1 /*
2 * Copyright (c) 2000, 2022, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciTypeFlow.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/loopnode.hpp"
34 #include "opto/phaseX.hpp"
35 #include "opto/runtime.hpp"
36 #include "opto/rootnode.hpp"
37 #include "opto/subnode.hpp"
38
39 // Portions of code courtesy of Clifford Click
40
41 // Optimization - Graph Style
42
43
44 #ifndef PRODUCT
45 extern int explicit_null_checks_elided;
46 #endif
47
48 //=============================================================================
49 //------------------------------Value------------------------------------------
50 // Return a tuple for whichever arm of the IF is reachable
51 const Type* IfNode::Value(PhaseGVN* phase) const {
52 if( !in(0) ) return Type::TOP;
53 if( phase->type(in(0)) == Type::TOP )
54 return Type::TOP;
55 const Type *t = phase->type(in(1));
56 if( t == Type::TOP ) // data is undefined
57 return TypeTuple::IFNEITHER; // unreachable altogether
58 if( t == TypeInt::ZERO ) // zero, or false
59 return TypeTuple::IFFALSE; // only false branch is reachable
60 if( t == TypeInt::ONE ) // 1, or true
61 return TypeTuple::IFTRUE; // only true branch is reachable
62 assert( t == TypeInt::BOOL, "expected boolean type" );
63
64 return TypeTuple::IFBOTH; // No progress
65 }
66
67 const RegMask &IfNode::out_RegMask() const {
68 return RegMask::Empty;
69 }
70
71 //------------------------------split_if---------------------------------------
72 // Look for places where we merge constants, then test on the merged value.
73 // If the IF test will be constant folded on the path with the constant, we
74 // win by splitting the IF to before the merge point.
75 static Node* split_if(IfNode *iff, PhaseIterGVN *igvn) {
76 // I could be a lot more general here, but I'm trying to squeeze this
77 // in before the Christmas '98 break so I'm gonna be kinda restrictive
78 // on the patterns I accept. CNC
79
80 // Look for a compare of a constant and a merged value
81 Node *i1 = iff->in(1);
82 if( !i1->is_Bool() ) return NULL;
83 BoolNode *b = i1->as_Bool();
84 Node *cmp = b->in(1);
85 if( !cmp->is_Cmp() ) return NULL;
86 i1 = cmp->in(1);
87 if( i1 == NULL || !i1->is_Phi() ) return NULL;
88 PhiNode *phi = i1->as_Phi();
89 Node *con2 = cmp->in(2);
90 if( !con2->is_Con() ) return NULL;
91 // See that the merge point contains some constants
92 Node *con1=NULL;
93 uint i4;
94 for( i4 = 1; i4 < phi->req(); i4++ ) {
95 con1 = phi->in(i4);
96 if( !con1 ) return NULL; // Do not optimize partially collapsed merges
97 if( con1->is_Con() ) break; // Found a constant
98 // Also allow null-vs-not-null checks
99 const TypePtr *tp = igvn->type(con1)->isa_ptr();
100 if( tp && tp->_ptr == TypePtr::NotNull )
101 break;
102 }
103 if( i4 >= phi->req() ) return NULL; // Found no constants
104
105 igvn->C->set_has_split_ifs(true); // Has chance for split-if
106
107 // Make sure that the compare can be constant folded away
108 Node *cmp2 = cmp->clone();
109 cmp2->set_req(1,con1);
110 cmp2->set_req(2,con2);
111 const Type *t = cmp2->Value(igvn);
112 // This compare is dead, so whack it!
113 igvn->remove_dead_node(cmp2);
114 if( !t->singleton() ) return NULL;
115
116 // No intervening control, like a simple Call
117 Node* r = iff->in(0);
118 if (!r->is_Region() || r->is_Loop() || phi->region() != r || r->as_Region()->is_copy()) {
119 return NULL;
120 }
121
122 // No other users of the cmp/bool
123 if (b->outcnt() != 1 || cmp->outcnt() != 1) {
124 //tty->print_cr("many users of cmp/bool");
125 return NULL;
126 }
127
128 // Make sure we can determine where all the uses of merged values go
129 for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) {
130 Node* u = r->fast_out(j);
131 if( u == r ) continue;
132 if( u == iff ) continue;
133 if( u->outcnt() == 0 ) continue; // use is dead & ignorable
134 if( !u->is_Phi() ) {
135 /*
136 if( u->is_Start() ) {
137 tty->print_cr("Region has inlined start use");
138 } else {
139 tty->print_cr("Region has odd use");
140 u->dump(2);
141 }*/
142 return NULL;
143 }
144 if( u != phi ) {
145 // CNC - do not allow any other merged value
146 //tty->print_cr("Merging another value");
147 //u->dump(2);
148 return NULL;
149 }
150 // Make sure we can account for all Phi uses
151 for (DUIterator_Fast kmax, k = u->fast_outs(kmax); k < kmax; k++) {
152 Node* v = u->fast_out(k); // User of the phi
153 // CNC - Allow only really simple patterns.
154 // In particular I disallow AddP of the Phi, a fairly common pattern
155 if (v == cmp) continue; // The compare is OK
156 if (v->is_ConstraintCast()) {
157 // If the cast is derived from data flow edges, it may not have a control edge.
158 // If so, it should be safe to split. But follow-up code can not deal with
159 // this (l. 359). So skip.
160 if (v->in(0) == NULL) {
161 return NULL;
162 }
163 if (v->in(0)->in(0) == iff) {
164 continue; // CastPP/II of the IfNode is OK
165 }
166 }
167 // Disabled following code because I cannot tell if exactly one
168 // path dominates without a real dominator check. CNC 9/9/1999
169 //uint vop = v->Opcode();
170 //if( vop == Op_Phi ) { // Phi from another merge point might be OK
171 // Node *r = v->in(0); // Get controlling point
172 // if( !r ) return NULL; // Degraded to a copy
173 // // Find exactly one path in (either True or False doms, but not IFF)
174 // int cnt = 0;
175 // for( uint i = 1; i < r->req(); i++ )
176 // if( r->in(i) && r->in(i)->in(0) == iff )
177 // cnt++;
178 // if( cnt == 1 ) continue; // Exactly one of True or False guards Phi
179 //}
180 if( !v->is_Call() ) {
181 /*
182 if( v->Opcode() == Op_AddP ) {
183 tty->print_cr("Phi has AddP use");
184 } else if( v->Opcode() == Op_CastPP ) {
185 tty->print_cr("Phi has CastPP use");
186 } else if( v->Opcode() == Op_CastII ) {
187 tty->print_cr("Phi has CastII use");
188 } else {
189 tty->print_cr("Phi has use I cant be bothered with");
190 }
191 */
192 }
193 return NULL;
194
195 /* CNC - Cut out all the fancy acceptance tests
196 // Can we clone this use when doing the transformation?
197 // If all uses are from Phis at this merge or constants, then YES.
198 if( !v->in(0) && v != cmp ) {
199 tty->print_cr("Phi has free-floating use");
200 v->dump(2);
201 return NULL;
202 }
203 for( uint l = 1; l < v->req(); l++ ) {
204 if( (!v->in(l)->is_Phi() || v->in(l)->in(0) != r) &&
205 !v->in(l)->is_Con() ) {
206 tty->print_cr("Phi has use");
207 v->dump(2);
208 return NULL;
209 } // End of if Phi-use input is neither Phi nor Constant
210 } // End of for all inputs to Phi-use
211 */
212 } // End of for all uses of Phi
213 } // End of for all uses of Region
214
215 // Only do this if the IF node is in a sane state
216 if (iff->outcnt() != 2)
217 return NULL;
218
219 // Got a hit! Do the Mondo Hack!
220 //
221 //ABC a1c def ghi B 1 e h A C a c d f g i
222 // R - Phi - Phi - Phi Rc - Phi - Phi - Phi Rx - Phi - Phi - Phi
223 // cmp - 2 cmp - 2 cmp - 2
224 // bool bool_c bool_x
225 // if if_c if_x
226 // T F T F T F
227 // ..s.. ..t .. ..s.. ..t.. ..s.. ..t..
228 //
229 // Split the paths coming into the merge point into 2 separate groups of
230 // merges. On the left will be all the paths feeding constants into the
231 // Cmp's Phi. On the right will be the remaining paths. The Cmp's Phi
232 // will fold up into a constant; this will let the Cmp fold up as well as
233 // all the control flow. Below the original IF we have 2 control
234 // dependent regions, 's' and 't'. Now we will merge the two paths
235 // just prior to 's' and 't' from the two IFs. At least 1 path (and quite
236 // likely 2 or more) will promptly constant fold away.
237 PhaseGVN *phase = igvn;
238
239 // Make a region merging constants and a region merging the rest
240 uint req_c = 0;
241 for (uint ii = 1; ii < r->req(); ii++) {
242 if (phi->in(ii) == con1) {
243 req_c++;
244 }
245 Node* proj = PhaseIdealLoop::find_predicate(r->in(ii));
246 if (proj != NULL) {
247 // Bail out if splitting through a region with a predicate input (could
248 // also be a loop header before loop opts creates a LoopNode for it).
249 return NULL;
250 }
251 }
252
253 // If all the defs of the phi are the same constant, we already have the desired end state.
254 // Skip the split that would create empty phi and region nodes.
255 if ((r->req() - req_c) == 1) {
256 return NULL;
257 }
258
259 // At this point we know that we can apply the split if optimization. If the region is still on the worklist,
260 // we should wait until it is processed. The region might be removed which makes this optimization redundant.
261 // This also avoids the creation of dead data loops when rewiring data nodes below when a region is dying.
262 if (igvn->_worklist.member(r)) {
263 igvn->_worklist.push(iff); // retry split if later again
264 return NULL;
265 }
266
267 Node *region_c = new RegionNode(req_c + 1);
268 Node *phi_c = con1;
269 uint len = r->req();
270 Node *region_x = new RegionNode(len - req_c);
271 Node *phi_x = PhiNode::make_blank(region_x, phi);
272 for (uint i = 1, i_c = 1, i_x = 1; i < len; i++) {
273 if (phi->in(i) == con1) {
274 region_c->init_req( i_c++, r ->in(i) );
275 } else {
276 region_x->init_req( i_x, r ->in(i) );
277 phi_x ->init_req( i_x++, phi->in(i) );
278 }
279 }
280
281 // Register the new RegionNodes but do not transform them. Cannot
282 // transform until the entire Region/Phi conglomerate has been hacked
283 // as a single huge transform.
