1 /*
2 * Copyright (c) 2000, 2023, 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/predicates.hpp"
36 #include "opto/runtime.hpp"
37 #include "opto/rootnode.hpp"
38 #include "opto/subnode.hpp"
39 #include "opto/subtypenode.hpp"
40
41 // Portions of code courtesy of Clifford Click
42
43 // Optimization - Graph Style
44
45
46 #ifndef PRODUCT
47 extern uint explicit_null_checks_elided;
48 #endif
49
50 IfNode::IfNode(Node* control, Node* bol, float p, float fcnt)
51 : MultiBranchNode(2),
52 _prob(p),
53 _fcnt(fcnt)
54 NOT_PRODUCT(COMMA _assertion_predicate_type(AssertionPredicateType::None)) {
55 init_node(control, bol);
56 }
57
58 #ifndef PRODUCT
59 IfNode::IfNode(Node* control, Node* bol, float p, float fcnt, AssertionPredicateType assertion_predicate_type)
60 : MultiBranchNode(2),
61 _prob(p),
62 _fcnt(fcnt),
63 _assertion_predicate_type(assertion_predicate_type) {
64 init_node(control, bol);
65 }
66 #endif // NOT_PRODUCT
67
68 //=============================================================================
69 //------------------------------Value------------------------------------------
70 // Return a tuple for whichever arm of the IF is reachable
71 const Type* IfNode::Value(PhaseGVN* phase) const {
72 if( !in(0) ) return Type::TOP;
73 if( phase->type(in(0)) == Type::TOP )
74 return Type::TOP;
75 const Type *t = phase->type(in(1));
76 if( t == Type::TOP ) // data is undefined
77 return TypeTuple::IFNEITHER; // unreachable altogether
78 if( t == TypeInt::ZERO ) // zero, or false
79 return TypeTuple::IFFALSE; // only false branch is reachable
80 if( t == TypeInt::ONE ) // 1, or true
81 return TypeTuple::IFTRUE; // only true branch is reachable
82 assert( t == TypeInt::BOOL, "expected boolean type" );
83
84 return TypeTuple::IFBOTH; // No progress
85 }
86
87 const RegMask &IfNode::out_RegMask() const {
88 return RegMask::Empty;
89 }
90
91 //------------------------------split_if---------------------------------------
92 // Look for places where we merge constants, then test on the merged value.
93 // If the IF test will be constant folded on the path with the constant, we
94 // win by splitting the IF to before the merge point.
95 static Node* split_if(IfNode *iff, PhaseIterGVN *igvn) {
96 // I could be a lot more general here, but I'm trying to squeeze this
97 // in before the Christmas '98 break so I'm gonna be kinda restrictive
98 // on the patterns I accept. CNC
99
100 // Look for a compare of a constant and a merged value
101 Node *i1 = iff->in(1);
102 if( !i1->is_Bool() ) return nullptr;
103 BoolNode *b = i1->as_Bool();
104 Node *cmp = b->in(1);
105 if( !cmp->is_Cmp() ) return nullptr;
106 i1 = cmp->in(1);
107 if( i1 == nullptr || !i1->is_Phi() ) return nullptr;
108 PhiNode *phi = i1->as_Phi();
109 Node *con2 = cmp->in(2);
110 if( !con2->is_Con() ) return nullptr;
111 // See that the merge point contains some constants
112 Node *con1=nullptr;
113 uint i4;
114 RegionNode* phi_region = phi->region();
115 for (i4 = 1; i4 < phi->req(); i4++ ) {
116 con1 = phi->in(i4);
117 // Do not optimize partially collapsed merges
118 if (con1 == nullptr || phi_region->in(i4) == nullptr || igvn->type(phi_region->in(i4)) == Type::TOP) {
119 igvn->_worklist.push(iff);
120 return nullptr;
121 }
122 if( con1->is_Con() ) break; // Found a constant
123 // Also allow null-vs-not-null checks
124 const TypePtr *tp = igvn->type(con1)->isa_ptr();
125 if( tp && tp->_ptr == TypePtr::NotNull )
126 break;
127 }
128 if( i4 >= phi->req() ) return nullptr; // Found no constants
129
130 igvn->C->set_has_split_ifs(true); // Has chance for split-if
131
132 // Make sure that the compare can be constant folded away
133 Node *cmp2 = cmp->clone();
134 cmp2->set_req(1,con1);
135 cmp2->set_req(2,con2);
136 const Type *t = cmp2->Value(igvn);
137 // This compare is dead, so whack it!
