(1, diff);
1670 Node* ratio_idx = gvn->transform(new (C, 3) MulINode(phi, ratio));
1671 hook->init_req(2, ratio_idx);
1672 Node* add = gvn->transform(new (C, 3) AddINode(ratio_idx, diff));
1673 set_subtree_ctrl(add);
1674 _igvn.replace_node( phi2, add );
1675 // Free up intermediate goo
1676 _igvn.remove_dead_node(hook);
1677 // Sometimes an induction variable is unused
1678 if (add->outcnt() == 0) {
1679 _igvn.remove_dead_node(add);
1680 }
1681 --i; // deleted this phi; rescan starting with next position
1682 continue;
1683 }
1684 }
1685 }
1687 //------------------------------counted_loop-----------------------------------
1688 // Convert to counted loops where possible
1689 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1691 // For grins, set the inner-loop flag here
1692 if (!_child) {
1693 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1694 }
1696 if (_head->is_CountedLoop() ||
1697 phase->is_counted_loop(_head, this)) {
1698 _has_sfpt = 1; // Indicate we do not need a safepoint here
1700 // Look for a safepoint to remove
1701 for (Node* n = tail(); n != _head; n = phase->idom(n))
1702 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1703 phase->is_deleteable_safept(n))
1704 phase->lazy_replace(n,n->in(TypeFunc::Control));
1706 // Look for induction variables
1707 phase->replace_parallel_iv(this);
1709 } else if (_parent != NULL && !_irreducible) {
1710 // Not a counted loop.
1711 // Look for a safepoint on the idom-path to remove, preserving the first one
1712 bool found = false;
1713 Node* n = tail();
1714 for (; n != _head && !found; n = phase->idom(n)) {
1715 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this)
1716 found = true; // Found one
1717 }
1718 // Skip past it and delete the others
1719 for (; n != _head; n = phase->idom(n)) {
1720 if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this &&
1721 phase->is_deleteable_safept(n))
1722 phase->lazy_replace(n,n->in(TypeFunc::Control));
1723 }
1724 }
1726 // Recursively
1727 if (_child) _child->counted_loop( phase );
1728 if (_next) _next ->counted_loop( phase );
1729 }
1731 #ifndef PRODUCT
1732 //------------------------------dump_head--------------------------------------
1733 // Dump 1 liner for loop header info
1734 void IdealLoopTree::dump_head( ) const {
1735 for (uint i=0; i<_nest; i++)
1736 tty->print(" ");
1737 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1738 if (_irreducible) tty->print(" IRREDUCIBLE");
1739 Node* entry = _head->in(LoopNode::EntryControl);
1740 if (LoopLimitCheck) {
1741 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1742 if (predicate != NULL ) {
1743 tty->print(" limit_check");
1744 entry = entry->in(0)->in(0);
1745 }
1746 }
1747 if (UseLoopPredicate) {
1748 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1749 if (entry != NULL) {
1750 tty->print(" predicated");
1751 }
1752 }
1753 if (_head->is_CountedLoop()) {
1754 CountedLoopNode *cl = _head->as_CountedLoop();
1755 tty->print(" counted");
1757 Node* init_n = cl->init_trip();
1758 if (init_n != NULL && init_n->is_Con())
1759 tty->print(" [%d,", cl->init_trip()->get_int());
1760 else
1761 tty->print(" [int,");
1762 Node* limit_n = cl->limit();
1763 if (limit_n != NULL && limit_n->is_Con())
1764 tty->print("%d),", cl->limit()->get_int());
1765 else
1766 tty->print("int),");
1767 int stride_con = cl->stride_con();
1768 if (stride_con > 0) tty->print("+");
1769 tty->print("%d", stride_con);
1771 if (cl->is_pre_loop ()) tty->print(" pre" );
1772 if (cl->is_main_loop()) tty->print(" main");
1773 if (cl->is_post_loop()) tty->print(" post");
1774 }
1775 tty->cr();
1776 }
1778 //------------------------------dump-------------------------------------------
1779 // Dump loops by loop tree
1780 void IdealLoopTree::dump( ) const {
1781 dump_head();
1782 if (_child) _child->dump();
1783 if (_next) _next ->dump();
1784 }
1786 #endif
1788 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1789 if (loop == root) {
1790 if (loop->_child != NULL) {
1791 log->begin_head("loop_tree");
1792 log->end_head();
1793 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1794 log->tail("loop_tree");
1795 assert(loop->_next == NULL, "what?");
1796 }
1797 } else {
1798 Node* head = loop->_head;
1799 log->begin_head("loop");
1800 log->print(" idx='%d' ", head->_idx);
1801 if (loop->_irreducible) log->print("irreducible='1' ");
1802 if (head->is_Loop()) {
1803 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1804 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1805 }
1806 if (head->is_CountedLoop()) {
1807 CountedLoopNode* cl = head->as_CountedLoop();
1808 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
1809 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1810 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
1811 }
1812 log->end_head();
1813 if( loop->_child ) log_loop_tree(root, loop->_child, log);
1814 log->tail("loop");
1815 if( loop->_next ) log_loop_tree(root, loop->_next, log);
1816 }
1817 }
1819 //---------------------collect_potentially_useful_predicates-----------------------
1820 // Helper function to collect potentially useful predicates to prevent them from
1821 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1822 void PhaseIdealLoop::collect_potentially_useful_predicates(
1823 IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1824 if (loop->_child) { // child
1825 collect_potentially_useful_predicates(loop->_child, useful_predicates);
1826 }
1828 // self (only loops that we can apply loop predication may use their predicates)
1829 if (loop->_head->is_Loop() &&
1830 !loop->_irreducible &&
1831 !loop->tail()->is_top()) {
1832 LoopNode* lpn = loop->_head->as_Loop();
1833 Node* entry = lpn->in(LoopNode::EntryControl);
1834 Node* predicate_proj = find_predicate(entry); // loop_limit_check first
1835 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
1836 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
1837 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1838 entry = entry->in(0)->in(0);
1839 }
1840 predicate_proj = find_predicate(entry); // Predicate
1841 if (predicate_proj != NULL ) {
1842 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
1843 }
1844 }
1846 if (loop->_next) { // sibling
1847 collect_potentially_useful_predicates(loop->_next, useful_predicates);
1848 }
1849 }
1851 //------------------------eliminate_useless_predicates-----------------------------
1852 // Eliminate all inserted predicates if they could not be used by loop predication.
1853 // Note: it will also eliminates loop limits check predicate since it also uses
1854 // Opaque1 node (see Parse::add_predicate()).
1855 void PhaseIdealLoop::eliminate_useless_predicates() {
1856 if (C->predicate_count() == 0)
1857 return; // no predicate left
1859 Unique_Node_List useful_predicates; // to store useful predicates
1860 if (C->has_loops()) {
1861 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
1862 }
1864 for (int i = C->predicate_count(); i > 0; i--) {
1865 Node * n = C->predicate_opaque1_node(i-1);
1866 assert(n->Opcode() == Op_Opaque1, "must be");
1867 if (!useful_predicates.member(n)) { // not in the useful list
1868 _igvn.replace_node(n, n->in(1));
1869 }
1870 }
1871 }
1873 //=============================================================================
1874 //----------------------------build_and_optimize-------------------------------
1875 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
1876 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
1877 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs) {
1878 ResourceMark rm;
1880 int old_progress = C->major_progress();
1881 uint orig_worklist_size = _igvn._worklist.size();
1883 // Reset major-progress flag for the driver's heuristics
1884 C->clear_major_progress();
1886 #ifndef PRODUCT
1887 // Capture for later assert
1888 uint unique = C->unique();
1889 _loop_invokes++;
1890 _loop_work += unique;
1891 #endif
1893 // True if the method has at least 1 irreducible loop
1894 _has_irreducible_loops = false;
1896 _created_loop_node = false;
1898 Arena *a = Thread::current()->resource_area();
1899 VectorSet visited(a);
1900 // Pre-grow the mapping from Nodes to IdealLoopTrees.