284 igvn->register_new_node_with_optimizer( region_c );
285 igvn->register_new_node_with_optimizer( region_x );
286 // Prevent the untimely death of phi_x. Currently he has no uses. He is
287 // about to get one. If this only use goes away, then phi_x will look dead.
288 // However, he will be picking up some more uses down below.
289 Node *hook = new Node(4);
290 hook->init_req(0, phi_x);
291 hook->init_req(1, phi_c);
292 phi_x = phase->transform( phi_x );
293
294 // Make the compare
295 Node *cmp_c = phase->makecon(t);
296 Node *cmp_x = cmp->clone();
297 cmp_x->set_req(1,phi_x);
298 cmp_x->set_req(2,con2);
299 cmp_x = phase->transform(cmp_x);
300 // Make the bool
301 Node *b_c = phase->transform(new BoolNode(cmp_c,b->_test._test));
302 Node *b_x = phase->transform(new BoolNode(cmp_x,b->_test._test));
303 // Make the IfNode
304 IfNode* iff_c = iff->clone()->as_If();
305 iff_c->set_req(0, region_c);
306 iff_c->set_req(1, b_c);
307 igvn->set_type_bottom(iff_c);
308 igvn->_worklist.push(iff_c);
309 hook->init_req(2, iff_c);
310
311 IfNode* iff_x = iff->clone()->as_If();
312 iff_x->set_req(0, region_x);
313 iff_x->set_req(1, b_x);
314 igvn->set_type_bottom(iff_x);
315 igvn->_worklist.push(iff_x);
316 hook->init_req(3, iff_x);
317
318 // Make the true/false arms
319 Node *iff_c_t = phase->transform(new IfTrueNode (iff_c));
320 Node *iff_c_f = phase->transform(new IfFalseNode(iff_c));
321 Node *iff_x_t = phase->transform(new IfTrueNode (iff_x));
322 Node *iff_x_f = phase->transform(new IfFalseNode(iff_x));
323
324 // Merge the TRUE paths
325 Node *region_s = new RegionNode(3);
326 igvn->_worklist.push(region_s);
327 region_s->init_req(1, iff_c_t);
328 region_s->init_req(2, iff_x_t);
329 igvn->register_new_node_with_optimizer( region_s );
330
331 // Merge the FALSE paths
332 Node *region_f = new RegionNode(3);
333 igvn->_worklist.push(region_f);
334 region_f->init_req(1, iff_c_f);
335 region_f->init_req(2, iff_x_f);
336 igvn->register_new_node_with_optimizer( region_f );
337
338 igvn->hash_delete(cmp);// Remove soon-to-be-dead node from hash table.
339 cmp->set_req(1,NULL); // Whack the inputs to cmp because it will be dead
340 cmp->set_req(2,NULL);
341 // Check for all uses of the Phi and give them a new home.
342 // The 'cmp' got cloned, but CastPP/IIs need to be moved.
343 Node *phi_s = NULL; // do not construct unless needed
344 Node *phi_f = NULL; // do not construct unless needed
345 for (DUIterator_Last i2min, i2 = phi->last_outs(i2min); i2 >= i2min; --i2) {
346 Node* v = phi->last_out(i2);// User of the phi
347 igvn->rehash_node_delayed(v); // Have to fixup other Phi users
348 uint vop = v->Opcode();
349 Node *proj = NULL;
350 if( vop == Op_Phi ) { // Remote merge point
351 Node *r = v->in(0);
352 for (uint i3 = 1; i3 < r->req(); i3++)
353 if (r->in(i3) && r->in(i3)->in(0) == iff) {
354 proj = r->in(i3);
355 break;
356 }
357 } else if( v->is_ConstraintCast() ) {
358 proj = v->in(0); // Controlling projection
359 } else {
360 assert( 0, "do not know how to handle this guy" );
361 }
362 guarantee(proj != NULL, "sanity");
363
364 Node *proj_path_data, *proj_path_ctrl;
365 if( proj->Opcode() == Op_IfTrue ) {
366 if( phi_s == NULL ) {
367 // Only construct phi_s if needed, otherwise provides
368 // interfering use.
369 phi_s = PhiNode::make_blank(region_s,phi);
370 phi_s->init_req( 1, phi_c );
371 phi_s->init_req( 2, phi_x );
372 hook->add_req(phi_s);
373 phi_s = phase->transform(phi_s);
374 }
375 proj_path_data = phi_s;
376 proj_path_ctrl = region_s;
377 } else {
378 if( phi_f == NULL ) {
379 // Only construct phi_f if needed, otherwise provides
380 // interfering use.
381 phi_f = PhiNode::make_blank(region_f,phi);
382 phi_f->init_req( 1, phi_c );
383 phi_f->init_req( 2, phi_x );
384 hook->add_req(phi_f);
385 phi_f = phase->transform(phi_f);
386 }
387 proj_path_data = phi_f;
388 proj_path_ctrl = region_f;
389 }
390
391 // Fixup 'v' for for the split
392 if( vop == Op_Phi ) { // Remote merge point
393 uint i;
394 for( i = 1; i < v->req(); i++ )
395 if( v->in(i) == phi )
396 break;
397 v->set_req(i, proj_path_data );
398 } else if( v->is_ConstraintCast() ) {
399 v->set_req(0, proj_path_ctrl );
400 v->set_req(1, proj_path_data );
401 } else
402 ShouldNotReachHere();
403 }
404
405 // Now replace the original iff's True/False with region_s/region_t.
406 // This makes the original iff go dead.
407 for (DUIterator_Last i3min, i3 = iff->last_outs(i3min); i3 >= i3min; --i3) {
408 Node* p = iff->last_out(i3);
409 assert( p->Opcode() == Op_IfTrue || p->Opcode() == Op_IfFalse, "" );
410 Node *u = (p->Opcode() == Op_IfTrue) ? region_s : region_f;
411 // Replace p with u
412 igvn->add_users_to_worklist(p);
413 for (DUIterator_Last lmin, l = p->last_outs(lmin); l >= lmin;) {
414 Node* x = p->last_out(l);
415 igvn->hash_delete(x);
416 uint uses_found = 0;
417 for( uint j = 0; j < x->req(); j++ ) {
418 if( x->in(j) == p ) {
419 x->set_req(j, u);
420 uses_found++;
421 }
422 }
423 l -= uses_found; // we deleted 1 or more copies of this edge
424 }
425 igvn->remove_dead_node(p);
426 }
427
428 // Force the original merge dead
429 igvn->hash_delete(r);
430 // First, remove region's dead users.
431 for (DUIterator_Last lmin, l = r->last_outs(lmin); l >= lmin;) {
432 Node* u = r->last_out(l);
433 if( u == r ) {
434 r->set_req(0, NULL);
435 } else {
436 assert(u->outcnt() == 0, "only dead users");
437 igvn->remove_dead_node(u);
438 }
439 l -= 1;
440 }
441 igvn->remove_dead_node(r);
442
443 // Now remove the bogus extra edges used to keep things alive
444 igvn->remove_dead_node( hook );
445
446 // Must return either the original node (now dead) or a new node
447 // (Do not return a top here, since that would break the uniqueness of top.)
448 return new ConINode(TypeInt::ZERO);
449 }
450
451 // if this IfNode follows a range check pattern return the projection
452 // for the failed path
453 ProjNode* IfNode::range_check_trap_proj(int& flip_test, Node*& l, Node*& r) {
454 if (outcnt() != 2) {
455 return NULL;
456 }
457 Node* b = in(1);
458 if (b == NULL || !b->is_Bool()) return NULL;
459 BoolNode* bn = b->as_Bool();
460 Node* cmp = bn->in(1);
461 if (cmp == NULL) return NULL;
462 if (cmp->Opcode() != Op_CmpU) return NULL;
463
464 l = cmp->in(1);
465 r = cmp->in(2);
466 flip_test = 1;
467 if (bn->_test._test == BoolTest::le) {
468 l = cmp->in(2);
469 r = cmp->in(1);
470 flip_test = 2;
471 } else if (bn->_test._test != BoolTest::lt) {
472 return NULL;
473 }
474 if (l->is_top()) return NULL; // Top input means dead test
475 if (r->Opcode() != Op_LoadRange && !is_RangeCheck()) return NULL;
476
477 // We have recognized one of these forms:
478 // Flip 1: If (Bool[<] CmpU(l, LoadRange)) ...
479 // Flip 2: If (Bool[<=] CmpU(LoadRange, l)) ...
480
481 ProjNode* iftrap = proj_out_or_null(flip_test == 2 ? true : false);
482 return iftrap;
483 }
484
485
486 //------------------------------is_range_check---------------------------------
487 // Return 0 if not a range check. Return 1 if a range check and set index and
488 // offset. Return 2 if we had to negate the test. Index is NULL if the check
489 // is versus a constant.
490 int RangeCheckNode::is_range_check(Node* &range, Node* &index, jint &offset) {
491 int flip_test = 0;
492 Node* l = NULL;
493 Node* r = NULL;
494 ProjNode* iftrap = range_check_trap_proj(flip_test, l, r);
495
496 if (iftrap == NULL) {
497 return 0;
498 }
499
500 // Make sure it's a real range check by requiring an uncommon trap
501 // along the OOB path. Otherwise, it's possible that the user wrote
502 // something which optimized to look like a range check but behaves
503 // in some other way.
504 if (iftrap->is_uncommon_trap_proj(Deoptimization::Reason_range_check) == NULL) {
505 return 0;
506 }
507
508 // Look for index+offset form
509 Node* ind = l;
510 jint off = 0;
511 if (l->is_top()) {
512 return 0;
513 } else if (l->Opcode() == Op_AddI) {
514 if ((off = l->in(1)->find_int_con(0)) != 0) {
515 ind = l->in(2)->uncast();
516 } else if ((off = l->in(2)->find_int_con(0)) != 0) {
517 ind = l->in(1)->uncast();
518 }
519 } else if ((off = l->find_int_con(-1)) >= 0) {
520 // constant offset with no variable index
521 ind = NULL;
522 } else {
523 // variable index with no constant offset (or dead negative index)
524 off = 0;
525 }
526
527 // Return all the values:
528 index = ind;
529 offset = off;
530 range = r;
531 return flip_test;
532 }
533
534 //------------------------------adjust_check-----------------------------------
535 // Adjust (widen) a prior range check
536 static void adjust_check(Node* proj, Node* range, Node* index,
537 int flip, jint off_lo, PhaseIterGVN* igvn) {
538 PhaseGVN *gvn = igvn;
539 // Break apart the old check
540 Node *iff = proj->in(0);
541 Node *bol = iff->in(1);
542 if( bol->is_top() ) return; // In case a partially dead range check appears
543 // bail (or bomb[ASSERT/DEBUG]) if NOT projection-->IfNode-->BoolNode
544 DEBUG_ONLY( if( !bol->is_Bool() ) { proj->dump(3); fatal("Expect projection-->IfNode-->BoolNode"); } )
545 if( !bol->is_Bool() ) return;
546
547 Node *cmp = bol->in(1);
548 // Compute a new check
549 Node *new_add = gvn->intcon(off_lo);
550 if( index ) {
551 new_add = off_lo ? gvn->transform(new AddINode( index, new_add )) : index;
552 }
553 Node *new_cmp = (flip == 1)
554 ? new CmpUNode( new_add, range )
555 : new CmpUNode( range, new_add );
556 new_cmp = gvn->transform(new_cmp);
557 // See if no need to adjust the existing check
558 if( new_cmp == cmp ) return;
559 // Else, adjust existing check
560 Node *new_bol = gvn->transform( new BoolNode( new_cmp, bol->as_Bool()->_test._test ) );
561 igvn->rehash_node_delayed( iff );
562 iff->set_req_X( 1, new_bol, igvn );
563 }
564
565 //------------------------------up_one_dom-------------------------------------
566 // Walk up the dominator tree one step. Return NULL at root or true
567 // complex merges. Skips through small diamonds.