138 igvn->remove_dead_node(cmp2);
139 if( !t->singleton() ) return nullptr;
140
141 // No intervening control, like a simple Call
142 Node* r = iff->in(0);
143 if (!r->is_Region() || r->is_Loop() || phi_region != r || r->as_Region()->is_copy()) {
144 return nullptr;
145 }
146
147 // No other users of the cmp/bool
148 if (b->outcnt() != 1 || cmp->outcnt() != 1) {
149 //tty->print_cr("many users of cmp/bool");
150 return nullptr;
151 }
152
153 // Make sure we can determine where all the uses of merged values go
154 for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) {
155 Node* u = r->fast_out(j);
156 if( u == r ) continue;
157 if( u == iff ) continue;
158 if( u->outcnt() == 0 ) continue; // use is dead & ignorable
159 if( !u->is_Phi() ) {
160 /*
161 if( u->is_Start() ) {
162 tty->print_cr("Region has inlined start use");
163 } else {
164 tty->print_cr("Region has odd use");
165 u->dump(2);
166 }*/
167 return nullptr;
168 }
169 if( u != phi ) {
170 // CNC - do not allow any other merged value
171 //tty->print_cr("Merging another value");
172 //u->dump(2);
173 return nullptr;
174 }
175 // Make sure we can account for all Phi uses
176 for (DUIterator_Fast kmax, k = u->fast_outs(kmax); k < kmax; k++) {
177 Node* v = u->fast_out(k); // User of the phi
178 // CNC - Allow only really simple patterns.
179 // In particular I disallow AddP of the Phi, a fairly common pattern
180 if (v == cmp) continue; // The compare is OK
181 if (v->is_ConstraintCast()) {
182 // If the cast is derived from data flow edges, it may not have a control edge.
183 // If so, it should be safe to split. But follow-up code can not deal with
184 // this (l. 359). So skip.
185 if (v->in(0) == nullptr) {
186 return nullptr;
187 }
188 if (v->in(0)->in(0) == iff) {
189 continue; // CastPP/II of the IfNode is OK
190 }
191 }
192 // Disabled following code because I cannot tell if exactly one
193 // path dominates without a real dominator check. CNC 9/9/1999
194 //uint vop = v->Opcode();
195 //if( vop == Op_Phi ) { // Phi from another merge point might be OK
196 // Node *r = v->in(0); // Get controlling point
197 // if( !r ) return nullptr; // Degraded to a copy
198 // // Find exactly one path in (either True or False doms, but not IFF)
199 // int cnt = 0;
200 // for( uint i = 1; i < r->req(); i++ )
201 // if( r->in(i) && r->in(i)->in(0) == iff )
202 // cnt++;
203 // if( cnt == 1 ) continue; // Exactly one of True or False guards Phi
204 //}
205 if( !v->is_Call() ) {
206 /*
207 if( v->Opcode() == Op_AddP ) {
208 tty->print_cr("Phi has AddP use");
209 } else if( v->Opcode() == Op_CastPP ) {
210 tty->print_cr("Phi has CastPP use");
211 } else if( v->Opcode() == Op_CastII ) {
212 tty->print_cr("Phi has CastII use");
213 } else {
214 tty->print_cr("Phi has use I can't be bothered with");
215 }
216 */
217 }
218 return nullptr;
219
220 /* CNC - Cut out all the fancy acceptance tests
221 // Can we clone this use when doing the transformation?
222 // If all uses are from Phis at this merge or constants, then YES.
223 if( !v->in(0) && v != cmp ) {
224 tty->print_cr("Phi has free-floating use");
225 v->dump(2);
226 return nullptr;
227 }
228 for( uint l = 1; l < v->req(); l++ ) {
229 if( (!v->in(l)->is_Phi() || v->in(l)->in(0) != r) &&
230 !v->in(l)->is_Con() ) {
231 tty->print_cr("Phi has use");
232 v->dump(2);
233 return nullptr;
234 } // End of if Phi-use input is neither Phi nor Constant
235 } // End of for all inputs to Phi-use
236 */
237 } // End of for all uses of Phi
238 } // End of for all uses of Region
239
240 // Only do this if the IF node is in a sane state
241 if (iff->outcnt() != 2)
242 return nullptr;
243
244 // Got a hit! Do the Mondo Hack!
245 //
246 //ABC a1c def ghi B 1 e h A C a c d f g i
247 // R - Phi - Phi - Phi Rc - Phi - Phi - Phi Rx - Phi - Phi - Phi
248 // cmp - 2 cmp - 2 cmp - 2
249 // bool bool_c bool_x
250 // if if_c if_x
251 // T F T F T F
252 // ..s.. ..t .. ..s.. ..t.. ..s.. ..t..
253 //
254 // Split the paths coming into the merge point into 2 separate groups of
255 // merges. On the left will be all the paths feeding constants into the
256 // Cmp's Phi. On the right will be the remaining paths. The Cmp's Phi
257 // will fold up into a constant; this will let the Cmp fold up as well as
258 // all the control flow. Below the original IF we have 2 control
259 // dependent regions, 's' and 't'. Now we will merge the two paths
260 // just prior to 's' and 't' from the two IFs. At least 1 path (and quite
261 // likely 2 or more) will promptly constant fold away.