1901 _nodes.map(C->unique(), NULL);
1902 memset(_nodes.adr(), 0, wordSize * C->unique());
1904 // Pre-build the top-level outermost loop tree entry
1905 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
1906 // Do not need a safepoint at the top level
1907 _ltree_root->_has_sfpt = 1;
1909 // Initialize Dominators.
1910 // Checked in clone_loop_predicate() during beautify_loops().
1911 _idom_size = 0;
1912 _idom = NULL;
1913 _dom_depth = NULL;
1914 _dom_stk = NULL;
1916 // Empty pre-order array
1917 allocate_preorders();
1919 // Build a loop tree on the fly. Build a mapping from CFG nodes to
1920 // IdealLoopTree entries. Data nodes are NOT walked.
1921 build_loop_tree();
1922 // Check for bailout, and return
1923 if (C->failing()) {
1924 return;
1925 }
1927 // No loops after all
1928 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
1930 // There should always be an outer loop containing the Root and Return nodes.
1931 // If not, we have a degenerate empty program. Bail out in this case.
1932 if (!has_node(C->root())) {
1933 if (!_verify_only) {
1934 C->clear_major_progress();
1935 C->record_method_not_compilable("empty program detected during loop optimization");
1936 }
1937 return;
1938 }
1940 // Nothing to do, so get out
1941 if( !C->has_loops() && !do_split_ifs && !_verify_me && !_verify_only ) {
1942 _igvn.optimize(); // Cleanup NeverBranches
1943 return;
1944 }
1946 // Set loop nesting depth
1947 _ltree_root->set_nest( 0 );
1949 // Split shared headers and insert loop landing pads.
1950 // Do not bother doing this on the Root loop of course.
1951 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
1952 C->print_method("Before beautify loops", 3);
1953 if( _ltree_root->_child->beautify_loops( this ) ) {
1954 // Re-build loop tree!
1955 _ltree_root->_child = NULL;
1956 _nodes.clear();
1957 reallocate_preorders();
1958 build_loop_tree();
1959 // Check for bailout, and return
1960 if (C->failing()) {
1961 return;
1962 }
1963 // Reset loop nesting depth
1964 _ltree_root->set_nest( 0 );
1966 C->print_method("After beautify loops", 3);
1967 }
1968 }
1970 // Build Dominators for elision of NULL checks & loop finding.
1971 // Since nodes do not have a slot for immediate dominator, make
1972 // a persistent side array for that info indexed on node->_idx.
1973 _idom_size = C->unique();
1974 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
1975 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
1976 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth
1977 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
1979 Dominators();
1981 if (!_verify_only) {
1982 // As a side effect, Dominators removed any unreachable CFG paths
1983 // into RegionNodes. It doesn't do this test against Root, so
1984 // we do it here.
1985 for( uint i = 1; i < C->root()->req(); i++ ) {
1986 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
1987 _igvn.hash_delete(C->root());
1988 C->root()->del_req(i);
1989 _igvn._worklist.push(C->root());
1990 i--; // Rerun same iteration on compressed edges
1991 }
1992 }
1994 // Given dominators, try to find inner loops with calls that must
1995 // always be executed (call dominates loop tail). These loops do
1996 // not need a separate safepoint.
1997 Node_List cisstack(a);
1998 _ltree_root->check_safepts(visited, cisstack);
1999 }
2001 // Walk the DATA nodes and place into loops. Find earliest control
2002 // node. For CFG nodes, the _nodes array starts out and remains
2003 // holding the associated IdealLoopTree pointer. For DATA nodes, the
2004 // _nodes array holds the earliest legal controlling CFG node.
2006 // Allocate stack with enough space to avoid frequent realloc
2007 int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats
2008 Node_Stack nstack( a, stack_size );
2010 visited.Clear();
2011 Node_List worklist(a);
2012 // Don't need C->root() on worklist since
2013 // it will be processed among C->top() inputs
2014 worklist.push( C->top() );
2015 visited.set( C->top()->_idx ); // Set C->top() as visited now
2016 build_loop_early( visited, worklist, nstack );
2018 // Given early legal placement, try finding counted loops. This placement
2019 // is good enough to discover most loop invariants.
2020 if( !_verify_me && !_verify_only )
2021 _ltree_root->counted_loop( this );
2023 // Find latest loop placement. Find ideal loop placement.
2024 visited.Clear();
2025 init_dom_lca_tags();
2026 // Need C->root() on worklist when processing outs
2027 worklist.push( C->root() );
2028 NOT_PRODUCT( C->verify_graph_edges(); )
2029 worklist.push( C->top() );
2030 build_loop_late( visited, worklist, nstack );
2032 if (_verify_only) {
2033 // restore major progress flag
2034 for (int i = 0; i < old_progress; i++)
2035 C->set_major_progress();
2036 assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2037 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2038 return;
2039 }
2041 // Some parser-inserted loop predicates could never be used by loop
2042 // predication or they were moved away from loop during some optimizations.
2043 // For example, peeling. Eliminate them before next loop optimizations.
2044 if (UseLoopPredicate || LoopLimitCheck) {
2045 eliminate_useless_predicates();
2046 }
2048 // clear out the dead code
2049 while(_deadlist.size()) {
2050 _igvn.remove_globally_dead_node(_deadlist.pop());
2051 }
2053 #ifndef PRODUCT
2054 C->verify_graph_edges();
2055 if (_verify_me) { // Nested verify pass?
2056 // Check to see if the verify mode is broken
2057 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2058 return;
2059 }
2060 if(VerifyLoopOptimizations) verify();
2061 if(TraceLoopOpts && C->has_loops()) {
2062 _ltree_root->dump();
2063 }
2064 #endif
2066 if (ReassociateInvariants) {
2067 // Reassociate invariants and prep for split_thru_phi
2068 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2069 IdealLoopTree* lpt = iter.current();
2070 if (!lpt->is_counted() || !lpt->is_inner()) continue;
2072 lpt->reassociate_invariants(this);
2074 // Because RCE opportunities can be masked by split_thru_phi,
2075 // look for RCE candidates and inhibit split_thru_phi
2076 // on just their loop-phi's for this pass of loop opts
2077 if (SplitIfBlocks && do_split_ifs) {
2078 if (lpt->policy_range_check(this)) {
2079 lpt->_rce_candidate = 1; // = true
2080 }
2081 }
2082 }
2083 }
2085 // Check for aggressive application of split-if and other transforms
2086 // that require basic-block info (like cloning through Phi's)
2087 if( SplitIfBlocks && do_split_ifs ) {
2088 visited.Clear();
2089 split_if_with_blocks( visited, nstack );
2090 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2091 }
2093 // Perform loop predication before iteration splitting
2094 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2095 _ltree_root->_child->loop_predication(this);
2096 }
2098 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2099 if (do_intrinsify_fill()) {
2100 C->set_major_progress();
2101 }
2102 }
2104 // Perform iteration-splitting on inner loops. Split iterations to avoid
2105 // range checks or one-shot null checks.
2107 // If split-if's didn't hack the graph too bad (no CFG changes)
2108 // then do loop opts.
2109 if (C->has_loops() && !C->major_progress()) {
2110 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2111 _ltree_root->_child->iteration_split( this, worklist );
2112 // No verify after peeling! GCM has hoisted code out of the loop.
2113 // After peeling, the hoisted code could sink inside the peeled area.
2114 // The peeling code does not try to recompute the best location for
2115 // all the code before the peeled area, so the verify pass will always
2116 // complain about it.
2117 }
2118 // Do verify graph edges in any case
2119 NOT_PRODUCT( C->verify_graph_edges(); );
2121 if (!do_split_ifs) {
2122 // We saw major progress in Split-If to get here. We forced a
2123 // pass with unrolling and not split-if, however more split-if's
2124 // might make progress. If the unrolling didn't make progress
2125 // then the major-progress flag got cleared and we won't try
2126 // another round of Split-If. In particular the ever-common
2127 // instance-of/check-cast pattern requires at least 2 rounds of
2128 // Split-If to clear out.