568 Node* IfNode::up_one_dom(Node *curr, bool linear_only) {
569 Node *dom = curr->in(0);
570 if( !dom ) // Found a Region degraded to a copy?
571 return curr->nonnull_req(); // Skip thru it
572
573 if( curr != dom ) // Normal walk up one step?
574 return dom;
575
576 // Use linear_only if we are still parsing, since we cannot
577 // trust the regions to be fully filled in.
578 if (linear_only)
579 return NULL;
580
581 if( dom->is_Root() )
582 return NULL;
583
584 // Else hit a Region. Check for a loop header
585 if( dom->is_Loop() )
586 return dom->in(1); // Skip up thru loops
587
588 // Check for small diamonds
589 Node *din1, *din2, *din3, *din4;
590 if( dom->req() == 3 && // 2-path merge point
591 (din1 = dom ->in(1)) && // Left path exists
592 (din2 = dom ->in(2)) && // Right path exists
593 (din3 = din1->in(0)) && // Left path up one
594 (din4 = din2->in(0)) ) { // Right path up one
595 if( din3->is_Call() && // Handle a slow-path call on either arm
596 (din3 = din3->in(0)) )
597 din3 = din3->in(0);
598 if( din4->is_Call() && // Handle a slow-path call on either arm
599 (din4 = din4->in(0)) )
600 din4 = din4->in(0);
601 if (din3 != NULL && din3 == din4 && din3->is_If()) // Regions not degraded to a copy
602 return din3; // Skip around diamonds
603 }
604
605 // Give up the search at true merges
606 return NULL; // Dead loop? Or hit root?
607 }
608
609
610 //------------------------------filtered_int_type--------------------------------
611 // Return a possibly more restrictive type for val based on condition control flow for an if
612 const TypeInt* IfNode::filtered_int_type(PhaseGVN* gvn, Node* val, Node* if_proj) {
613 assert(if_proj &&
614 (if_proj->Opcode() == Op_IfTrue || if_proj->Opcode() == Op_IfFalse), "expecting an if projection");
615 if (if_proj->in(0) && if_proj->in(0)->is_If()) {
616 IfNode* iff = if_proj->in(0)->as_If();
617 if (iff->in(1) && iff->in(1)->is_Bool()) {
618 BoolNode* bol = iff->in(1)->as_Bool();
619 if (bol->in(1) && bol->in(1)->is_Cmp()) {
620 const CmpNode* cmp = bol->in(1)->as_Cmp();
621 if (cmp->in(1) == val) {
622 const TypeInt* cmp2_t = gvn->type(cmp->in(2))->isa_int();
623 if (cmp2_t != NULL) {
624 jint lo = cmp2_t->_lo;
625 jint hi = cmp2_t->_hi;
626 BoolTest::mask msk = if_proj->Opcode() == Op_IfTrue ? bol->_test._test : bol->_test.negate();
627 switch (msk) {
628 case BoolTest::ne: {
629 // If val is compared to its lower or upper bound, we can narrow the type
630 const TypeInt* val_t = gvn->type(val)->isa_int();
631 if (val_t != NULL && !val_t->singleton() && cmp2_t->is_con()) {
632 if (val_t->_lo == lo) {
633 return TypeInt::make(val_t->_lo + 1, val_t->_hi, val_t->_widen);
634 } else if (val_t->_hi == hi) {
635 return TypeInt::make(val_t->_lo, val_t->_hi - 1, val_t->_widen);
636 }
637 }
638 // Can't refine type
639 return NULL;
640 }
641 case BoolTest::eq:
642 return cmp2_t;
643 case BoolTest::lt:
644 lo = TypeInt::INT->_lo;
645 if (hi != min_jint) {
646 hi = hi - 1;
647 }
648 break;
649 case BoolTest::le:
650 lo = TypeInt::INT->_lo;
651 break;
652 case BoolTest::gt:
653 if (lo != max_jint) {
654 lo = lo + 1;
655 }
656 hi = TypeInt::INT->_hi;
657 break;
658 case BoolTest::ge:
659 // lo unchanged
660 hi = TypeInt::INT->_hi;
661 break;
662 default:
663 break;
664 }
665 const TypeInt* rtn_t = TypeInt::make(lo, hi, cmp2_t->_widen);
666 return rtn_t;
667 }
668 }
669 }
670 }
671 }
672 return NULL;
673 }
674
675 //------------------------------fold_compares----------------------------
676 // See if a pair of CmpIs can be converted into a CmpU. In some cases
677 // the direction of this if is determined by the preceding if so it
678 // can be eliminate entirely.
679 //
680 // Given an if testing (CmpI n v) check for an immediately control
681 // dependent if that is testing (CmpI n v2) and has one projection
682 // leading to this if and the other projection leading to a region
683 // that merges one of this ifs control projections.
684 //
685 // If
686 // / |
687 // / |
688 // / |
689 // If |
690 // /\ |
691 // / \ |
692 // / \ |
693 // / Region
694 //
695 // Or given an if testing (CmpI n v) check for a dominating if that is
696 // testing (CmpI n v2), both having one projection leading to an
697 // uncommon trap. Allow Another independent guard in between to cover
698 // an explicit range check:
699 // if (index < 0 || index >= array.length) {
700 // which may need a null check to guard the LoadRange
701 //
702 // If
703 // / \
704 // / \
705 // / \
706 // If unc
707 // /\
708 // / \
709 // / \
710 // / unc
711 //
712
713 // Is the comparison for this If suitable for folding?
714 bool IfNode::cmpi_folds(PhaseIterGVN* igvn, bool fold_ne) {
715 return in(1) != NULL &&
716 in(1)->is_Bool() &&
717 in(1)->in(1) != NULL &&
718 in(1)->in(1)->Opcode() == Op_CmpI &&
719 in(1)->in(1)->in(2) != NULL &&
720 in(1)->in(1)->in(2) != igvn->C->top() &&
721 (in(1)->as_Bool()->_test.is_less() ||
722 in(1)->as_Bool()->_test.is_greater() ||
723 (fold_ne && in(1)->as_Bool()->_test._test == BoolTest::ne));
724 }
725
726 // Is a dominating control suitable for folding with this if?
727 bool IfNode::is_ctrl_folds(Node* ctrl, PhaseIterGVN* igvn) {
728 return ctrl != NULL &&
729 ctrl->is_Proj() &&
730 ctrl->in(0) != NULL &&
731 ctrl->in(0)->Opcode() == Op_If &&
732 ctrl->in(0)->outcnt() == 2 &&
733 ctrl->in(0)->as_If()->cmpi_folds(igvn, true) &&
734 // Must compare same value
735 ctrl->in(0)->in(1)->in(1)->in(1) != NULL &&
736 ctrl->in(0)->in(1)->in(1)->in(1) == in(1)->in(1)->in(1);
737 }
738
739 // Do this If and the dominating If share a region?
740 bool IfNode::has_shared_region(ProjNode* proj, ProjNode*& success, ProjNode*& fail) {
741 ProjNode* otherproj = proj->other_if_proj();
742 Node* otherproj_ctrl_use = otherproj->unique_ctrl_out_or_null();
743 RegionNode* region = (otherproj_ctrl_use != NULL && otherproj_ctrl_use->is_Region()) ? otherproj_ctrl_use->as_Region() : NULL;
744 success = NULL;
745 fail = NULL;
746
747 if (otherproj->outcnt() == 1 && region != NULL && !region->has_phi()) {
748 for (int i = 0; i < 2; i++) {
749 ProjNode* proj = proj_out(i);
750 if (success == NULL && proj->outcnt() == 1 && proj->unique_out() == region) {
751 success = proj;
752 } else if (fail == NULL) {
753 fail = proj;
754 } else {
755 success = fail = NULL;
756 }
757 }
758 }
759 return success != NULL && fail != NULL;
760 }
761
762 bool IfNode::is_dominator_unc(CallStaticJavaNode* dom_unc, CallStaticJavaNode* unc) {
763 // Different methods and methods containing jsrs are not supported.
764 ciMethod* method = unc->jvms()->method();
765 ciMethod* dom_method = dom_unc->jvms()->method();
766 if (method != dom_method || method->has_jsrs()) {
767 return false;
768 }
769 // Check that both traps are in the same activation of the method (instead
770 // of two activations being inlined through different call sites) by verifying
771 // that the call stacks are equal for both JVMStates.
772 JVMState* dom_caller = dom_unc->jvms()->caller();
773 JVMState* caller = unc->jvms()->caller();
774 if ((dom_caller == NULL) != (caller == NULL)) {
775 // The current method must either be inlined into both dom_caller and
776 // caller or must not be inlined at all (top method). Bail out otherwise.
777 return false;
778 } else if (dom_caller != NULL && !dom_caller->same_calls_as(caller)) {
779 return false;
780 }
781 // Check that the bci of the dominating uncommon trap dominates the bci
782 // of the dominated uncommon trap. Otherwise we may not re-execute
783 // the dominated check after deoptimization from the merged uncommon trap.