262 PhaseGVN *phase = igvn;
263
264 // Make a region merging constants and a region merging the rest
265 uint req_c = 0;
266 for (uint ii = 1; ii < r->req(); ii++) {
267 if (phi->in(ii) == con1) {
268 req_c++;
269 }
270 if (Node::may_be_loop_entry(r->in(ii))) {
271 // Bail out if splitting through a region with a Parse Predicate input (could
272 // also be a loop header before loop opts creates a LoopNode for it).
273 return nullptr;
274 }
275 }
276
277 // If all the defs of the phi are the same constant, we already have the desired end state.
278 // Skip the split that would create empty phi and region nodes.
279 if ((r->req() - req_c) == 1) {
280 return nullptr;
281 }
282
283 // At this point we know that we can apply the split if optimization. If the region is still on the worklist,
284 // we should wait until it is processed. The region might be removed which makes this optimization redundant.
285 // This also avoids the creation of dead data loops when rewiring data nodes below when a region is dying.
286 if (igvn->_worklist.member(r)) {
287 igvn->_worklist.push(iff); // retry split if later again
288 return nullptr;
289 }
290
291 Node *region_c = new RegionNode(req_c + 1);
292 Node *phi_c = con1;
293 uint len = r->req();
294 Node *region_x = new RegionNode(len - req_c);
295 Node *phi_x = PhiNode::make_blank(region_x, phi);
296 for (uint i = 1, i_c = 1, i_x = 1; i < len; i++) {
297 if (phi->in(i) == con1) {
298 region_c->init_req( i_c++, r ->in(i) );
299 } else {
300 region_x->init_req( i_x, r ->in(i) );
301 phi_x ->init_req( i_x++, phi->in(i) );
302 }
303 }
304
305 // Register the new RegionNodes but do not transform them. Cannot
306 // transform until the entire Region/Phi conglomerate has been hacked
307 // as a single huge transform.
308 igvn->register_new_node_with_optimizer( region_c );
309 igvn->register_new_node_with_optimizer( region_x );
310 // Prevent the untimely death of phi_x. Currently he has no uses. He is
311 // about to get one. If this only use goes away, then phi_x will look dead.
312 // However, he will be picking up some more uses down below.
313 Node *hook = new Node(4);
314 hook->init_req(0, phi_x);
315 hook->init_req(1, phi_c);
316 phi_x = phase->transform( phi_x );
317
318 // Make the compare
319 Node *cmp_c = phase->makecon(t);
320 Node *cmp_x = cmp->clone();
321 cmp_x->set_req(1,phi_x);
322 cmp_x->set_req(2,con2);
323 cmp_x = phase->transform(cmp_x);
324 // Make the bool
325 Node *b_c = phase->transform(new BoolNode(cmp_c,b->_test._test));
326 Node *b_x = phase->transform(new BoolNode(cmp_x,b->_test._test));
327 // Make the IfNode
328 IfNode* iff_c = iff->clone()->as_If();
329 iff_c->set_req(0, region_c);
330 iff_c->set_req(1, b_c);
331 igvn->set_type_bottom(iff_c);
332 igvn->_worklist.push(iff_c);
333 hook->init_req(2, iff_c);
334
335 IfNode* iff_x = iff->clone()->as_If();
336 iff_x->set_req(0, region_x);
337 iff_x->set_req(1, b_x);
338 igvn->set_type_bottom(iff_x);
339 igvn->_worklist.push(iff_x);
340 hook->init_req(3, iff_x);
341
342 // Make the true/false arms
343 Node *iff_c_t = phase->transform(new IfTrueNode (iff_c));
344 Node *iff_c_f = phase->transform(new IfFalseNode(iff_c));
345 Node *iff_x_t = phase->transform(new IfTrueNode (iff_x));
346 Node *iff_x_f = phase->transform(new IfFalseNode(iff_x));
347
348 // Merge the TRUE paths
349 Node *region_s = new RegionNode(3);
350 igvn->_worklist.push(region_s);
351 region_s->init_req(1, iff_c_t);
352 region_s->init_req(2, iff_x_t);
353 igvn->register_new_node_with_optimizer( region_s );
354
355 // Merge the FALSE paths
356 Node *region_f = new RegionNode(3);
357 igvn->_worklist.push(region_f);
358 region_f->init_req(1, iff_c_f);
359 region_f->init_req(2, iff_x_f);
360 igvn->register_new_node_with_optimizer( region_f );
361
362 igvn->hash_delete(cmp);// Remove soon-to-be-dead node from hash table.
363 cmp->set_req(1,nullptr); // Whack the inputs to cmp because it will be dead
364 cmp->set_req(2,nullptr);
365 // Check for all uses of the Phi and give them a new home.
366 // The 'cmp' got cloned, but CastPP/IIs need to be moved.