2129 C->set_major_progress();
2130 }
2132 // Repeat loop optimizations if new loops were seen
2133 if (created_loop_node()) {
2134 C->set_major_progress();
2135 }
2137 // Keep loop predicates and perform optimizations with them
2138 // until no more loop optimizations could be done.
2139 // After that switch predicates off and do more loop optimizations.
2140 if (!C->major_progress() && (C->predicate_count() > 0)) {
2141 C->cleanup_loop_predicates(_igvn);
2142 #ifndef PRODUCT
2143 if (TraceLoopOpts) {
2144 tty->print_cr("PredicatesOff");
2145 }
2146 #endif
2147 C->set_major_progress();
2148 }
2150 // Convert scalar to superword operations at the end of all loop opts.
2151 if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2152 // SuperWord transform
2153 SuperWord sw(this);
2154 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2155 IdealLoopTree* lpt = iter.current();
2156 if (lpt->is_counted()) {
2157 sw.transform_loop(lpt);
2158 }
2159 }
2160 }
2162 // Cleanup any modified bits
2163 _igvn.optimize();
2165 // disable assert until issue with split_flow_path is resolved (6742111)
2166 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2167 // "shouldn't introduce irreducible loops");
2169 if (C->log() != NULL) {
2170 log_loop_tree(_ltree_root, _ltree_root, C->log());
2171 }
2172 }
2174 #ifndef PRODUCT
2175 //------------------------------print_statistics-------------------------------
2176 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2177 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2178 void PhaseIdealLoop::print_statistics() {
2179 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2180 }
2182 //------------------------------verify-----------------------------------------
2183 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2184 static int fail; // debug only, so its multi-thread dont care
2185 void PhaseIdealLoop::verify() const {
2186 int old_progress = C->major_progress();
2187 ResourceMark rm;
2188 PhaseIdealLoop loop_verify( _igvn, this );
2189 VectorSet visited(Thread::current()->resource_area());
2191 fail = 0;
2192 verify_compare( C->root(), &loop_verify, visited );
2193 assert( fail == 0, "verify loops failed" );
2194 // Verify loop structure is the same
2195 _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2196 // Reset major-progress. It was cleared by creating a verify version of
2197 // PhaseIdealLoop.
2198 for( int i=0; i<old_progress; i++ )
2199 C->set_major_progress();
2200 }
2202 //------------------------------verify_compare---------------------------------
2203 // Make sure me and the given PhaseIdealLoop agree on key data structures
2204 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2205 if( !n ) return;
2206 if( visited.test_set( n->_idx ) ) return;
2207 if( !_nodes[n->_idx] ) { // Unreachable
2208 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2209 return;
2210 }
2212 uint i;
2213 for( i = 0; i < n->req(); i++ )
2214 verify_compare( n->in(i), loop_verify, visited );
2216 // Check the '_nodes' block/loop structure
2217 i = n->_idx;
2218 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl
2219 if( _nodes[i] != loop_verify->_nodes[i] &&
2220 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2221 tty->print("Mismatched control setting for: ");
2222 n->dump();
2223 if( fail++ > 10 ) return;
2224 Node *c = get_ctrl_no_update(n);
2225 tty->print("We have it as: ");
2226 if( c->in(0) ) c->dump();
2227 else tty->print_cr("N%d",c->_idx);
2228 tty->print("Verify thinks: ");
2229 if( loop_verify->has_ctrl(n) )
2230 loop_verify->get_ctrl_no_update(n)->dump();
2231 else
2232 loop_verify->get_loop_idx(n)->dump();
2233 tty->cr();
2234 }
2235 } else { // We have a loop
2236 IdealLoopTree *us = get_loop_idx(n);
2237 if( loop_verify->has_ctrl(n) ) {
2238 tty->print("Mismatched loop setting for: ");
2239 n->dump();
2240 if( fail++ > 10 ) return;
2241 tty->print("We have it as: ");
2242 us->dump();
2243 tty->print("Verify thinks: ");
2244 loop_verify->get_ctrl_no_update(n)->dump();
2245 tty->cr();
2246 } else if (!C->major_progress()) {
2247 // Loop selection can be messed up if we did a major progress
2248 // operation, like split-if. Do not verify in that case.
2249 IdealLoopTree *them = loop_verify->get_loop_idx(n);
2250 if( us->_head != them->_head || us->_tail != them->_tail ) {
2251 tty->print("Unequals loops for: ");
2252 n->dump();
2253 if( fail++ > 10 ) return;
2254 tty->print("We have it as: ");
2255 us->dump();
2256 tty->print("Verify thinks: ");
2257 them->dump();
2258 tty->cr();
2259 }
2260 }
2261 }
2263 // Check for immediate dominators being equal
2264 if( i >= _idom_size ) {
2265 if( !n->is_CFG() ) return;
2266 tty->print("CFG Node with no idom: ");
2267 n->dump();
2268 return;
2269 }
2270 if( !n->is_CFG() ) return;
2271 if( n == C->root() ) return; // No IDOM here
2273 assert(n->_idx == i, "sanity");
2274 Node *id = idom_no_update(n);
2275 if( id != loop_verify->idom_no_update(n) ) {
2276 tty->print("Unequals idoms for: ");
2277 n->dump();
2278 if( fail++ > 10 ) return;
2279 tty->print("We have it as: ");
2280 id->dump();
2281 tty->print("Verify thinks: ");
2282 loop_verify->idom_no_update(n)->dump();
2283 tty->cr();
2284 }
2286 }
2288 //------------------------------verify_tree------------------------------------
2289 // Verify that tree structures match. Because the CFG can change, siblings
2290 // within the loop tree can be reordered. We attempt to deal with that by
2291 // reordering the verify's loop tree if possible.
2292 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2293 assert( _parent == parent, "Badly formed loop tree" );
2295 // Siblings not in same order? Attempt to re-order.
2296 if( _head != loop->_head ) {
2297 // Find _next pointer to update
2298 IdealLoopTree **pp = &loop->_parent->_child;
2299 while( *pp != loop )
2300 pp = &((*pp)->_next);
2301 // Find proper sibling to be next
2302 IdealLoopTree **nn = &loop->_next;
2303 while( (*nn) && (*nn)->_head != _head )
2304 nn = &((*nn)->_next);
2306 // Check for no match.
2307 if( !(*nn) ) {
2308 // Annoyingly, irreducible loops can pick different headers
2309 // after a major_progress operation, so the rest of the loop
2310 // tree cannot be matched.
2311 if (_irreducible && Compile::current()->major_progress()) return;
2312 assert( 0, "failed to match loop tree" );
2313 }
2315 // Move (*nn) to (*pp)
2316 IdealLoopTree *hit = *nn;
2317 *nn = hit->_next;
2318 hit->_next = loop;
2319 *pp = loop;
2320 loop = hit;
2321 // Now try again to verify
2322 }
2324 assert( _head == loop->_head , "mismatched loop head" );
2325 Node *tail = _tail; // Inline a non-updating version of
2326 while( !tail->in(0) ) // the 'tail()' call.
2327 tail = tail->in(1);
2328 assert( tail == loop->_tail, "mismatched loop tail" );
2330 // Counted loops that are guarded should be able to find their guards
2331 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2332 CountedLoopNode *cl = _head->as_CountedLoop();
2333 Node *init = cl->init_trip();
2334 Node *ctrl = cl->in(LoopNode::EntryControl);
2335 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2336 Node *iff = ctrl->in(0);
2337 assert( iff->Opcode() == Op_If, "" );
2338 Node *bol = iff->in(1);
2339 assert( bol->Opcode() == Op_Bool, "" );
2340 Node *cmp = bol->in(1);
2341 assert( cmp->Opcode() == Op_CmpI, "" );
2342 Node *add = cmp->in(1);
2343 Node *opaq;
2344 if( add->Opcode() == Op_Opaque1 ) {
2345 opaq = add;
2346 } else {
2347 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2348 assert( add == init, "" );
2349 opaq = cmp->in(2);
2350 }
2351 assert( opaq->Opcode() == Op_Opaque1, "" );
2353 }
2355 if (_child != NULL) _child->verify_tree(loop->_child, this);
2356 if (_next != NULL) _next ->verify_tree(loop->_next, parent);
2357 // Innermost loops need to verify loop bodies,
2358 // but only if no 'major_progress'
2359 int fail = 0;
2360 if (!Compile::current()->major_progress() && _child == NULL) {
2361 for( uint i = 0; i < _body.size(); i++ ) {
2362 Node *n = _body.at(i);
2363 if (n->outcnt() == 0) continue; // Ignore dead
2364 uint j;
2365 for( j = 0; j < loop->_body.size(); j++ )
2366 if( loop->_body.at(j) == n )
2367 break;
2368 if( j == loop->_body.size() ) { // Not found in loop body
2369 // Last ditch effort to avoid assertion: Its possible that we
2370 // have some users (so outcnt not zero) but are still dead.