784 ciTypeFlow* flow = dom_method->get_flow_analysis();
785 int bci = unc->jvms()->bci();
786 int dom_bci = dom_unc->jvms()->bci();
787 if (!flow->is_dominated_by(bci, dom_bci)) {
788 return false;
789 }
790
791 return true;
792 }
793
794 // Return projection that leads to an uncommon trap if any
795 ProjNode* IfNode::uncommon_trap_proj(CallStaticJavaNode*& call) const {
796 for (int i = 0; i < 2; i++) {
797 call = proj_out(i)->is_uncommon_trap_proj(Deoptimization::Reason_none);
798 if (call != NULL) {
799 return proj_out(i);
800 }
801 }
802 return NULL;
803 }
804
805 // Do this If and the dominating If both branch out to an uncommon trap
806 bool IfNode::has_only_uncommon_traps(ProjNode* proj, ProjNode*& success, ProjNode*& fail, PhaseIterGVN* igvn) {
807 ProjNode* otherproj = proj->other_if_proj();
808 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(Deoptimization::Reason_none);
809
810 if (otherproj->outcnt() == 1 && dom_unc != NULL) {
811 // We need to re-execute the folded Ifs after deoptimization from the merged traps
812 if (!dom_unc->jvms()->should_reexecute()) {
813 return false;
814 }
815
816 CallStaticJavaNode* unc = NULL;
817 ProjNode* unc_proj = uncommon_trap_proj(unc);
818 if (unc_proj != NULL && unc_proj->outcnt() == 1) {
819 if (dom_unc == unc) {
820 // Allow the uncommon trap to be shared through a region
821 RegionNode* r = unc->in(0)->as_Region();
822 if (r->outcnt() != 2 || r->req() != 3 || r->find_edge(otherproj) == -1 || r->find_edge(unc_proj) == -1) {
823 return false;
824 }
825 assert(r->has_phi() == NULL, "simple region shouldn't have a phi");
826 } else if (dom_unc->in(0) != otherproj || unc->in(0) != unc_proj) {
827 return false;
828 }
829
830 if (!is_dominator_unc(dom_unc, unc)) {
831 return false;
832 }
833
834 // See merge_uncommon_traps: the reason of the uncommon trap
835 // will be changed and the state of the dominating If will be
836 // used. Checked that we didn't apply this transformation in a
837 // previous compilation and it didn't cause too many traps
838 ciMethod* dom_method = dom_unc->jvms()->method();
839 int dom_bci = dom_unc->jvms()->bci();
840 if (!igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_unstable_fused_if) &&
841 !igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_range_check)) {
842 success = unc_proj;
843 fail = unc_proj->other_if_proj();
844 return true;
845 }
846 }
847 }
848 return false;
849 }
850
851 // Check that the 2 CmpI can be folded into as single CmpU and proceed with the folding
852 bool IfNode::fold_compares_helper(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) {
853 Node* this_cmp = in(1)->in(1);
854 BoolNode* this_bool = in(1)->as_Bool();
855 IfNode* dom_iff = proj->in(0)->as_If();
856 BoolNode* dom_bool = dom_iff->in(1)->as_Bool();
857 Node* lo = dom_iff->in(1)->in(1)->in(2);
858 Node* hi = this_cmp->in(2);
859 Node* n = this_cmp->in(1);
860 ProjNode* otherproj = proj->other_if_proj();
861
862 const TypeInt* lo_type = IfNode::filtered_int_type(igvn, n, otherproj);
863 const TypeInt* hi_type = IfNode::filtered_int_type(igvn, n, success);
864
865 BoolTest::mask lo_test = dom_bool->_test._test;
866 BoolTest::mask hi_test = this_bool->_test._test;
867 BoolTest::mask cond = hi_test;
868
869 // convert:
870 //
871 // dom_bool = x {<,<=,>,>=} a
872 // / \
873 // proj = {True,False} / \ otherproj = {False,True}
874 // /
875 // this_bool = x {<,<=} b
876 // / \
877 // fail = {True,False} / \ success = {False,True}
878 // /
879 //
880 // (Second test guaranteed canonicalized, first one may not have
881 // been canonicalized yet)
882 //
883 // into:
884 //
885 // cond = (x - lo) {<u,<=u,>u,>=u} adjusted_lim
886 // / \
887 // fail / \ success
888 // /
889 //
890
891 // Figure out which of the two tests sets the upper bound and which
892 // sets the lower bound if any.
893 Node* adjusted_lim = NULL;
894 if (lo_type != NULL && hi_type != NULL && hi_type->_lo > lo_type->_hi &&
895 hi_type->_hi == max_jint && lo_type->_lo == min_jint && lo_test != BoolTest::ne) {
896 assert((dom_bool->_test.is_less() && !proj->_con) ||
897 (dom_bool->_test.is_greater() && proj->_con), "incorrect test");
898
899 // this_bool = <
900 // dom_bool = >= (proj = True) or dom_bool = < (proj = False)
901 // x in [a, b[ on the fail (= True) projection, b > a-1 (because of hi_type->_lo > lo_type->_hi test above):
902 // lo = a, hi = b, adjusted_lim = b-a, cond = <u
903 // dom_bool = > (proj = True) or dom_bool = <= (proj = False)
904 // x in ]a, b[ on the fail (= True) projection, b > a:
905 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <u
906 // this_bool = <=
907 // dom_bool = >= (proj = True) or dom_bool = < (proj = False)
908 // x in [a, b] on the fail (= True) projection, b+1 > a-1:
909 // lo = a, hi = b, adjusted_lim = b-a+1, cond = <u
910 // lo = a, hi = b, adjusted_lim = b-a, cond = <=u doesn't work because b = a - 1 is possible, then b-a = -1
911 // dom_bool = > (proj = True) or dom_bool = <= (proj = False)
912 // x in ]a, b] on the fail (= True) projection b+1 > a:
913 // lo = a+1, hi = b, adjusted_lim = b-a, cond = <u
914 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <=u doesn't work because a = b is possible, then b-a-1 = -1
915
916 if (hi_test == BoolTest::lt) {
917 if (lo_test == BoolTest::gt || lo_test == BoolTest::le) {
918 lo = igvn->transform(new AddINode(lo, igvn->intcon(1)));
919 }
920 } else if (hi_test == BoolTest::le) {
921 if (lo_test == BoolTest::ge || lo_test == BoolTest::lt) {
922 adjusted_lim = igvn->transform(new SubINode(hi, lo));
923 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1)));
924 cond = BoolTest::lt;
925 } else if (lo_test == BoolTest::gt || lo_test == BoolTest::le) {
926 adjusted_lim = igvn->transform(new SubINode(hi, lo));
927 lo = igvn->transform(new AddINode(lo, igvn->intcon(1)));
928 cond = BoolTest::lt;
929 } else {
930 assert(false, "unhandled lo_test: %d", lo_test);
931 return false;
932 }
933 } else {
934 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled hi_test: %d", hi_test);
935 return false;
936 }
937 // this test was canonicalized
938 assert(this_bool->_test.is_less() && fail->_con, "incorrect test");
939 } else if (lo_type != NULL && hi_type != NULL && lo_type->_lo > hi_type->_hi &&
940 lo_type->_hi == max_jint && hi_type->_lo == min_jint && lo_test != BoolTest::ne) {
941
942 // this_bool = <
943 // dom_bool = < (proj = True) or dom_bool = >= (proj = False)
944 // x in [b, a[ on the fail (= False) projection, a > b-1 (because of lo_type->_lo > hi_type->_hi above):
945 // lo = b, hi = a, adjusted_lim = a-b, cond = >=u
946 // dom_bool = <= (proj = True) or dom_bool = > (proj = False)
947 // x in [b, a] on the fail (= False) projection, a+1 > b-1:
948 // lo = b, hi = a, adjusted_lim = a-b+1, cond = >=u
949 // lo = b, hi = a, adjusted_lim = a-b, cond = >u doesn't work because a = b - 1 is possible, then b-a = -1
950 // this_bool = <=
951 // dom_bool = < (proj = True) or dom_bool = >= (proj = False)
952 // x in ]b, a[ on the fail (= False) projection, a > b:
953 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >=u
954 // dom_bool = <= (proj = True) or dom_bool = > (proj = False)
955 // x in ]b, a] on the fail (= False) projection, a+1 > b:
956 // lo = b+1, hi = a, adjusted_lim = a-b, cond = >=u
957 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >u doesn't work because a = b is possible, then b-a-1 = -1
958
959 swap(lo, hi);
960 swap(lo_type, hi_type);
961 swap(lo_test, hi_test);
962
963 assert((dom_bool->_test.is_less() && proj->_con) ||
964 (dom_bool->_test.is_greater() && !proj->_con), "incorrect test");
965
966 cond = (hi_test == BoolTest::le || hi_test == BoolTest::gt) ? BoolTest::gt : BoolTest::ge;
967
968 if (lo_test == BoolTest::lt) {
969 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) {
970 cond = BoolTest::ge;
971 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) {
972 adjusted_lim = igvn->transform(new SubINode(hi, lo));
973 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1)));
974 cond = BoolTest::ge;
975 } else {
976 assert(false, "unhandled hi_test: %d", hi_test);
977 return false;
978 }
979 } else if (lo_test == BoolTest::le) {
980 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) {
981 lo = igvn->transform(new AddINode(lo, igvn->intcon(1)));
982 cond = BoolTest::ge;
983 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) {
984 adjusted_lim = igvn->transform(new SubINode(hi, lo));
985 lo = igvn->transform(new AddINode(lo, igvn->intcon(1)));
986 cond = BoolTest::ge;
987 } else {
988 assert(false, "unhandled hi_test: %d", hi_test);
989 return false;
990 }
991 } else {
992 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled lo_test: %d", lo_test);
993 return false;
994 }
995 // this test was canonicalized
996 assert(this_bool->_test.is_less() && !fail->_con, "incorrect test");
997 } else {
998 const TypeInt* failtype = filtered_int_type(igvn, n, proj);
999 if (failtype != NULL) {
1000 const TypeInt* type2 = filtered_int_type(igvn, n, fail);
1001 if (type2 != NULL) {
1002 failtype = failtype->join(type2)->is_int();
1003 if (failtype->_lo > failtype->_hi) {
1004 // previous if determines the result of this if so
1005 // replace Bool with constant
1006 igvn->_worklist.push(in(1));
1007 igvn->replace_input_of(this, 1, igvn->intcon(success->_con));
1008 return true;
1009 }
1010 }
1011 }
1012 lo = NULL;
1013 hi = NULL;
1014 }
1015
1016 if (lo && hi) {
1017 Node* hook = new Node(1);
1018 hook->init_req(0, lo); // Add a use to lo to prevent him from dying
1019 // Merge the two compares into a single unsigned compare by building (CmpU (n - lo) (hi - lo))
1020 Node* adjusted_val = igvn->transform(new SubINode(n, lo));
1021 if (adjusted_lim == NULL) {
1022 adjusted_lim = igvn->transform(new SubINode(hi, lo));
1023 }
1024 hook->destruct(igvn);
1025
1026 int lo = igvn->type(adjusted_lim)->is_int()->_lo;
1027 if (lo < 0) {
1028 // If range check elimination applies to this comparison, it includes code to protect from overflows that may
1029 // cause the main loop to be skipped entirely. Delay this transformation.