367 Node *phi_s = nullptr; // do not construct unless needed
368 Node *phi_f = nullptr; // do not construct unless needed
369 for (DUIterator_Last i2min, i2 = phi->last_outs(i2min); i2 >= i2min; --i2) {
370 Node* v = phi->last_out(i2);// User of the phi
371 igvn->rehash_node_delayed(v); // Have to fixup other Phi users
372 uint vop = v->Opcode();
373 Node *proj = nullptr;
374 if( vop == Op_Phi ) { // Remote merge point
375 Node *r = v->in(0);
376 for (uint i3 = 1; i3 < r->req(); i3++)
377 if (r->in(i3) && r->in(i3)->in(0) == iff) {
378 proj = r->in(i3);
379 break;
380 }
381 } else if( v->is_ConstraintCast() ) {
382 proj = v->in(0); // Controlling projection
383 } else {
384 assert( 0, "do not know how to handle this guy" );
385 }
386 guarantee(proj != nullptr, "sanity");
387
388 Node *proj_path_data, *proj_path_ctrl;
389 if( proj->Opcode() == Op_IfTrue ) {
390 if( phi_s == nullptr ) {
391 // Only construct phi_s if needed, otherwise provides
392 // interfering use.
393 phi_s = PhiNode::make_blank(region_s,phi);
394 phi_s->init_req( 1, phi_c );
395 phi_s->init_req( 2, phi_x );
396 hook->add_req(phi_s);
397 phi_s = phase->transform(phi_s);
398 }
399 proj_path_data = phi_s;
400 proj_path_ctrl = region_s;
401 } else {
402 if( phi_f == nullptr ) {
403 // Only construct phi_f if needed, otherwise provides
404 // interfering use.
405 phi_f = PhiNode::make_blank(region_f,phi);
406 phi_f->init_req( 1, phi_c );
407 phi_f->init_req( 2, phi_x );
408 hook->add_req(phi_f);
409 phi_f = phase->transform(phi_f);
410 }
411 proj_path_data = phi_f;
412 proj_path_ctrl = region_f;
413 }
414
415 // Fixup 'v' for for the split
416 if( vop == Op_Phi ) { // Remote merge point
417 uint i;
418 for( i = 1; i < v->req(); i++ )
419 if( v->in(i) == phi )
420 break;
421 v->set_req(i, proj_path_data );
422 } else if( v->is_ConstraintCast() ) {
423 v->set_req(0, proj_path_ctrl );
424 v->set_req(1, proj_path_data );
425 } else
426 ShouldNotReachHere();
427 }
428
429 // Now replace the original iff's True/False with region_s/region_t.
430 // This makes the original iff go dead.
431 for (DUIterator_Last i3min, i3 = iff->last_outs(i3min); i3 >= i3min; --i3) {
432 Node* p = iff->last_out(i3);
433 assert( p->Opcode() == Op_IfTrue || p->Opcode() == Op_IfFalse, "" );
434 Node *u = (p->Opcode() == Op_IfTrue) ? region_s : region_f;
435 // Replace p with u
436 igvn->add_users_to_worklist(p);
437 for (DUIterator_Last lmin, l = p->last_outs(lmin); l >= lmin;) {
438 Node* x = p->last_out(l);
439 igvn->hash_delete(x);
440 uint uses_found = 0;
441 for( uint j = 0; j < x->req(); j++ ) {
442 if( x->in(j) == p ) {
443 x->set_req(j, u);
444 uses_found++;
445 }
446 }
447 l -= uses_found; // we deleted 1 or more copies of this edge
448 }
449 igvn->remove_dead_node(p);
450 }
451
452 // Force the original merge dead
453 igvn->hash_delete(r);
454 // First, remove region's dead users.
455 for (DUIterator_Last lmin, l = r->last_outs(lmin); l >= lmin;) {
456 Node* u = r->last_out(l);
457 if( u == r ) {
458 r->set_req(0, nullptr);
459 } else {
460 assert(u->outcnt() == 0, "only dead users");
461 igvn->remove_dead_node(u);
462 }
463 l -= 1;
464 }
465 igvn->remove_dead_node(r);
466
467 // Now remove the bogus extra edges used to keep things alive
468 igvn->remove_dead_node( hook );
469
470 // Must return either the original node (now dead) or a new node
471 // (Do not return a top here, since that would break the uniqueness of top.)
472 return new ConINode(TypeInt::ZERO);
473 }
474
475 IfNode* IfNode::make_with_same_profile(IfNode* if_node_profile, Node* ctrl, BoolNode* bol) {
476 // Assert here that we only try to create a clone from an If node with the same profiling if that actually makes sense.
477 // Some If node subtypes should not be cloned in this way. In theory, we should not clone BaseCountedLoopEndNodes.
478 // But they can end up being used as normal If nodes when peeling a loop - they serve as zero-trip guard.