2371 // Try to find from root.
2372 if (Compile::current()->root()->find(n->_idx)) {
2373 fail++;
2374 tty->print("We have that verify does not: ");
2375 n->dump();
2376 }
2377 }
2378 }
2379 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2380 Node *n = loop->_body.at(i2);
2381 if (n->outcnt() == 0) continue; // Ignore dead
2382 uint j;
2383 for( j = 0; j < _body.size(); j++ )
2384 if( _body.at(j) == n )
2385 break;
2386 if( j == _body.size() ) { // Not found in loop body
2387 // Last ditch effort to avoid assertion: Its possible that we
2388 // have some users (so outcnt not zero) but are still dead.
2389 // Try to find from root.
2390 if (Compile::current()->root()->find(n->_idx)) {
2391 fail++;
2392 tty->print("Verify has that we do not: ");
2393 n->dump();
2394 }
2395 }
2396 }
2397 assert( !fail, "loop body mismatch" );
2398 }
2399 }
2401 #endif
2403 //------------------------------set_idom---------------------------------------
2404 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2405 uint idx = d->_idx;
2406 if (idx >= _idom_size) {
2407 uint newsize = _idom_size<<1;
2408 while( idx >= newsize ) {
2409 newsize <<= 1;
2410 }
2411 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
2412 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2413 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2414 _idom_size = newsize;
2415 }
2416 _idom[idx] = n;
2417 _dom_depth[idx] = dom_depth;
2418 }
2420 //------------------------------recompute_dom_depth---------------------------------------
2421 // The dominator tree is constructed with only parent pointers.
2422 // This recomputes the depth in the tree by first tagging all
2423 // nodes as "no depth yet" marker. The next pass then runs up
2424 // the dom tree from each node marked "no depth yet", and computes
2425 // the depth on the way back down.
2426 void PhaseIdealLoop::recompute_dom_depth() {
2427 uint no_depth_marker = C->unique();
2428 uint i;
2429 // Initialize depth to "no depth yet"
2430 for (i = 0; i < _idom_size; i++) {
2431 if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2432 _dom_depth[i] = no_depth_marker;
2433 }
2434 }
2435 if (_dom_stk == NULL) {
2436 uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size.
2437 if (init_size < 10) init_size = 10;
2438 _dom_stk = new GrowableArray<uint>(init_size);
2439 }
2440 // Compute new depth for each node.
2441 for (i = 0; i < _idom_size; i++) {
2442 uint j = i;
2443 // Run up the dom tree to find a node with a depth
2444 while (_dom_depth[j] == no_depth_marker) {
2445 _dom_stk->push(j);
2446 j = _idom[j]->_idx;
2447 }
2448 // Compute the depth on the way back down this tree branch
2449 uint dd = _dom_depth[j] + 1;
2450 while (_dom_stk->length() > 0) {
2451 uint j = _dom_stk->pop();
2452 _dom_depth[j] = dd;
2453 dd++;
2454 }
2455 }
2456 }
2458 //------------------------------sort-------------------------------------------
2459 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
2460 // loop tree, not the root.
2461 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2462 if( !innermost ) return loop; // New innermost loop
2464 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2465 assert( loop_preorder, "not yet post-walked loop" );
2466 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
2467 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
2469 // Insert at start of list
2470 while( l ) { // Insertion sort based on pre-order
2471 if( l == loop ) return innermost; // Already on list!
2472 int l_preorder = get_preorder(l->_head); // Cache pre-order number
2473 assert( l_preorder, "not yet post-walked l" );
2474 // Check header pre-order number to figure proper nesting
2475 if( loop_preorder > l_preorder )
2476 break; // End of insertion
2477 // If headers tie (e.g., shared headers) check tail pre-order numbers.
2478 // Since I split shared headers, you'd think this could not happen.
2479 // BUT: I must first do the preorder numbering before I can discover I
2480 // have shared headers, so the split headers all get the same preorder
2481 // number as the RegionNode they split from.
2482 if( loop_preorder == l_preorder &&
2483 get_preorder(loop->_tail) < get_preorder(l->_tail) )
2484 break; // Also check for shared headers (same pre#)
2485 pp = &l->_parent; // Chain up list
2486 l = *pp;
2487 }
2488 // Link into list
2489 // Point predecessor to me
2490 *pp = loop;
2491 // Point me to successor
2492 IdealLoopTree *p = loop->_parent;
2493 loop->_parent = l; // Point me to successor
2494 if( p ) sort( p, innermost ); // Insert my parents into list as well
2495 return innermost;
2496 }
2498 //------------------------------build_loop_tree--------------------------------
2499 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
2500 // bits. The _nodes[] array is mapped by Node index and holds a NULL for
2501 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2502 // tightest enclosing IdealLoopTree for post-walked.
2503 //
2504 // During my forward walk I do a short 1-layer lookahead to see if I can find
2505 // a loop backedge with that doesn't have any work on the backedge. This
2506 // helps me construct nested loops with shared headers better.
2507 //
2508 // Once I've done the forward recursion, I do the post-work. For each child
2509 // I check to see if there is a backedge. Backedges define a loop! I
2510 // insert an IdealLoopTree at the target of the backedge.
2511 //
2512 // During the post-work I also check to see if I have several children
2513 // belonging to different loops. If so, then this Node is a decision point
2514 // where control flow can choose to change loop nests. It is at this
2515 // decision point where I can figure out how loops are nested. At this
2516 // time I can properly order the different loop nests from my children.
2517 // Note that there may not be any backedges at the decision point!
2518 //
2519 // Since the decision point can be far removed from the backedges, I can't
2520 // order my loops at the time I discover them. Thus at the decision point
2521 // I need to inspect loop header pre-order numbers to properly nest my
2522 // loops. This means I need to sort my childrens' loops by pre-order.
2523 // The sort is of size number-of-control-children, which generally limits
2524 // it to size 2 (i.e., I just choose between my 2 target loops).
2525 void PhaseIdealLoop::build_loop_tree() {
2526 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2527 GrowableArray <Node *> bltstack(C->unique() >> 1);
2528 Node *n = C->root();
2529 bltstack.push(n);
2530 int pre_order = 1;
2531 int stack_size;
2533 while ( ( stack_size = bltstack.length() ) != 0 ) {
2534 n = bltstack.top(); // Leave node on stack
2535 if ( !is_visited(n) ) {
2536 // ---- Pre-pass Work ----
2537 // Pre-walked but not post-walked nodes need a pre_order number.
2539 set_preorder_visited( n, pre_order ); // set as visited
2541 // ---- Scan over children ----
2542 // Scan first over control projections that lead to loop headers.
2543 // This helps us find inner-to-outer loops with shared headers better.
2545 // Scan children's children for loop headers.
2546 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2547 Node* m = n->raw_out(i); // Child
2548 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2549 // Scan over children's children to find loop
2550 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2551 Node* l = m->fast_out(j);
2552 if( is_visited(l) && // Been visited?
2553 !is_postvisited(l) && // But not post-visited
2554 get_preorder(l) < pre_order ) { // And smaller pre-order
2555 // Found! Scan the DFS down this path before doing other paths
2556 bltstack.push(m);
2557 break;
2558 }
2559 }
2560 }
2561 }
2562 pre_order++;
2563 }
2564 else if ( !is_postvisited(n) ) {
2565 // Note: build_loop_tree_impl() adds out edges on rare occasions,
2566 // such as com.sun.rsasign.am::a.