1030 // Example:
1031 // for (int i = 0; i < limit; i++) {
1032 // if (i < max_jint && i > min_jint) {...
1033 // }
1034 // Comparisons folded as:
1035 // i - min_jint - 1 <u -2
1036 // when RC applies, main loop limit becomes:
1037 // min(limit, max(-2 + min_jint + 1, min_jint))
1038 // = min(limit, min_jint)
1039 // = min_jint
1040 if (!igvn->C->post_loop_opts_phase()) {
1041 if (adjusted_val->outcnt() == 0) {
1042 igvn->remove_dead_node(adjusted_val);
1043 }
1044 if (adjusted_lim->outcnt() == 0) {
1045 igvn->remove_dead_node(adjusted_lim);
1046 }
1047 igvn->C->record_for_post_loop_opts_igvn(this);
1048 return false;
1049 }
1050 }
1051
1052 Node* newcmp = igvn->transform(new CmpUNode(adjusted_val, adjusted_lim));
1053 Node* newbool = igvn->transform(new BoolNode(newcmp, cond));
1054
1055 igvn->replace_input_of(dom_iff, 1, igvn->intcon(proj->_con));
1056 igvn->_worklist.push(in(1));
1057 igvn->replace_input_of(this, 1, newbool);
1058
1059 return true;
1060 }
1061 return false;
1062 }
1063
1064 // Merge the branches that trap for this If and the dominating If into
1065 // a single region that branches to the uncommon trap for the
1066 // dominating If
1067 Node* IfNode::merge_uncommon_traps(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) {
1068 Node* res = this;
1069 assert(success->in(0) == this, "bad projection");
1070
1071 ProjNode* otherproj = proj->other_if_proj();
1072
1073 CallStaticJavaNode* unc = success->is_uncommon_trap_proj(Deoptimization::Reason_none);
1074 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(Deoptimization::Reason_none);
1075
1076 if (unc != dom_unc) {
1077 Node* r = new RegionNode(3);
1078
1079 r->set_req(1, otherproj);
1080 r->set_req(2, success);
1081 r = igvn->transform(r);
1082 assert(r->is_Region(), "can't go away");
1083
1084 // Make both If trap at the state of the first If: once the CmpI
1085 // nodes are merged, if we trap we don't know which of the CmpI
1086 // nodes would have caused the trap so we have to restart
1087 // execution at the first one
1088 igvn->replace_input_of(dom_unc, 0, r);
1089 igvn->replace_input_of(unc, 0, igvn->C->top());
1090 }
1091 int trap_request = dom_unc->uncommon_trap_request();
1092 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1093 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1094
1095 int flip_test = 0;
1096 Node* l = NULL;
1097 Node* r = NULL;
1098
1099 if (success->in(0)->as_If()->range_check_trap_proj(flip_test, l, r) != NULL) {
1100 // If this looks like a range check, change the trap to
1101 // Reason_range_check so the compiler recognizes it as a range
1102 // check and applies the corresponding optimizations
1103 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_range_check, action);
1104
1105 improve_address_types(l, r, fail, igvn);
1106
1107 res = igvn->transform(new RangeCheckNode(in(0), in(1), _prob, _fcnt));
1108 } else if (unc != dom_unc) {
1109 // If we trap we won't know what CmpI would have caused the trap
1110 // so use a special trap reason to mark this pair of CmpI nodes as
1111 // bad candidate for folding. On recompilation we won't fold them
1112 // and we may trap again but this time we'll know what branch
1113 // traps
1114 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_unstable_fused_if, action);
1115 }
1116 igvn->replace_input_of(dom_unc, TypeFunc::Parms, igvn->intcon(trap_request));
1117 return res;
1118 }
1119
1120 // If we are turning 2 CmpI nodes into a CmpU that follows the pattern
1121 // of a rangecheck on index i, on 64 bit the compares may be followed
1122 // by memory accesses using i as index. In that case, the CmpU tells
1123 // us something about the values taken by i that can help the compiler
1124 // (see Compile::conv_I2X_index())
1125 void IfNode::improve_address_types(Node* l, Node* r, ProjNode* fail, PhaseIterGVN* igvn) {
1126 #ifdef _LP64
1127 ResourceMark rm;
1128 Node_Stack stack(2);
1129
1130 assert(r->Opcode() == Op_LoadRange, "unexpected range check");
1131 const TypeInt* array_size = igvn->type(r)->is_int();
1132
1133 stack.push(l, 0);
1134
1135 while(stack.size() > 0) {
1136 Node* n = stack.node();
1137 uint start = stack.index();
1138
1139 uint i = start;
1140 for (; i < n->outcnt(); i++) {
1141 Node* use = n->raw_out(i);
1142 if (stack.size() == 1) {
1143 if (use->Opcode() == Op_ConvI2L) {
1144 const TypeLong* bounds = use->as_Type()->type()->is_long();
1145 if (bounds->_lo <= array_size->_lo && bounds->_hi >= array_size->_hi &&
1146 (bounds->_lo != array_size->_lo || bounds->_hi != array_size->_hi)) {
1147 stack.set_index(i+1);
1148 stack.push(use, 0);
1149 break;
1150 }
1151 }
1152 } else if (use->is_Mem()) {
1153 Node* ctrl = use->in(0);
1154 for (int i = 0; i < 10 && ctrl != NULL && ctrl != fail; i++) {
1155 ctrl = up_one_dom(ctrl);
1156 }
1157 if (ctrl == fail) {
1158 Node* init_n = stack.node_at(1);
1159 assert(init_n->Opcode() == Op_ConvI2L, "unexpected first node");
1160 // Create a new narrow ConvI2L node that is dependent on the range check
1161 Node* new_n = igvn->C->conv_I2X_index(igvn, l, array_size, fail);
1162
1163 // The type of the ConvI2L may be widen and so the new
1164 // ConvI2L may not be better than an existing ConvI2L
1165 if (new_n != init_n) {
1166 for (uint j = 2; j < stack.size(); j++) {
1167 Node* n = stack.node_at(j);
1168 Node* clone = n->clone();
1169 int rep = clone->replace_edge(init_n, new_n, igvn);
1170 assert(rep > 0, "can't find expected node?");
1171 clone = igvn->transform(clone);
1172 init_n = n;
1173 new_n = clone;
1174 }
1175 igvn->hash_delete(use);
1176 int rep = use->replace_edge(init_n, new_n, igvn);
1177 assert(rep > 0, "can't find expected node?");
1178 igvn->transform(use);
1179 if (init_n->outcnt() == 0) {
1180 igvn->_worklist.push(init_n);
1181 }
1182 }
1183 }
1184 } else if (use->in(0) == NULL && (igvn->type(use)->isa_long() ||
1185 igvn->type(use)->isa_ptr())) {
1186 stack.set_index(i+1);
1187 stack.push(use, 0);
1188 break;
1189 }
1190 }
1191 if (i == n->outcnt()) {
1192 stack.pop();
1193 }
1194 }
1195 #endif
1196 }
1197
1198 bool IfNode::is_cmp_with_loadrange(ProjNode* proj) {
1199 if (in(1) != NULL &&
1200 in(1)->in(1) != NULL &&
1201 in(1)->in(1)->in(2) != NULL) {
1202 Node* other = in(1)->in(1)->in(2);
1203 if (other->Opcode() == Op_LoadRange &&
1204 ((other->in(0) != NULL && other->in(0) == proj) ||
1205 (other->in(0) == NULL &&
1206 other->in(2) != NULL &&
1207 other->in(2)->is_AddP() &&
1208 other->in(2)->in(1) != NULL &&
1209 other->in(2)->in(1)->Opcode() == Op_CastPP &&
1210 other->in(2)->in(1)->in(0) == proj))) {
1211 return true;
1212 }
1213 }
1214 return false;
1215 }
1216
1217 bool IfNode::is_null_check(ProjNode* proj, PhaseIterGVN* igvn) {
1218 Node* other = in(1)->in(1)->in(2);
1219 if (other->in(MemNode::Address) != NULL &&
1220 proj->in(0)->in(1) != NULL &&
1221 proj->in(0)->in(1)->is_Bool() &&
1222 proj->in(0)->in(1)->in(1) != NULL &&
1223 proj->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1224 proj->in(0)->in(1)->in(1)->in(2) != NULL &&
1225 proj->in(0)->in(1)->in(1)->in(1) == other->in(MemNode::Address)->in(AddPNode::Address)->uncast() &&
1226 igvn->type(proj->in(0)->in(1)->in(1)->in(2)) == TypePtr::NULL_PTR) {
1227 return true;
1228 }
1229 return false;
1230 }
1231
1232 // Check that the If that is in between the 2 integer comparisons has
1233 // no side effect
1234 bool IfNode::is_side_effect_free_test(ProjNode* proj, PhaseIterGVN* igvn) {
1235 if (proj == NULL) {
1236 return false;
1237 }
1238 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1239 if (unc != NULL && proj->outcnt() <= 2) {
1240 if (proj->outcnt() == 1 ||
1241 // Allow simple null check from LoadRange
1242 (is_cmp_with_loadrange(proj) && is_null_check(proj, igvn))) {
1243 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1244 CallStaticJavaNode* dom_unc = proj->in(0)->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1245 assert(dom_unc != NULL, "is_uncommon_trap_if_pattern returned NULL");
1246
1247 // reroute_side_effect_free_unc changes the state of this
1248 // uncommon trap to restart execution at the previous
1249 // CmpI. Check that this change in a previous compilation didn't
1250 // cause too many traps.