479 // Allow them as well.
480 assert(if_node_profile->Opcode() == Op_If || if_node_profile->is_RangeCheck()
481 || if_node_profile->is_BaseCountedLoopEnd(), "should not clone other nodes");
482 if (if_node_profile->is_RangeCheck()) {
483 // RangeCheck nodes could be further optimized.
484 return new RangeCheckNode(ctrl, bol, if_node_profile->_prob, if_node_profile->_fcnt);
485 } else {
486 // Not a RangeCheckNode? Fall back to IfNode.
487 return new IfNode(ctrl, bol, if_node_profile->_prob, if_node_profile->_fcnt);
488 }
489 }
490
491 // if this IfNode follows a range check pattern return the projection
492 // for the failed path
493 ProjNode* IfNode::range_check_trap_proj(int& flip_test, Node*& l, Node*& r) {
494 if (outcnt() != 2) {
495 return nullptr;
496 }
497 Node* b = in(1);
498 if (b == nullptr || !b->is_Bool()) return nullptr;
499 BoolNode* bn = b->as_Bool();
500 Node* cmp = bn->in(1);
501 if (cmp == nullptr) return nullptr;
502 if (cmp->Opcode() != Op_CmpU) return nullptr;
503
504 l = cmp->in(1);
505 r = cmp->in(2);
506 flip_test = 1;
507 if (bn->_test._test == BoolTest::le) {
508 l = cmp->in(2);
509 r = cmp->in(1);
510 flip_test = 2;
511 } else if (bn->_test._test != BoolTest::lt) {
512 return nullptr;
513 }
514 if (l->is_top()) return nullptr; // Top input means dead test
515 if (r->Opcode() != Op_LoadRange && !is_RangeCheck()) return nullptr;
516
517 // We have recognized one of these forms:
518 // Flip 1: If (Bool[<] CmpU(l, LoadRange)) ...
519 // Flip 2: If (Bool[<=] CmpU(LoadRange, l)) ...
520
521 ProjNode* iftrap = proj_out_or_null(flip_test == 2 ? true : false);
522 return iftrap;
523 }
524
525
526 //------------------------------is_range_check---------------------------------
527 // Return 0 if not a range check. Return 1 if a range check and set index and
528 // offset. Return 2 if we had to negate the test. Index is null if the check
529 // is versus a constant.
530 int RangeCheckNode::is_range_check(Node* &range, Node* &index, jint &offset) {
531 int flip_test = 0;
532 Node* l = nullptr;
533 Node* r = nullptr;
534 ProjNode* iftrap = range_check_trap_proj(flip_test, l, r);
535
536 if (iftrap == nullptr) {
537 return 0;
538 }
539
540 // Make sure it's a real range check by requiring an uncommon trap
541 // along the OOB path. Otherwise, it's possible that the user wrote
542 // something which optimized to look like a range check but behaves
543 // in some other way.
544 if (iftrap->is_uncommon_trap_proj(Deoptimization::Reason_range_check) == nullptr) {
545 return 0;
546 }
547
548 // Look for index+offset form
549 Node* ind = l;
550 jint off = 0;
551 if (l->is_top()) {
552 return 0;
553 } else if (l->Opcode() == Op_AddI) {
554 if ((off = l->in(1)->find_int_con(0)) != 0) {
555 ind = l->in(2)->uncast();
556 } else if ((off = l->in(2)->find_int_con(0)) != 0) {
557 ind = l->in(1)->uncast();
558 }
559 } else if ((off = l->find_int_con(-1)) >= 0) {
560 // constant offset with no variable index
561 ind = nullptr;
562 } else {
563 // variable index with no constant offset (or dead negative index)
564 off = 0;
565 }
566
567 // Return all the values:
568 index = ind;
569 offset = off;
570 range = r;
571 return flip_test;
572 }
573
574 //------------------------------adjust_check-----------------------------------
575 // Adjust (widen) a prior range check
576 static void adjust_check(IfProjNode* proj, Node* range, Node* index,
577 int flip, jint off_lo, PhaseIterGVN* igvn) {
578 PhaseGVN *gvn = igvn;
579 // Break apart the old check
580 Node *iff = proj->in(0);
581 Node *bol = iff->in(1);
582 if( bol->is_top() ) return; // In case a partially dead range check appears
583 // bail (or bomb[ASSERT/DEBUG]) if NOT projection-->IfNode-->BoolNode
584 DEBUG_ONLY( if (!bol->is_Bool()) { proj->dump(3); fatal("Expect projection-->IfNode-->BoolNode"); } )
585 if (!bol->is_Bool()) return;
586
587 Node *cmp = bol->in(1);
588 // Compute a new check
589 Node *new_add = gvn->intcon(off_lo);
590 if (index) {
591 new_add = off_lo ? gvn->transform(new AddINode(index, new_add)) : index;
592 }
593 Node *new_cmp = (flip == 1)
594 ? new CmpUNode(new_add, range)
595 : new CmpUNode(range, new_add);
596 new_cmp = gvn->transform(new_cmp);
597 // See if no need to adjust the existing check
598 if (new_cmp == cmp) return;
599 // Else, adjust existing check
600 Node* new_bol = gvn->transform(new BoolNode(new_cmp, bol->as_Bool()->_test._test));
601 igvn->rehash_node_delayed(iff);
602 iff->set_req_X(1, new_bol, igvn);
603 // As part of range check smearing, this range check is widened. Loads and range check Cast nodes that are control
604 // dependent on this range check now depend on multiple dominating range checks. These control dependent nodes end up
605 // at the lowest/nearest dominating check in the graph. To ensure that these Loads/Casts do not float above any of the
606 // dominating checks (even when the lowest dominating check is later replaced by yet another dominating check), we
607 // need to pin them at the lowest dominating check.