2567 // For non-recursive version, first, process current children.
2568 // On next iteration, check if additional children were added.
2569 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2570 Node* u = n->raw_out(k);
2571 if ( u->is_CFG() && !is_visited(u) ) {
2572 bltstack.push(u);
2573 }
2574 }
2575 if ( bltstack.length() == stack_size ) {
2576 // There were no additional children, post visit node now
2577 (void)bltstack.pop(); // Remove node from stack
2578 pre_order = build_loop_tree_impl( n, pre_order );
2579 // Check for bailout
2580 if (C->failing()) {
2581 return;
2582 }
2583 // Check to grow _preorders[] array for the case when
2584 // build_loop_tree_impl() adds new nodes.
2585 check_grow_preorders();
2586 }
2587 }
2588 else {
2589 (void)bltstack.pop(); // Remove post-visited node from stack
2590 }
2591 }
2592 }
2594 //------------------------------build_loop_tree_impl---------------------------
2595 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2596 // ---- Post-pass Work ----
2597 // Pre-walked but not post-walked nodes need a pre_order number.
2599 // Tightest enclosing loop for this Node
2600 IdealLoopTree *innermost = NULL;
2602 // For all children, see if any edge is a backedge. If so, make a loop
2603 // for it. Then find the tightest enclosing loop for the self Node.
2604 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2605 Node* m = n->fast_out(i); // Child
2606 if( n == m ) continue; // Ignore control self-cycles
2607 if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2609 IdealLoopTree *l; // Child's loop
2610 if( !is_postvisited(m) ) { // Child visited but not post-visited?
2611 // Found a backedge
2612 assert( get_preorder(m) < pre_order, "should be backedge" );
2613 // Check for the RootNode, which is already a LoopNode and is allowed
2614 // to have multiple "backedges".
2615 if( m == C->root()) { // Found the root?
2616 l = _ltree_root; // Root is the outermost LoopNode
2617 } else { // Else found a nested loop
2618 // Insert a LoopNode to mark this loop.
2619 l = new IdealLoopTree(this, m, n);
2620 } // End of Else found a nested loop
2621 if( !has_loop(m) ) // If 'm' does not already have a loop set
2622 set_loop(m, l); // Set loop header to loop now
2624 } else { // Else not a nested loop
2625 if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2626 l = get_loop(m); // Get previously determined loop
2627 // If successor is header of a loop (nest), move up-loop till it
2628 // is a member of some outer enclosing loop. Since there are no
2629 // shared headers (I've split them already) I only need to go up
2630 // at most 1 level.
2631 while( l && l->_head == m ) // Successor heads loop?
2632 l = l->_parent; // Move up 1 for me
2633 // If this loop is not properly parented, then this loop
2634 // has no exit path out, i.e. its an infinite loop.
2635 if( !l ) {
2636 // Make loop "reachable" from root so the CFG is reachable. Basically
2637 // insert a bogus loop exit that is never taken. 'm', the loop head,
2638 // points to 'n', one (of possibly many) fall-in paths. There may be
2639 // many backedges as well.
2641 // Here I set the loop to be the root loop. I could have, after
2642 // inserting a bogus loop exit, restarted the recursion and found my
2643 // new loop exit. This would make the infinite loop a first-class
2644 // loop and it would then get properly optimized. What's the use of
2645 // optimizing an infinite loop?
2646 l = _ltree_root; // Oops, found infinite loop
2648 if (!_verify_only) {
2649 // Insert the NeverBranch between 'm' and it's control user.
2650 NeverBranchNode *iff = new (C, 1) NeverBranchNode( m );
2651 _igvn.register_new_node_with_optimizer(iff);
2652 set_loop(iff, l);
2653 Node *if_t = new (C, 1) CProjNode( iff, 0 );
2654 _igvn.register_new_node_with_optimizer(if_t);
2655 set_loop(if_t, l);
2657 Node* cfg = NULL; // Find the One True Control User of m
2658 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2659 Node* x = m->fast_out(j);
2660 if (x->is_CFG() && x != m && x != iff)
2661 { cfg = x; break; }
2662 }
2663 assert(cfg != NULL, "must find the control user of m");
2664 uint k = 0; // Probably cfg->in(0)
2665 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2666 cfg->set_req( k, if_t ); // Now point to NeverBranch
2668 // Now create the never-taken loop exit
2669 Node *if_f = new (C, 1) CProjNode( iff, 1 );
2670 _igvn.register_new_node_with_optimizer(if_f);
2671 set_loop(if_f, l);
2672 // Find frame ptr for Halt. Relies on the optimizer
2673 // V-N'ing. Easier and quicker than searching through
2674 // the program structure.
2675 Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr );
2676 _igvn.register_new_node_with_optimizer(frame);
2677 // Halt & Catch Fire
2678 Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame );
2679 _igvn.register_new_node_with_optimizer(halt);
2680 set_loop(halt, l);
2681 C->root()->add_req(halt);
2682 }
2683 set_loop(C->root(), _ltree_root);
2684 }
2685 }
2686 // Weeny check for irreducible. This child was already visited (this
2687 // IS the post-work phase). Is this child's loop header post-visited
2688 // as well? If so, then I found another entry into the loop.
2689 if (!_verify_only) {
2690 while( is_postvisited(l->_head) ) {
2691 // found irreducible
2692 l->_irreducible = 1; // = true
2693 l = l->_parent;
2694 _has_irreducible_loops = true;
2695 // Check for bad CFG here to prevent crash, and bailout of compile
2696 if (l == NULL) {
2697 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2698 return pre_order;
2699 }
2700 }
2701 }
2703 // This Node might be a decision point for loops. It is only if
2704 // it's children belong to several different loops. The sort call
2705 // does a trivial amount of work if there is only 1 child or all
2706 // children belong to the same loop. If however, the children
2707 // belong to different loops, the sort call will properly set the
2708 // _parent pointers to show how the loops nest.
2709 //
2710 // In any case, it returns the tightest enclosing loop.
2711 innermost = sort( l, innermost );
2712 }
2714 // Def-use info will have some dead stuff; dead stuff will have no
2715 // loop decided on.
2717 // Am I a loop header? If so fix up my parent's child and next ptrs.
2718 if( innermost && innermost->_head == n ) {
2719 assert( get_loop(n) == innermost, "" );
2720 IdealLoopTree *p = innermost->_parent;
2721 IdealLoopTree *l = innermost;
2722 while( p && l->_head == n ) {
2723 l->_next = p->_child; // Put self on parents 'next child'
2724 p->_child = l; // Make self as first child of parent
2725 l = p; // Now walk up the parent chain
2726 p = l->_parent;
2727 }
2728 } else {
2729 // Note that it is possible for a LoopNode to reach here, if the
2730 // backedge has been made unreachable (hence the LoopNode no longer
2731 // denotes a Loop, and will eventually be removed).
2733 // Record tightest enclosing loop for self. Mark as post-visited.
2734 set_loop(n, innermost);
2735 // Also record has_call flag early on
2736 if( innermost ) {
2737 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
2738 // Do not count uncommon calls
2739 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
2740 Node *iff = n->in(0)->in(0);
2741 if( !iff->is_If() ||
2742 (n->in(0)->Opcode() == Op_IfFalse &&
2743 (1.0 - iff->as_If()->_prob) >= 0.01) ||
2744 (iff->as_If()->_prob >= 0.01) )
2745 innermost->_has_call = 1;
2746 }
2747 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
2748 // Disable loop optimizations if the loop has a scalar replaceable
2749 // allocation. This disabling may cause a potential performance lost
2750 // if the allocation is not eliminated for some reason.
2751 innermost->_allow_optimizations = false;
2752 innermost->_has_call = 1; // = true
2753 }
2754 }
2755 }
2757 // Flag as post-visited now
2758 set_postvisited(n);
2759 return pre_order;
2760 }
2763 //------------------------------build_loop_early-------------------------------
2764 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
2765 // First pass computes the earliest controlling node possible. This is the
2766 // controlling input with the deepest dominating depth.