1251 int trap_request = unc->uncommon_trap_request();
1252 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1253
1254 if (igvn->C->too_many_traps(dom_unc->jvms()->method(), dom_unc->jvms()->bci(), reason)) {
1255 return false;
1256 }
1257
1258 if (!is_dominator_unc(dom_unc, unc)) {
1259 return false;
1260 }
1261
1262 return true;
1263 }
1264 }
1265 return false;
1266 }
1267
1268 // Make the If between the 2 integer comparisons trap at the state of
1269 // the first If: the last CmpI is the one replaced by a CmpU and the
1270 // first CmpI is eliminated, so the test between the 2 CmpI nodes
1271 // won't be guarded by the first CmpI anymore. It can trap in cases
1272 // where the first CmpI would have prevented it from executing: on a
1273 // trap, we need to restart execution at the state of the first CmpI
1274 void IfNode::reroute_side_effect_free_unc(ProjNode* proj, ProjNode* dom_proj, PhaseIterGVN* igvn) {
1275 CallStaticJavaNode* dom_unc = dom_proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1276 ProjNode* otherproj = proj->other_if_proj();
1277 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1278 Node* call_proj = dom_unc->unique_ctrl_out();
1279 Node* halt = call_proj->unique_ctrl_out();
1280
1281 Node* new_unc = dom_unc->clone();
1282 call_proj = call_proj->clone();
1283 halt = halt->clone();
1284 Node* c = otherproj->clone();
1285
1286 c = igvn->transform(c);
1287 new_unc->set_req(TypeFunc::Parms, unc->in(TypeFunc::Parms));
1288 new_unc->set_req(0, c);
1289 new_unc = igvn->transform(new_unc);
1290 call_proj->set_req(0, new_unc);
1291 call_proj = igvn->transform(call_proj);
1292 halt->set_req(0, call_proj);
1293 halt = igvn->transform(halt);
1294
1295 igvn->replace_node(otherproj, igvn->C->top());
1296 igvn->C->root()->add_req(halt);
1297 }
1298
1299 Node* IfNode::fold_compares(PhaseIterGVN* igvn) {
1300 if (Opcode() != Op_If) return NULL;
1301
1302 if (cmpi_folds(igvn)) {
1303 Node* ctrl = in(0);
1304 if (is_ctrl_folds(ctrl, igvn) && ctrl->outcnt() == 1) {
1305 // A integer comparison immediately dominated by another integer
1306 // comparison
1307 ProjNode* success = NULL;
1308 ProjNode* fail = NULL;
1309 ProjNode* dom_cmp = ctrl->as_Proj();
1310 if (has_shared_region(dom_cmp, success, fail) &&
1311 // Next call modifies graph so must be last
1312 fold_compares_helper(dom_cmp, success, fail, igvn)) {
1313 return this;
1314 }
1315 if (has_only_uncommon_traps(dom_cmp, success, fail, igvn) &&
1316 // Next call modifies graph so must be last
1317 fold_compares_helper(dom_cmp, success, fail, igvn)) {
1318 return merge_uncommon_traps(dom_cmp, success, fail, igvn);
1319 }
1320 return NULL;
1321 } else if (ctrl->in(0) != NULL &&
1322 ctrl->in(0)->in(0) != NULL) {
1323 ProjNode* success = NULL;
1324 ProjNode* fail = NULL;
1325 Node* dom = ctrl->in(0)->in(0);
1326 ProjNode* dom_cmp = dom->isa_Proj();
1327 ProjNode* other_cmp = ctrl->isa_Proj();
1328
1329 // Check if it's an integer comparison dominated by another
1330 // integer comparison with another test in between
1331 if (is_ctrl_folds(dom, igvn) &&
1332 has_only_uncommon_traps(dom_cmp, success, fail, igvn) &&
1333 is_side_effect_free_test(other_cmp, igvn) &&
1334 // Next call modifies graph so must be last
1335 fold_compares_helper(dom_cmp, success, fail, igvn)) {
1336 reroute_side_effect_free_unc(other_cmp, dom_cmp, igvn);
1337 return merge_uncommon_traps(dom_cmp, success, fail, igvn);
1338 }
1339 }
1340 }
1341 return NULL;
1342 }
1343
1344 //------------------------------remove_useless_bool----------------------------
1345 // Check for people making a useless boolean: things like
1346 // if( (x < y ? true : false) ) { ... }
1347 // Replace with if( x < y ) { ... }
1348 static Node *remove_useless_bool(IfNode *iff, PhaseGVN *phase) {
1349 Node *i1 = iff->in(1);
1350 if( !i1->is_Bool() ) return NULL;
1351 BoolNode *bol = i1->as_Bool();
1352
1353 Node *cmp = bol->in(1);
1354 if( cmp->Opcode() != Op_CmpI ) return NULL;
1355
1356 // Must be comparing against a bool
1357 const Type *cmp2_t = phase->type( cmp->in(2) );
1358 if( cmp2_t != TypeInt::ZERO &&
1359 cmp2_t != TypeInt::ONE )
1360 return NULL;
1361
1362 // Find a prior merge point merging the boolean
1363 i1 = cmp->in(1);
1364 if( !i1->is_Phi() ) return NULL;
1365 PhiNode *phi = i1->as_Phi();
1366 if( phase->type( phi ) != TypeInt::BOOL )
1367 return NULL;
1368
1369 // Check for diamond pattern
1370 int true_path = phi->is_diamond_phi();
1371 if( true_path == 0 ) return NULL;
1372
1373 // Make sure that iff and the control of the phi are different. This
1374 // should really only happen for dead control flow since it requires
1375 // an illegal cycle.
1376 if (phi->in(0)->in(1)->in(0) == iff) return NULL;
1377
1378 // phi->region->if_proj->ifnode->bool->cmp
1379 BoolNode *bol2 = phi->in(0)->in(1)->in(0)->in(1)->as_Bool();
1380
1381 // Now get the 'sense' of the test correct so we can plug in
1382 // either iff2->in(1) or its complement.
1383 int flip = 0;
1384 if( bol->_test._test == BoolTest::ne ) flip = 1-flip;
1385 else if( bol->_test._test != BoolTest::eq ) return NULL;
1386 if( cmp2_t == TypeInt::ZERO ) flip = 1-flip;
1387
1388 const Type *phi1_t = phase->type( phi->in(1) );
1389 const Type *phi2_t = phase->type( phi->in(2) );
1390 // Check for Phi(0,1) and flip
1391 if( phi1_t == TypeInt::ZERO ) {
1392 if( phi2_t != TypeInt::ONE ) return NULL;
1393 flip = 1-flip;
1394 } else {
1395 // Check for Phi(1,0)
1396 if( phi1_t != TypeInt::ONE ) return NULL;
1397 if( phi2_t != TypeInt::ZERO ) return NULL;
1398 }
1399 if( true_path == 2 ) {
1400 flip = 1-flip;
1401 }
1402
1403 Node* new_bol = (flip ? phase->transform( bol2->negate(phase) ) : bol2);
1404 assert(new_bol != iff->in(1), "must make progress");
1405 iff->set_req_X(1, new_bol, phase);
1406 // Intervening diamond probably goes dead
1407 phase->C->set_major_progress();
1408 return iff;
1409 }
1410
1411 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff);
1412
1413 struct RangeCheck {
1414 Node* ctl;
1415 jint off;
1416 };
1417
1418 Node* IfNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
1419 if (remove_dead_region(phase, can_reshape)) return this;
1420 // No Def-Use info?
1421 if (!can_reshape) return NULL;
1422
1423 // Don't bother trying to transform a dead if
1424 if (in(0)->is_top()) return NULL;
1425 // Don't bother trying to transform an if with a dead test
1426 if (in(1)->is_top()) return NULL;
1427 // Another variation of a dead test
1428 if (in(1)->is_Con()) return NULL;
1429 // Another variation of a dead if
1430 if (outcnt() < 2) return NULL;
1431
1432 // Canonicalize the test.
1433 Node* idt_if = idealize_test(phase, this);
1434 if (idt_if != NULL) return idt_if;
1435
1436 // Try to split the IF
1437 PhaseIterGVN *igvn = phase->is_IterGVN();
1438 Node *s = split_if(this, igvn);
1439 if (s != NULL) return s;
1440
1441 return NodeSentinel;
1442 }
1443
1444 //------------------------------Ideal------------------------------------------
1445 // Return a node which is more "ideal" than the current node. Strip out
1446 // control copies
1447 Node* IfNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1448 Node* res = Ideal_common(phase, can_reshape);
1449 if (res != NodeSentinel) {
1450 return res;
1451 }
1452
1453 // Check for people making a useless boolean: things like
1454 // if( (x < y ? true : false) ) { ... }
1455 // Replace with if( x < y ) { ... }
1456 Node* bol2 = remove_useless_bool(this, phase);
1457 if (bol2) return bol2;
1458
1459 if (in(0) == NULL) return NULL; // Dead loop?
1460
1461 PhaseIterGVN* igvn = phase->is_IterGVN();
1462 Node* result = fold_compares(igvn);
1463 if (result != NULL) {
1464 return result;
1465 }
1466
1467 // Scan for an equivalent test
1468 int dist = 4; // Cutoff limit for search
1469 if (is_If() && in(1)->is_Bool()) {
1470 Node* cmp = in(1)->in(1);
1471 if (cmp->Opcode() == Op_CmpP &&
1472 cmp->in(2) != NULL && // make sure cmp is not already dead
1473 cmp->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1474 dist = 64; // Limit for null-pointer scans
1475 }
1476 }
1477
1478 Node* prev_dom = search_identical(dist);
1479
1480 if (prev_dom != NULL) {
1481 // Replace dominated IfNode
1482 return dominated_by(prev_dom, igvn);
1483 }
1484
1485 return simple_subsuming(igvn);
1486 }
1487
1488 //------------------------------dominated_by-----------------------------------
1489 Node* IfNode::dominated_by(Node* prev_dom, PhaseIterGVN *igvn) {
1490 #ifndef PRODUCT
1491 if (TraceIterativeGVN) {
1492 tty->print(" Removing IfNode: "); this->dump();
1493 }
1494 #endif
1495
1496 igvn->hash_delete(this); // Remove self to prevent spurious V-N
1497 Node *idom = in(0);
1498 // Need opcode to decide which way 'this' test goes
1499 int prev_op = prev_dom->Opcode();
1500 Node *top = igvn->C->top(); // Shortcut to top
1501
1502 // Loop predicates may have depending checks which should not
1503 // be skipped. For example, range check predicate has two checks
1504 // for lower and upper bounds.
1505 ProjNode* unc_proj = proj_out(1 - prev_dom->as_Proj()->_con)->as_Proj();
1506 if (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
1507 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL) {
1508 prev_dom = idom;
1509 }
1510
1511 // Now walk the current IfNode's projections.
1512 // Loop ends when 'this' has no more uses.
1513 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) {
1514 Node *ifp = last_out(i); // Get IfTrue/IfFalse
1515 igvn->add_users_to_worklist(ifp);
1516 // Check which projection it is and set target.
1517 // Data-target is either the dominating projection of the same type
1518 // or TOP if the dominating projection is of opposite type.