608 proj->pin_array_access_nodes(igvn);
609 }
610
611 //------------------------------up_one_dom-------------------------------------
612 // Walk up the dominator tree one step. Return null at root or true
613 // complex merges. Skips through small diamonds.
614 Node* IfNode::up_one_dom(Node *curr, bool linear_only) {
615 Node *dom = curr->in(0);
616 if( !dom ) // Found a Region degraded to a copy?
617 return curr->nonnull_req(); // Skip thru it
618
619 if( curr != dom ) // Normal walk up one step?
620 return dom;
621
622 // Use linear_only if we are still parsing, since we cannot
623 // trust the regions to be fully filled in.
624 if (linear_only)
625 return nullptr;
626
627 if( dom->is_Root() )
628 return nullptr;
629
630 // Else hit a Region. Check for a loop header
631 if( dom->is_Loop() )
632 return dom->in(1); // Skip up thru loops
633
634 // Check for small diamonds
635 Node *din1, *din2, *din3, *din4;
636 if( dom->req() == 3 && // 2-path merge point
637 (din1 = dom ->in(1)) && // Left path exists
638 (din2 = dom ->in(2)) && // Right path exists
639 (din3 = din1->in(0)) && // Left path up one
640 (din4 = din2->in(0)) ) { // Right path up one
641 if( din3->is_Call() && // Handle a slow-path call on either arm
642 (din3 = din3->in(0)) )
643 din3 = din3->in(0);
644 if( din4->is_Call() && // Handle a slow-path call on either arm
645 (din4 = din4->in(0)) )
646 din4 = din4->in(0);
647 if (din3 != nullptr && din3 == din4 && din3->is_If()) // Regions not degraded to a copy
648 return din3; // Skip around diamonds
649 }
650
651 // Give up the search at true merges
652 return nullptr; // Dead loop? Or hit root?
653 }
654
655
656 //------------------------------filtered_int_type--------------------------------
657 // Return a possibly more restrictive type for val based on condition control flow for an if
658 const TypeInt* IfNode::filtered_int_type(PhaseGVN* gvn, Node* val, Node* if_proj) {
659 assert(if_proj &&
660 (if_proj->Opcode() == Op_IfTrue || if_proj->Opcode() == Op_IfFalse), "expecting an if projection");
661 if (if_proj->in(0) && if_proj->in(0)->is_If()) {
662 IfNode* iff = if_proj->in(0)->as_If();
663 if (iff->in(1) && iff->in(1)->is_Bool()) {
664 BoolNode* bol = iff->in(1)->as_Bool();
665 if (bol->in(1) && bol->in(1)->is_Cmp()) {
666 const CmpNode* cmp = bol->in(1)->as_Cmp();
667 if (cmp->in(1) == val) {
668 const TypeInt* cmp2_t = gvn->type(cmp->in(2))->isa_int();
669 if (cmp2_t != nullptr) {
670 jint lo = cmp2_t->_lo;
671 jint hi = cmp2_t->_hi;
672 BoolTest::mask msk = if_proj->Opcode() == Op_IfTrue ? bol->_test._test : bol->_test.negate();
673 switch (msk) {
674 case BoolTest::ne: {
675 // If val is compared to its lower or upper bound, we can narrow the type
676 const TypeInt* val_t = gvn->type(val)->isa_int();
677 if (val_t != nullptr && !val_t->singleton() && cmp2_t->is_con()) {
678 if (val_t->_lo == lo) {
679 return TypeInt::make(val_t->_lo + 1, val_t->_hi, val_t->_widen);
680 } else if (val_t->_hi == hi) {
681 return TypeInt::make(val_t->_lo, val_t->_hi - 1, val_t->_widen);
682 }
683 }
684 // Can't refine type
685 return nullptr;
686 }
687 case BoolTest::eq:
688 return cmp2_t;
689 case BoolTest::lt:
690 lo = TypeInt::INT->_lo;
691 if (hi != min_jint) {
692 hi = hi - 1;
693 }
694 break;
695 case BoolTest::le:
696 lo = TypeInt::INT->_lo;
697 break;
698 case BoolTest::gt:
699 if (lo != max_jint) {
700 lo = lo + 1;
701 }
702 hi = TypeInt::INT->_hi;
703 break;
704 case BoolTest::ge:
705 // lo unchanged
706 hi = TypeInt::INT->_hi;
707 break;
708 default:
709 break;
710 }
711 const TypeInt* rtn_t = TypeInt::make(lo, hi, cmp2_t->_widen);
712 return rtn_t;
713 }
714 }
715 }
716 }
717 }
718 return nullptr;
719 }
720
721 //------------------------------fold_compares----------------------------
722 // See if a pair of CmpIs can be converted into a CmpU. In some cases
723 // the direction of this if is determined by the preceding if so it
724 // can be eliminate entirely.