2767 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
2768 while (worklist.size() != 0) {
2769 // Use local variables nstack_top_n & nstack_top_i to cache values
2770 // on nstack's top.
2771 Node *nstack_top_n = worklist.pop();
2772 uint nstack_top_i = 0;
2773 //while_nstack_nonempty:
2774 while (true) {
2775 // Get parent node and next input's index from stack's top.
2776 Node *n = nstack_top_n;
2777 uint i = nstack_top_i;
2778 uint cnt = n->req(); // Count of inputs
2779 if (i == 0) { // Pre-process the node.
2780 if( has_node(n) && // Have either loop or control already?
2781 !has_ctrl(n) ) { // Have loop picked out already?
2782 // During "merge_many_backedges" we fold up several nested loops
2783 // into a single loop. This makes the members of the original
2784 // loop bodies pointing to dead loops; they need to move up
2785 // to the new UNION'd larger loop. I set the _head field of these
2786 // dead loops to NULL and the _parent field points to the owning
2787 // loop. Shades of UNION-FIND algorithm.
2788 IdealLoopTree *ilt;
2789 while( !(ilt = get_loop(n))->_head ) {
2790 // Normally I would use a set_loop here. But in this one special
2791 // case, it is legal (and expected) to change what loop a Node
2792 // belongs to.
2793 _nodes.map(n->_idx, (Node*)(ilt->_parent) );
2794 }
2795 // Remove safepoints ONLY if I've already seen I don't need one.
2796 // (the old code here would yank a 2nd safepoint after seeing a
2797 // first one, even though the 1st did not dominate in the loop body
2798 // and thus could be avoided indefinitely)
2799 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
2800 is_deleteable_safept(n)) {
2801 Node *in = n->in(TypeFunc::Control);
2802 lazy_replace(n,in); // Pull safepoint now
2803 // Carry on with the recursion "as if" we are walking
2804 // only the control input
2805 if( !visited.test_set( in->_idx ) ) {
2806 worklist.push(in); // Visit this guy later, using worklist
2807 }
2808 // Get next node from nstack:
2809 // - skip n's inputs processing by setting i > cnt;
2810 // - we also will not call set_early_ctrl(n) since
2811 // has_node(n) == true (see the condition above).
2812 i = cnt + 1;
2813 }
2814 }
2815 } // if (i == 0)
2817 // Visit all inputs
2818 bool done = true; // Assume all n's inputs will be processed
2819 while (i < cnt) {
2820 Node *in = n->in(i);
2821 ++i;
2822 if (in == NULL) continue;
2823 if (in->pinned() && !in->is_CFG())
2824 set_ctrl(in, in->in(0));
2825 int is_visited = visited.test_set( in->_idx );
2826 if (!has_node(in)) { // No controlling input yet?
2827 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
2828 assert( !is_visited, "visit only once" );
2829 nstack.push(n, i); // Save parent node and next input's index.
2830 nstack_top_n = in; // Process current input now.
2831 nstack_top_i = 0;
2832 done = false; // Not all n's inputs processed.
2833 break; // continue while_nstack_nonempty;
2834 } else if (!is_visited) {
2835 // This guy has a location picked out for him, but has not yet
2836 // been visited. Happens to all CFG nodes, for instance.
2837 // Visit him using the worklist instead of recursion, to break
2838 // cycles. Since he has a location already we do not need to
2839 // find his location before proceeding with the current Node.
2840 worklist.push(in); // Visit this guy later, using worklist
2841 }
2842 }
2843 if (done) {
2844 // All of n's inputs have been processed, complete post-processing.
2846 // Compute earliest point this Node can go.
2847 // CFG, Phi, pinned nodes already know their controlling input.
2848 if (!has_node(n)) {
2849 // Record earliest legal location
2850 set_early_ctrl( n );
2851 }
2852 if (nstack.is_empty()) {
2853 // Finished all nodes on stack.
2854 // Process next node on the worklist.
2855 break;
2856 }
2857 // Get saved parent node and next input's index.
2858 nstack_top_n = nstack.node();
2859 nstack_top_i = nstack.index();
2860 nstack.pop();
2861 }
2862 } // while (true)
2863 }
2864 }
2866 //------------------------------dom_lca_internal--------------------------------
2867 // Pair-wise LCA
2868 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
2869 if( !n1 ) return n2; // Handle NULL original LCA
2870 assert( n1->is_CFG(), "" );
2871 assert( n2->is_CFG(), "" );
2872 // find LCA of all uses
2873 uint d1 = dom_depth(n1);
2874 uint d2 = dom_depth(n2);
2875 while (n1 != n2) {
2876 if (d1 > d2) {
2877 n1 = idom(n1);
2878 d1 = dom_depth(n1);
2879 } else if (d1 < d2) {
2880 n2 = idom(n2);
2881 d2 = dom_depth(n2);
2882 } else {
2883 // Here d1 == d2. Due to edits of the dominator-tree, sections
2884 // of the tree might have the same depth. These sections have
2885 // to be searched more carefully.
2887 // Scan up all the n1's with equal depth, looking for n2.
2888 Node *t1 = idom(n1);
2889 while (dom_depth(t1) == d1) {
2890 if (t1 == n2) return n2;
2891 t1 = idom(t1);
2892 }
2893 // Scan up all the n2's with equal depth, looking for n1.
2894 Node *t2 = idom(n2);
2895 while (dom_depth(t2) == d2) {
2896 if (t2 == n1) return n1;
2897 t2 = idom(t2);
2898 }
2899 // Move up to a new dominator-depth value as well as up the dom-tree.
2900 n1 = t1;
2901 n2 = t2;
2902 d1 = dom_depth(n1);
2903 d2 = dom_depth(n2);
2904 }
2905 }
2906 return n1;
2907 }
2909 //------------------------------compute_idom-----------------------------------
2910 // Locally compute IDOM using dom_lca call. Correct only if the incoming
2911 // IDOMs are correct.
2912 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
2913 assert( region->is_Region(), "" );
2914 Node *LCA = NULL;
2915 for( uint i = 1; i < region->req(); i++ ) {
2916 if( region->in(i) != C->top() )
2917 LCA = dom_lca( LCA, region->in(i) );
2918 }
2919 return LCA;
2920 }
2922 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
2923 bool had_error = false;
2924 #ifdef ASSERT
2925 if (early != C->root()) {
2926 // Make sure that there's a dominance path from use to LCA
2927 Node* d = use;
2928 while (d != LCA) {
2929 d = idom(d);
2930 if (d == C->root()) {
2931 tty->print_cr("*** Use %d isn't dominated by def %s", use->_idx, n->_idx);
2932 n->dump();
2933 use->dump();
2934 had_error = true;
2935 break;
2936 }
2937 }
2938 }
2939 #endif
2940 return had_error;
2941 }
2944 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
2945 // Compute LCA over list of uses
2946 bool had_error = false;
2947 Node *LCA = NULL;
2948 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
2949 Node* c = n->fast_out(i);
2950 if (_nodes[c->_idx] == NULL)
2951 continue; // Skip the occasional dead node
2952 if( c->is_Phi() ) { // For Phis, we must land above on the path
2953 for( uint j=1; j<c->req(); j++ ) {// For all inputs
2954 if( c->in(j) == n ) { // Found matching input?
2955 Node *use = c->in(0)->in(j);
2956 if (_verify_only && use->is_top()) continue;
2957 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2958 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2959 }
2960 }
2961 } else {
2962 // For CFG data-users, use is in the block just prior
2963 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
2964 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
2965 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
2966 }
2967 }
2968 assert(!had_error, "bad dominance");
2969 return LCA;
2970 }
2972 //------------------------------get_late_ctrl----------------------------------
2973 // Compute latest legal control.