1519 // Data-target will be used as the new control edge for the non-CFG
1520 // nodes like Casts and Loads.
1521 Node *data_target = (ifp->Opcode() == prev_op) ? prev_dom : top;
1522 // Control-target is just the If's immediate dominator or TOP.
1523 Node *ctrl_target = (ifp->Opcode() == prev_op) ? idom : top;
1524
1525 // For each child of an IfTrue/IfFalse projection, reroute.
1526 // Loop ends when projection has no more uses.
1527 for (DUIterator_Last jmin, j = ifp->last_outs(jmin); j >= jmin; --j) {
1528 Node* s = ifp->last_out(j); // Get child of IfTrue/IfFalse
1529 if (s->depends_only_on_test() && igvn->no_dependent_zero_check(s)) {
1530 // For control producers.
1531 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
1532 igvn->replace_input_of(s, 0, data_target); // Move child to data-target
1533 } else {
1534 // Find the control input matching this def-use edge.
1535 // For Regions it may not be in slot 0.
1536 uint l;
1537 for (l = 0; s->in(l) != ifp; l++) { }
1538 igvn->replace_input_of(s, l, ctrl_target);
1539 }
1540 } // End for each child of a projection
1541
1542 igvn->remove_dead_node(ifp);
1543 } // End for each IfTrue/IfFalse child of If
1544
1545 // Kill the IfNode
1546 igvn->remove_dead_node(this);
1547
1548 // Must return either the original node (now dead) or a new node
1549 // (Do not return a top here, since that would break the uniqueness of top.)
1550 return new ConINode(TypeInt::ZERO);
1551 }
1552
1553 Node* IfNode::search_identical(int dist) {
1554 // Setup to scan up the CFG looking for a dominating test
1555 Node* dom = in(0);
1556 Node* prev_dom = this;
1557 int op = Opcode();
1558 // Search up the dominator tree for an If with an identical test
1559 while (dom->Opcode() != op || // Not same opcode?
1560 dom->in(1) != in(1) || // Not same input 1?
1561 prev_dom->in(0) != dom) { // One path of test does not dominate?
1562 if (dist < 0) return NULL;
1563
1564 dist--;
1565 prev_dom = dom;
1566 dom = up_one_dom(dom);
1567 if (!dom) return NULL;
1568 }
1569
1570 // Check that we did not follow a loop back to ourselves
1571 if (this == dom) {
1572 return NULL;
1573 }
1574
1575 #ifndef PRODUCT
1576 if (dist > 2) { // Add to count of NULL checks elided
1577 explicit_null_checks_elided++;
1578 }
1579 #endif
1580
1581 return prev_dom;
1582 }
1583
1584
1585 static int subsuming_bool_test_encode(Node*);
1586
1587 // Check if dominating test is subsuming 'this' one.
1588 //
1589 // cmp
1590 // / \
1591 // (r1) bool \
1592 // / bool (r2)
1593 // (dom) if \
1594 // \ )
1595 // (pre) if[TF] /
1596 // \ /
1597 // if (this)
1598 // \r1
1599 // r2\ eqT eqF neT neF ltT ltF leT leF gtT gtF geT geF
1600 // eq t f f t f - - f f - - f
1601 // ne f t t f t - - t t - - t
1602 // lt f - - f t f - f f - f t
1603 // le t - - t t - t f f t - t
1604 // gt f - - f f - f t t f - f
1605 // ge t - - t f t - t t - t f
1606 //
1607 Node* IfNode::simple_subsuming(PhaseIterGVN* igvn) {
1608 // Table encoding: N/A (na), True-branch (tb), False-branch (fb).
1609 static enum { na, tb, fb } s_short_circuit_map[6][12] = {
1610 /*rel: eq+T eq+F ne+T ne+F lt+T lt+F le+T le+F gt+T gt+F ge+T ge+F*/
1611 /*eq*/{ tb, fb, fb, tb, fb, na, na, fb, fb, na, na, fb },
1612 /*ne*/{ fb, tb, tb, fb, tb, na, na, tb, tb, na, na, tb },
1613 /*lt*/{ fb, na, na, fb, tb, fb, na, fb, fb, na, fb, tb },
1614 /*le*/{ tb, na, na, tb, tb, na, tb, fb, fb, tb, na, tb },
1615 /*gt*/{ fb, na, na, fb, fb, na, fb, tb, tb, fb, na, fb },
1616 /*ge*/{ tb, na, na, tb, fb, tb, na, tb, tb, na, tb, fb }};
1617
1618 Node* pre = in(0);
1619 if (!pre->is_IfTrue() && !pre->is_IfFalse()) {
1620 return NULL;
1621 }
1622 Node* dom = pre->in(0);
1623 if (!dom->is_If()) {
1624 return NULL;
1625 }
1626 Node* bol = in(1);
1627 if (!bol->is_Bool()) {
1628 return NULL;
1629 }
1630 Node* cmp = in(1)->in(1);
1631 if (!cmp->is_Cmp()) {
1632 return NULL;
1633 }
1634
1635 if (!dom->in(1)->is_Bool()) {
1636 return NULL;
1637 }
1638 if (dom->in(1)->in(1) != cmp) { // Not same cond?
1639 return NULL;
1640 }
1641
1642 int drel = subsuming_bool_test_encode(dom->in(1));
1643 int trel = subsuming_bool_test_encode(bol);
1644 int bout = pre->is_IfFalse() ? 1 : 0;
1645
1646 if (drel < 0 || trel < 0) {
1647 return NULL;
1648 }
1649 int br = s_short_circuit_map[trel][2*drel+bout];
1650 if (br == na) {
1651 return NULL;
1652 }
1653 #ifndef PRODUCT
1654 if (TraceIterativeGVN) {
1655 tty->print(" Subsumed IfNode: "); dump();
1656 }
1657 #endif
1658 // Replace condition with constant True(1)/False(0).
1659 bool is_always_true = br == tb;
1660 set_req(1, igvn->intcon(is_always_true ? 1 : 0));
1661
1662 // Update any data dependencies to the directly dominating test. This subsumed test is not immediately removed by igvn
1663 // and therefore subsequent optimizations might miss these data dependencies otherwise. There might be a dead loop
1664 // ('always_taken_proj' == 'pre') that is cleaned up later. Skip this case to make the iterator work properly.
1665 Node* always_taken_proj = proj_out(is_always_true);
1666 if (always_taken_proj != pre) {
1667 for (DUIterator_Fast imax, i = always_taken_proj->fast_outs(imax); i < imax; i++) {
1668 Node* u = always_taken_proj->fast_out(i);
1669 if (!u->is_CFG()) {
1670 igvn->replace_input_of(u, 0, pre);
1671 --i;
1672 --imax;
1673 }
1674 }
1675 }
1676
1677 if (bol->outcnt() == 0) {
1678 igvn->remove_dead_node(bol); // Kill the BoolNode.
1679 }
1680 return this;
1681 }
1682
1683 // Map BoolTest to local table encoding. The BoolTest (e)numerals
1684 // { eq = 0, ne = 4, le = 5, ge = 7, lt = 3, gt = 1 }
1685 // are mapped to table indices, while the remaining (e)numerals in BoolTest
1686 // { overflow = 2, no_overflow = 6, never = 8, illegal = 9 }
1687 // are ignored (these are not modeled in the table).
1688 //
1689 static int subsuming_bool_test_encode(Node* node) {
1690 precond(node->is_Bool());
1691 BoolTest::mask x = node->as_Bool()->_test._test;
1692 switch (x) {
1693 case BoolTest::eq: return 0;
1694 case BoolTest::ne: return 1;
1695 case BoolTest::lt: return 2;
1696 case BoolTest::le: return 3;
1697 case BoolTest::gt: return 4;
1698 case BoolTest::ge: return 5;
1699 case BoolTest::overflow:
1700 case BoolTest::no_overflow:
1701 case BoolTest::never:
1702 case BoolTest::illegal:
1703 default:
1704 return -1;
1705 }
1706 }
1707
1708 //------------------------------Identity---------------------------------------
1709 // If the test is constant & we match, then we are the input Control
1710 Node* IfProjNode::Identity(PhaseGVN* phase) {
1711 // Can only optimize if cannot go the other way
1712 const TypeTuple *t = phase->type(in(0))->is_tuple();
1713 if (t == TypeTuple::IFNEITHER || (always_taken(t) &&
1714 // During parsing (GVN) we don't remove dead code aggressively.
1715 // Cut off dead branch and let PhaseRemoveUseless take care of it.
1716 (!phase->is_IterGVN() ||
1717 // During IGVN, first wait for the dead branch to be killed.
1718 // Otherwise, the IfNode's control will have two control uses (the IfNode
1719 // that doesn't go away because it still has uses and this branch of the
1720 // If) which breaks other optimizations. Node::has_special_unique_user()
1721 // will cause this node to be reprocessed once the dead branch is killed.
1722 in(0)->outcnt() == 1))) {
1723 // IfNode control
1724 if (in(0)->is_BaseCountedLoopEnd()) {
1725 // CountedLoopEndNode may be eliminated by if subsuming, replace CountedLoopNode with LoopNode to
1726 // avoid mismatching between CountedLoopNode and CountedLoopEndNode in the following optimization.
1727 Node* head = unique_ctrl_out_or_null();
1728 if (head != NULL && head->is_BaseCountedLoop() && head->in(LoopNode::LoopBackControl) == this) {
1729 Node* new_head = new LoopNode(head->in(LoopNode::EntryControl), this);
1730 phase->is_IterGVN()->register_new_node_with_optimizer(new_head);
1731 phase->is_IterGVN()->replace_node(head, new_head);
1732 }
1733 }
1734 return in(0)->in(0);
1735 }
1736 // no progress
1737 return this;
1738 }
1739
1740 #ifndef PRODUCT
1741 //-------------------------------related---------------------------------------
1742 // An IfProjNode's related node set consists of its input (an IfNode) including
1743 // the IfNode's condition, plus all of its outputs at level 1. In compact mode,
1744 // the restrictions for IfNode apply (see IfNode::rel).
1745 void IfProjNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
1746 Node* ifNode = this->in(0);
1747 in_rel->append(ifNode);
1748 if (compact) {
1749 ifNode->collect_nodes(in_rel, 3, false, true);
1750 } else {
1751 ifNode->collect_nodes_in_all_data(in_rel, false);
1752 }
1753 this->collect_nodes(out_rel, -1, false, false);
1754 }
1755
1756 //------------------------------dump_spec--------------------------------------
1757 void IfNode::dump_spec(outputStream *st) const {
1758 st->print("P=%f, C=%f",_prob,_fcnt);
1759 }
1760
1761 //-------------------------------related---------------------------------------
1762 // For an IfNode, the set of related output nodes is just the output nodes till
1763 // depth 2, i.e, the IfTrue/IfFalse projection nodes plus the nodes they refer.