725 //
726 // Given an if testing (CmpI n v) check for an immediately control
727 // dependent if that is testing (CmpI n v2) and has one projection
728 // leading to this if and the other projection leading to a region
729 // that merges one of this ifs control projections.
730 //
731 // If
732 // / |
733 // / |
734 // / |
735 // If |
736 // /\ |
737 // / \ |
738 // / \ |
739 // / Region
740 //
741 // Or given an if testing (CmpI n v) check for a dominating if that is
742 // testing (CmpI n v2), both having one projection leading to an
743 // uncommon trap. Allow Another independent guard in between to cover
744 // an explicit range check:
745 // if (index < 0 || index >= array.length) {
746 // which may need a null check to guard the LoadRange
747 //
748 // If
749 // / \
750 // / \
751 // / \
752 // If unc
753 // /\
754 // / \
755 // / \
756 // / unc
757 //
758
759 // Is the comparison for this If suitable for folding?
760 bool IfNode::cmpi_folds(PhaseIterGVN* igvn, bool fold_ne) {
761 return in(1) != nullptr &&
762 in(1)->is_Bool() &&
763 in(1)->in(1) != nullptr &&
764 in(1)->in(1)->Opcode() == Op_CmpI &&
765 in(1)->in(1)->in(2) != nullptr &&
766 in(1)->in(1)->in(2) != igvn->C->top() &&
767 (in(1)->as_Bool()->_test.is_less() ||
768 in(1)->as_Bool()->_test.is_greater() ||
769 (fold_ne && in(1)->as_Bool()->_test._test == BoolTest::ne));
770 }
771
772 // Is a dominating control suitable for folding with this if?
773 bool IfNode::is_ctrl_folds(Node* ctrl, PhaseIterGVN* igvn) {
774 return ctrl != nullptr &&
775 ctrl->is_Proj() &&
776 ctrl->outcnt() == 1 && // No side-effects
777 ctrl->in(0) != nullptr &&
778 ctrl->in(0)->Opcode() == Op_If &&
779 ctrl->in(0)->outcnt() == 2 &&
780 ctrl->in(0)->as_If()->cmpi_folds(igvn, true) &&
781 // Must compare same value
782 ctrl->in(0)->in(1)->in(1)->in(1) != nullptr &&
783 ctrl->in(0)->in(1)->in(1)->in(1) != igvn->C->top() &&
784 ctrl->in(0)->in(1)->in(1)->in(1) == in(1)->in(1)->in(1);
785 }
786
787 // Do this If and the dominating If share a region?
788 bool IfNode::has_shared_region(ProjNode* proj, ProjNode*& success, ProjNode*& fail) {
789 ProjNode* otherproj = proj->other_if_proj();
790 Node* otherproj_ctrl_use = otherproj->unique_ctrl_out_or_null();
791 RegionNode* region = (otherproj_ctrl_use != nullptr && otherproj_ctrl_use->is_Region()) ? otherproj_ctrl_use->as_Region() : nullptr;
792 success = nullptr;
793 fail = nullptr;
794
795 if (otherproj->outcnt() == 1 && region != nullptr && !region->has_phi()) {
796 for (int i = 0; i < 2; i++) {
797 ProjNode* proj = proj_out(i);
798 if (success == nullptr && proj->outcnt() == 1 && proj->unique_out() == region) {
799 success = proj;
800 } else if (fail == nullptr) {
801 fail = proj;
802 } else {
803 success = fail = nullptr;
804 }
805 }
806 }
807 return success != nullptr && fail != nullptr;
808 }
809
810 bool IfNode::is_dominator_unc(CallStaticJavaNode* dom_unc, CallStaticJavaNode* unc) {
811 // Different methods and methods containing jsrs are not supported.