2974 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
2975 assert(early != NULL, "early control should not be NULL");
2977 Node* LCA = compute_lca_of_uses(n, early);
2978 #ifdef ASSERT
2979 if (LCA == C->root() && LCA != early) {
2980 // def doesn't dominate uses so print some useful debugging output
2981 compute_lca_of_uses(n, early, true);
2982 }
2983 #endif
2985 // if this is a load, check for anti-dependent stores
2986 // We use a conservative algorithm to identify potential interfering
2987 // instructions and for rescheduling the load. The users of the memory
2988 // input of this load are examined. Any use which is not a load and is
2989 // dominated by early is considered a potentially interfering store.
2990 // This can produce false positives.
2991 if (n->is_Load() && LCA != early) {
2992 Node_List worklist;
2994 Node *mem = n->in(MemNode::Memory);
2995 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2996 Node* s = mem->fast_out(i);
2997 worklist.push(s);
2998 }
2999 while(worklist.size() != 0 && LCA != early) {
3000 Node* s = worklist.pop();
3001 if (s->is_Load()) {
3002 continue;
3003 } else if (s->is_MergeMem()) {
3004 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3005 Node* s1 = s->fast_out(i);
3006 worklist.push(s1);
3007 }
3008 } else {
3009 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3010 assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3011 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3012 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3013 }
3014 }
3015 }
3016 }
3018 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3019 return LCA;
3020 }
3022 // true if CFG node d dominates CFG node n
3023 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3024 if (d == n)
3025 return true;
3026 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3027 uint dd = dom_depth(d);
3028 while (dom_depth(n) >= dd) {
3029 if (n == d)
3030 return true;
3031 n = idom(n);
3032 }
3033 return false;
3034 }
3036 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3037 // Pair-wise LCA with tags.
3038 // Tag each index with the node 'tag' currently being processed
3039 // before advancing up the dominator chain using idom().
3040 // Later calls that find a match to 'tag' know that this path has already
3041 // been considered in the current LCA (which is input 'n1' by convention).
3042 // Since get_late_ctrl() is only called once for each node, the tag array
3043 // does not need to be cleared between calls to get_late_ctrl().
3044 // Algorithm trades a larger constant factor for better asymptotic behavior
3045 //
3046 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3047 uint d1 = dom_depth(n1);
3048 uint d2 = dom_depth(n2);
3050 do {
3051 if (d1 > d2) {
3052 // current lca is deeper than n2
3053 _dom_lca_tags.map(n1->_idx, tag);
3054 n1 = idom(n1);
3055 d1 = dom_depth(n1);
3056 } else if (d1 < d2) {
3057 // n2 is deeper than current lca
3058 Node *memo = _dom_lca_tags[n2->_idx];
3059 if( memo == tag ) {
3060 return n1; // Return the current LCA
3061 }
3062 _dom_lca_tags.map(n2->_idx, tag);
3063 n2 = idom(n2);
3064 d2 = dom_depth(n2);
3065 } else {
3066 // Here d1 == d2. Due to edits of the dominator-tree, sections
3067 // of the tree might have the same depth. These sections have
3068 // to be searched more carefully.
3070 // Scan up all the n1's with equal depth, looking for n2.
3071 _dom_lca_tags.map(n1->_idx, tag);
3072 Node *t1 = idom(n1);
3073 while (dom_depth(t1) == d1) {
3074 if (t1 == n2) return n2;
3075 _dom_lca_tags.map(t1->_idx, tag);
3076 t1 = idom(t1);
3077 }
3078 // Scan up all the n2's with equal depth, looking for n1.
3079 _dom_lca_tags.map(n2->_idx, tag);
3080 Node *t2 = idom(n2);
3081 while (dom_depth(t2) == d2) {
3082 if (t2 == n1) return n1;
3083 _dom_lca_tags.map(t2->_idx, tag);
3084 t2 = idom(t2);
3085 }
3086 // Move up to a new dominator-depth value as well as up the dom-tree.
3087 n1 = t1;
3088 n2 = t2;
3089 d1 = dom_depth(n1);
3090 d2 = dom_depth(n2);
3091 }
3092 } while (n1 != n2);
3093 return n1;
3094 }
3096 //------------------------------init_dom_lca_tags------------------------------
3097 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3098 // Intended use does not involve any growth for the array, so it could
3099 // be of fixed size.
3100 void PhaseIdealLoop::init_dom_lca_tags() {
3101 uint limit = C->unique() + 1;
3102 _dom_lca_tags.map( limit, NULL );
3103 #ifdef ASSERT
3104 for( uint i = 0; i < limit; ++i ) {
3105 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3106 }
3107 #endif // ASSERT
3108 }
3110 //------------------------------clear_dom_lca_tags------------------------------
3111 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3112 // Intended use does not involve any growth for the array, so it could
3113 // be of fixed size.
3114 void PhaseIdealLoop::clear_dom_lca_tags() {
3115 uint limit = C->unique() + 1;
3116 _dom_lca_tags.map( limit, NULL );
3117 _dom_lca_tags.clear();
3118 #ifdef ASSERT
3119 for( uint i = 0; i < limit; ++i ) {
3120 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3121 }
3122 #endif // ASSERT
3123 }
3125 //------------------------------build_loop_late--------------------------------
3126 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3127 // Second pass finds latest legal placement, and ideal loop placement.
3128 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3129 while (worklist.size() != 0) {
3130 Node *n = worklist.pop();
3131 // Only visit once
3132 if (visited.test_set(n->_idx)) continue;
3133 uint cnt = n->outcnt();
3134 uint i = 0;
3135 while (true) {
3136 assert( _nodes[n->_idx], "no dead nodes" );
3137 // Visit all children
3138 if (i < cnt) {
3139 Node* use = n->raw_out(i);
3140 ++i;
3141 // Check for dead uses. Aggressively prune such junk. It might be
3142 // dead in the global sense, but still have local uses so I cannot
3143 // easily call 'remove_dead_node'.
3144 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3145 // Due to cycles, we might not hit the same fixed point in the verify
3146 // pass as we do in the regular pass. Instead, visit such phis as
3147 // simple uses of the loop head.
3148 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3149 if( !visited.test(use->_idx) )
3150 worklist.push(use);
3151 } else if( !visited.test_set(use->_idx) ) {
3152 nstack.push(n, i); // Save parent and next use's index.
3153 n = use; // Process all children of current use.
3154 cnt = use->outcnt();
3155 i = 0;
3156 }
3157 } else {
3158 // Do not visit around the backedge of loops via data edges.
3159 // push dead code onto a worklist
3160 _deadlist.push(use);
3161 }
3162 } else {
3163 // All of n's children have been processed, complete post-processing.
3164 build_loop_late_post(n);
3165 if (nstack.is_empty()) {
3166 // Finished all nodes on stack.
3167 // Process next node on the worklist.
3168 break;
3169 }
3170 // Get saved parent node and next use's index. Visit the rest of uses.
3171 n = nstack.node();
3172 cnt = n->outcnt();
3173 i = nstack.index();
3174 nstack.pop();
3175 }
3176 }
3177 }
3178 }
3180 //------------------------------build_loop_late_post---------------------------
3181 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3182 // Second pass finds latest legal placement, and ideal loop placement.
3183 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3185 if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress() && !_verify_only) {
3186 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
3187 }
3189 // CFG and pinned nodes already handled
3190 if( n->in(0) ) {
3191 if( n->in(0)->is_top() ) return; // Dead?
3193 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3194 // _must_ be pinned (they have to observe their control edge of course).
3195 // Unlike Stores (which modify an unallocable resource, the memory
3196 // state), Mods/Loads can float around. So free them up.
3197 bool pinned = true;
3198 switch( n->Opcode() ) {
3199 case Op_DivI:
3200 case Op_DivF:
3201 case Op_DivD:
3202 case Op_ModI:
3203 case Op_ModF:
3204 case Op_ModD:
3205 case Op_LoadB: // Same with Loads; they can sink
3206 case Op_LoadUS: // during loop optimizations.