1764 // The related input nodes contain no control nodes, but all data nodes
1765 // pertaining to the condition. In compact mode, the input nodes are collected
1766 // up to a depth of 3.
1767 void IfNode::related(GrowableArray <Node *> *in_rel, GrowableArray <Node *> *out_rel, bool compact) const {
1768 if (compact) {
1769 this->collect_nodes(in_rel, 3, false, true);
1770 } else {
1771 this->collect_nodes_in_all_data(in_rel, false);
1772 }
1773 this->collect_nodes(out_rel, -2, false, false);
1774 }
1775 #endif
1776
1777 //------------------------------idealize_test----------------------------------
1778 // Try to canonicalize tests better. Peek at the Cmp/Bool/If sequence and
1779 // come up with a canonical sequence. Bools getting 'eq', 'gt' and 'ge' forms
1780 // converted to 'ne', 'le' and 'lt' forms. IfTrue/IfFalse get swapped as
1781 // needed.
1782 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff) {
1783 assert(iff->in(0) != NULL, "If must be live");
1784
1785 if (iff->outcnt() != 2) return NULL; // Malformed projections.
1786 Node* old_if_f = iff->proj_out(false);
1787 Node* old_if_t = iff->proj_out(true);
1788
1789 // CountedLoopEnds want the back-control test to be TRUE, irregardless of
1790 // whether they are testing a 'gt' or 'lt' condition. The 'gt' condition
1791 // happens in count-down loops
1792 if (iff->is_BaseCountedLoopEnd()) return NULL;
1793 if (!iff->in(1)->is_Bool()) return NULL; // Happens for partially optimized IF tests
1794 BoolNode *b = iff->in(1)->as_Bool();
1795 BoolTest bt = b->_test;
1796 // Test already in good order?
1797 if( bt.is_canonical() )
1798 return NULL;
1799
1800 // Flip test to be canonical. Requires flipping the IfFalse/IfTrue and
1801 // cloning the IfNode.
1802 Node* new_b = phase->transform( new BoolNode(b->in(1), bt.negate()) );
1803 if( !new_b->is_Bool() ) return NULL;
1804 b = new_b->as_Bool();
1805
1806 PhaseIterGVN *igvn = phase->is_IterGVN();
1807 assert( igvn, "Test is not canonical in parser?" );
1808
1809 // The IF node never really changes, but it needs to be cloned
1810 iff = iff->clone()->as_If();
1811 iff->set_req(1, b);
1812 iff->_prob = 1.0-iff->_prob;
1813
1814 Node *prior = igvn->hash_find_insert(iff);
1815 if( prior ) {
1816 igvn->remove_dead_node(iff);
1817 iff = (IfNode*)prior;
1818 } else {
1819 // Cannot call transform on it just yet
1820 igvn->set_type_bottom(iff);
1821 }
1822 igvn->_worklist.push(iff);
1823
1824 // Now handle projections. Cloning not required.
1825 Node* new_if_f = (Node*)(new IfFalseNode( iff ));
1826 Node* new_if_t = (Node*)(new IfTrueNode ( iff ));
1827
1828 igvn->register_new_node_with_optimizer(new_if_f);
1829 igvn->register_new_node_with_optimizer(new_if_t);
1830 // Flip test, so flip trailing control
1831 igvn->replace_node(old_if_f, new_if_t);
1832 igvn->replace_node(old_if_t, new_if_f);
1833
1834 // Progress
1835 return iff;
1836 }
1837
1838 Node* RangeCheckNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1839 Node* res = Ideal_common(phase, can_reshape);
1840 if (res != NodeSentinel) {
1841 return res;
1842 }
1843
1844 PhaseIterGVN *igvn = phase->is_IterGVN();
1845 // Setup to scan up the CFG looking for a dominating test
1846 Node* prev_dom = this;
1847
1848 // Check for range-check vs other kinds of tests
1849 Node* index1;
1850 Node* range1;
1851 jint offset1;
1852 int flip1 = is_range_check(range1, index1, offset1);
1853 if (flip1) {
1854 Node* dom = in(0);
1855 // Try to remove extra range checks. All 'up_one_dom' gives up at merges
1856 // so all checks we inspect post-dominate the top-most check we find.
1857 // If we are going to fail the current check and we reach the top check
1858 // then we are guaranteed to fail, so just start interpreting there.
1859 // We 'expand' the top 3 range checks to include all post-dominating
1860 // checks.
1861
1862 // The top 3 range checks seen
1863 const int NRC = 3;
1864 RangeCheck prev_checks[NRC];
1865 int nb_checks = 0;
1866
1867 // Low and high offsets seen so far
1868 jint off_lo = offset1;
1869 jint off_hi = offset1;
1870
1871 bool found_immediate_dominator = false;
1872
1873 // Scan for the top checks and collect range of offsets
1874 for (int dist = 0; dist < 999; dist++) { // Range-Check scan limit
1875 if (dom->Opcode() == Op_RangeCheck && // Not same opcode?
1876 prev_dom->in(0) == dom) { // One path of test does dominate?
1877 if (dom == this) return NULL; // dead loop
1878 // See if this is a range check
1879 Node* index2;
1880 Node* range2;
1881 jint offset2;
1882 int flip2 = dom->as_RangeCheck()->is_range_check(range2, index2, offset2);
1883 // See if this is a _matching_ range check, checking against
1884 // the same array bounds.
1885 if (flip2 == flip1 && range2 == range1 && index2 == index1 &&
1886 dom->outcnt() == 2) {
1887 if (nb_checks == 0 && dom->in(1) == in(1)) {
1888 // Found an immediately dominating test at the same offset.
1889 // This kind of back-to-back test can be eliminated locally,
1890 // and there is no need to search further for dominating tests.
1891 assert(offset2 == offset1, "Same test but different offsets");
1892 found_immediate_dominator = true;
1893 break;
1894 }
1895 // Gather expanded bounds
1896 off_lo = MIN2(off_lo,offset2);
1897 off_hi = MAX2(off_hi,offset2);
1898 // Record top NRC range checks
1899 prev_checks[nb_checks%NRC].ctl = prev_dom;
1900 prev_checks[nb_checks%NRC].off = offset2;
1901 nb_checks++;
1902 }
1903 }
1904 prev_dom = dom;
1905 dom = up_one_dom(dom);
1906 if (!dom) break;
1907 }
1908
1909 if (!found_immediate_dominator) {
1910 // Attempt to widen the dominating range check to cover some later
1911 // ones. Since range checks "fail" by uncommon-trapping to the
1912 // interpreter, widening a check can make us speculatively enter
1913 // the interpreter. If we see range-check deopt's, do not widen!
1914 if (!phase->C->allow_range_check_smearing()) return NULL;
1915
1916 // Didn't find prior covering check, so cannot remove anything.
1917 if (nb_checks == 0) {
1918 return NULL;
1919 }
1920 // Constant indices only need to check the upper bound.
1921 // Non-constant indices must check both low and high.
1922 int chk0 = (nb_checks - 1) % NRC;
1923 if (index1) {
1924 if (nb_checks == 1) {
1925 return NULL;
1926 } else {
1927 // If the top range check's constant is the min or max of
1928 // all constants we widen the next one to cover the whole
1929 // range of constants.
1930 RangeCheck rc0 = prev_checks[chk0];
1931 int chk1 = (nb_checks - 2) % NRC;
1932 RangeCheck rc1 = prev_checks[chk1];
1933 if (rc0.off == off_lo) {
1934 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn);
1935 prev_dom = rc1.ctl;
1936 } else if (rc0.off == off_hi) {
1937 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn);
1938 prev_dom = rc1.ctl;
1939 } else {
1940 // If the top test's constant is not the min or max of all
1941 // constants, we need 3 range checks. We must leave the
1942 // top test unchanged because widening it would allow the
1943 // accesses it protects to successfully read/write out of
1944 // bounds.
1945 if (nb_checks == 2) {
1946 return NULL;
1947 }
1948 int chk2 = (nb_checks - 3) % NRC;
1949 RangeCheck rc2 = prev_checks[chk2];
1950 // The top range check a+i covers interval: -a <= i < length-a
1951 // The second range check b+i covers interval: -b <= i < length-b
1952 if (rc1.off <= rc0.off) {
1953 // if b <= a, we change the second range check to:
1954 // -min_of_all_constants <= i < length-min_of_all_constants
1955 // Together top and second range checks now cover:
1956 // -min_of_all_constants <= i < length-a
1957 // which is more restrictive than -b <= i < length-b:
1958 // -b <= -min_of_all_constants <= i < length-a <= length-b
1959 // The third check is then changed to:
1960 // -max_of_all_constants <= i < length-max_of_all_constants
1961 // so 2nd and 3rd checks restrict allowed values of i to:
1962 // -min_of_all_constants <= i < length-max_of_all_constants
1963 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn);
1964 adjust_check(rc2.ctl, range1, index1, flip1, off_hi, igvn);
1965 } else {
1966 // if b > a, we change the second range check to:
1967 // -max_of_all_constants <= i < length-max_of_all_constants
1968 // Together top and second range checks now cover:
1969 // -a <= i < length-max_of_all_constants
1970 // which is more restrictive than -b <= i < length-b:
1971 // -b < -a <= i < length-max_of_all_constants <= length-b
1972 // The third check is then changed to:
1973 // -max_of_all_constants <= i < length-max_of_all_constants
1974 // so 2nd and 3rd checks restrict allowed values of i to:
1975 // -min_of_all_constants <= i < length-max_of_all_constants
1976 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn);
1977 adjust_check(rc2.ctl, range1, index1, flip1, off_lo, igvn);
1978 }
1979 prev_dom = rc2.ctl;
1980 }
1981 }
1982 } else {
1983 RangeCheck rc0 = prev_checks[chk0];
1984 // 'Widen' the offset of the 1st and only covering check
1985 adjust_check(rc0.ctl, range1, index1, flip1, off_hi, igvn);
1986 // Test is now covered by prior checks, dominate it out
1987 prev_dom = rc0.ctl;
1988 }
1989 }
1990 } else {
1991 prev_dom = search_identical(4);
1992
1993 if (prev_dom == NULL) {
1994 return NULL;
1995 }
1996 }
1997
1998 // Replace dominated IfNode
1999 return dominated_by(prev_dom, igvn);
2000 }