3207 case Op_LoadD:
3208 case Op_LoadF:
3209 case Op_LoadI:
3210 case Op_LoadKlass:
3211 case Op_LoadNKlass:
3212 case Op_LoadL:
3213 case Op_LoadS:
3214 case Op_LoadP:
3215 case Op_LoadN:
3216 case Op_LoadRange:
3217 case Op_LoadD_unaligned:
3218 case Op_LoadL_unaligned:
3219 case Op_StrComp: // Does a bunch of load-like effects
3220 case Op_StrEquals:
3221 case Op_StrIndexOf:
3222 case Op_AryEq:
3223 pinned = false;
3224 }
3225 if( pinned ) {
3226 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3227 if( !chosen_loop->_child ) // Inner loop?
3228 chosen_loop->_body.push(n); // Collect inner loops
3229 return;
3230 }
3231 } else { // No slot zero
3232 if( n->is_CFG() ) { // CFG with no slot 0 is dead
3233 _nodes.map(n->_idx,0); // No block setting, it's globally dead
3234 return;
3235 }
3236 assert(!n->is_CFG() || n->outcnt() == 0, "");
3237 }
3239 // Do I have a "safe range" I can select over?
3240 Node *early = get_ctrl(n);// Early location already computed
3242 // Compute latest point this Node can go
3243 Node *LCA = get_late_ctrl( n, early );
3244 // LCA is NULL due to uses being dead
3245 if( LCA == NULL ) {
3246 #ifdef ASSERT
3247 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3248 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3249 }
3250 #endif
3251 _nodes.map(n->_idx, 0); // This node is useless
3252 _deadlist.push(n);
3253 return;
3254 }
3255 assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3257 Node *legal = LCA; // Walk 'legal' up the IDOM chain
3258 Node *least = legal; // Best legal position so far
3259 while( early != legal ) { // While not at earliest legal
3260 #ifdef ASSERT
3261 if (legal->is_Start() && !early->is_Root()) {
3262 // Bad graph. Print idom path and fail.
3263 tty->print_cr( "Bad graph detected in build_loop_late");
3264 tty->print("n: ");n->dump(); tty->cr();
3265 tty->print("early: ");early->dump(); tty->cr();
3266 int ct = 0;
3267 Node *dbg_legal = LCA;
3268 while(!dbg_legal->is_Start() && ct < 100) {
3269 tty->print("idom[%d] ",ct); dbg_legal->dump(); tty->cr();
3270 ct++;
3271 dbg_legal = idom(dbg_legal);
3272 }
3273 assert(false, "Bad graph detected in build_loop_late");
3274 }
3275 #endif
3276 // Find least loop nesting depth
3277 legal = idom(legal); // Bump up the IDOM tree
3278 // Check for lower nesting depth
3279 if( get_loop(legal)->_nest < get_loop(least)->_nest )
3280 least = legal;
3281 }
3282 assert(early == legal || legal != C->root(), "bad dominance of inputs");
3284 // Try not to place code on a loop entry projection
3285 // which can inhibit range check elimination.
3286 if (least != early) {
3287 Node* ctrl_out = least->unique_ctrl_out();
3288 if (ctrl_out && ctrl_out->is_CountedLoop() &&
3289 least == ctrl_out->in(LoopNode::EntryControl)) {
3290 Node* least_dom = idom(least);
3291 if (get_loop(least_dom)->is_member(get_loop(least))) {
3292 least = least_dom;
3293 }
3294 }
3295 }
3297 #ifdef ASSERT
3298 // If verifying, verify that 'verify_me' has a legal location
3299 // and choose it as our location.
3300 if( _verify_me ) {
3301 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3302 Node *legal = LCA;
3303 while( early != legal ) { // While not at earliest legal
3304 if( legal == v_ctrl ) break; // Check for prior good location
3305 legal = idom(legal) ;// Bump up the IDOM tree
3306 }
3307 // Check for prior good location
3308 if( legal == v_ctrl ) least = legal; // Keep prior if found
3309 }
3310 #endif
3312 // Assign discovered "here or above" point
3313 least = find_non_split_ctrl(least);
3314 set_ctrl(n, least);
3316 // Collect inner loop bodies
3317 IdealLoopTree *chosen_loop = get_loop(least);
3318 if( !chosen_loop->_child ) // Inner loop?
3319 chosen_loop->_body.push(n);// Collect inner loops
3320 }
3322 #ifndef PRODUCT
3323 //------------------------------dump-------------------------------------------
3324 void PhaseIdealLoop::dump( ) const {
3325 ResourceMark rm;
3326 Arena* arena = Thread::current()->resource_area();
3327 Node_Stack stack(arena, C->unique() >> 2);
3328 Node_List rpo_list;
3329 VectorSet visited(arena);
3330 visited.set(C->top()->_idx);
3331 rpo( C->root(), stack, visited, rpo_list );
3332 // Dump root loop indexed by last element in PO order
3333 dump( _ltree_root, rpo_list.size(), rpo_list );
3334 }
3336 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3337 loop->dump_head();
3339 // Now scan for CFG nodes in the same loop
3340 for( uint j=idx; j > 0; j-- ) {
3341 Node *n = rpo_list[j-1];
3342 if( !_nodes[n->_idx] ) // Skip dead nodes
3343 continue;
3344 if( get_loop(n) != loop ) { // Wrong loop nest
3345 if( get_loop(n)->_head == n && // Found nested loop?
3346 get_loop(n)->_parent == loop )
3347 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly
3348 continue;
3349 }
3351 // Dump controlling node
3352 for( uint x = 0; x < loop->_nest; x++ )
3353 tty->print(" ");
3354 tty->print("C");
3355 if( n == C->root() ) {
3356 n->dump();
3357 } else {
3358 Node* cached_idom = idom_no_update(n);
3359 Node *computed_idom = n->in(0);
3360 if( n->is_Region() ) {
3361 computed_idom = compute_idom(n);
3362 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3363 // any MultiBranch ctrl node), so apply a similar transform to
3364 // the cached idom returned from idom_no_update.
3365 cached_idom = find_non_split_ctrl(cached_idom);
3366 }
3367 tty->print(" ID:%d",computed_idom->_idx);
3368 n->dump();
3369 if( cached_idom != computed_idom ) {
3370 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
3371 computed_idom->_idx, cached_idom->_idx);
3372 }
3373 }
3374 // Dump nodes it controls
3375 for( uint k = 0; k < _nodes.Size(); k++ ) {
3376 // (k < C->unique() && get_ctrl(find(k)) == n)
3377 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3378 Node *m = C->root()->find(k);
3379 if( m && m->outcnt() > 0 ) {
3380 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3381 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
3382 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3383 }
3384 for( uint j = 0; j < loop->_nest; j++ )
3385 tty->print(" ");
3386 tty->print(" ");
3387 m->dump();
3388 }
3389 }
3390 }
3391 }
3392 }
3394 // Collect a R-P-O for the whole CFG.
3395 // Result list is in post-order (scan backwards for RPO)
3396 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3397 stk.push(start, 0);
3398 visited.set(start->_idx);
3400 while (stk.is_nonempty()) {
3401 Node* m = stk.node();
3402 uint idx = stk.index();
3403 if (idx < m->outcnt()) {
3404 stk.set_index(idx + 1);
3405 Node* n = m->raw_out(idx);
3406 if (n->is_CFG() && !visited.test_set(n->_idx)) {
3407 stk.push(n, 0);
3408 }
3409 } else {
3410 rpo_list.push(m);
3411 stk.pop();
3412 }
3413 }
3414 }
3415 #endif
3418 //=============================================================================
3419 //------------------------------LoopTreeIterator-----------------------------------
3421 // Advance to next loop tree using a preorder, left-to-right traversal.
3422 void LoopTreeIterator::next() {
3423 assert(!done(), "must not be done.");
3424 if (_curnt->_child != NULL) {
3425 _curnt = _curnt->_child;
3426 } else if (_curnt->_next != NULL) {
3427 _curnt = _curnt->_next;
3428 } else {
3429 while (_curnt != _root && _curnt->_next == NULL) {
3430 _curnt = _curnt->_parent;
3431 }
3432 if (_curnt == _root) {
3433 _curnt = NULL;
3434 assert(done(), "must be done.");
3435 } else {
3436 assert(_curnt->_next != NULL, "must be more to do");
3437 _curnt = _curnt->_next;
3438 }
3439 }
3440 }