812 ciMethod* method = unc->jvms()->method();
813 ciMethod* dom_method = dom_unc->jvms()->method();
814 if (method != dom_method || method->has_jsrs()) {
815 return false;
816 }
817 // Check that both traps are in the same activation of the method (instead
818 // of two activations being inlined through different call sites) by verifying
819 // that the call stacks are equal for both JVMStates.
820 JVMState* dom_caller = dom_unc->jvms()->caller();
821 JVMState* caller = unc->jvms()->caller();
822 if ((dom_caller == nullptr) != (caller == nullptr)) {
823 // The current method must either be inlined into both dom_caller and
824 // caller or must not be inlined at all (top method). Bail out otherwise.
825 return false;
826 } else if (dom_caller != nullptr && !dom_caller->same_calls_as(caller)) {
827 return false;
828 }
829 // Check that the bci of the dominating uncommon trap dominates the bci
830 // of the dominated uncommon trap. Otherwise we may not re-execute
831 // the dominated check after deoptimization from the merged uncommon trap.
832 ciTypeFlow* flow = dom_method->get_flow_analysis();
833 int bci = unc->jvms()->bci();
834 int dom_bci = dom_unc->jvms()->bci();
835 if (!flow->is_dominated_by(bci, dom_bci)) {
836 return false;
837 }
838
839 return true;
840 }
841
842 // Return projection that leads to an uncommon trap if any
843 ProjNode* IfNode::uncommon_trap_proj(CallStaticJavaNode*& call, Deoptimization::DeoptReason reason) const {
844 for (int i = 0; i < 2; i++) {
845 call = proj_out(i)->is_uncommon_trap_proj(reason);
846 if (call != nullptr) {
847 return proj_out(i);
848 }
849 }
850 return nullptr;
851 }
852
853 // Do this If and the dominating If both branch out to an uncommon trap
854 bool IfNode::has_only_uncommon_traps(ProjNode* proj, ProjNode*& success, ProjNode*& fail, PhaseIterGVN* igvn) {
855 ProjNode* otherproj = proj->other_if_proj();
856 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj();
857
858 if (otherproj->outcnt() == 1 && dom_unc != nullptr) {
859 // We need to re-execute the folded Ifs after deoptimization from the merged traps
860 if (!dom_unc->jvms()->should_reexecute()) {
861 return false;
862 }
863
864 CallStaticJavaNode* unc = nullptr;
865 ProjNode* unc_proj = uncommon_trap_proj(unc);
866 if (unc_proj != nullptr && unc_proj->outcnt() == 1) {
867 if (dom_unc == unc) {
868 // Allow the uncommon trap to be shared through a region
869 RegionNode* r = unc->in(0)->as_Region();
870 if (r->outcnt() != 2 || r->req() != 3 || r->find_edge(otherproj) == -1 || r->find_edge(unc_proj) == -1) {
871 return false;
872 }
873 assert(r->has_phi() == nullptr, "simple region shouldn't have a phi");
874 } else if (dom_unc->in(0) != otherproj || unc->in(0) != unc_proj) {
875 return false;
876 }
877
878 if (!is_dominator_unc(dom_unc, unc)) {
879 return false;
880 }
881
882 // See merge_uncommon_traps: the reason of the uncommon trap
883 // will be changed and the state of the dominating If will be
884 // used. Checked that we didn't apply this transformation in a
885 // previous compilation and it didn't cause too many traps
886 ciMethod* dom_method = dom_unc->jvms()->method();
887 int dom_bci = dom_unc->jvms()->bci();
888 if (!igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_unstable_fused_if) &&
889 !igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_range_check) &&
890 // Return true if c2 manages to reconcile with UnstableIf optimization. See the comments for it.
891 igvn->C->remove_unstable_if_trap(dom_unc, true/*yield*/)) {
892 success = unc_proj;
893 fail = unc_proj->other_if_proj();
894 return true;
895 }
896 }
897 }
898 return false;
899 }
900
901 // Check that the 2 CmpI can be folded into as single CmpU and proceed with the folding
902 bool IfNode::fold_compares_helper(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) {
903 Node* this_cmp = in(1)->in(1);
904 BoolNode* this_bool = in(1)->as_Bool();
905 IfNode* dom_iff = proj->in(0)->as_If();
906 BoolNode* dom_bool = dom_iff->in(1)->as_Bool();
907 Node* lo = dom_iff->in(1)->in(1)->in(2);
908 Node* hi = this_cmp->in(2);
909 Node* n = this_cmp->in(1);
910 ProjNode* otherproj = proj->other_if_proj();
911
912 const TypeInt* lo_type = IfNode::filtered_int_type(igvn, n, otherproj);
913 const TypeInt* hi_type = IfNode::filtered_int_type(igvn, n, success);
914
915 BoolTest::mask lo_test = dom_bool->_test._test;
916 BoolTest::mask hi_test = this_bool->_test._test;
917 BoolTest::mask cond = hi_test;
918
919 // convert:
920 //
921 // dom_bool = x {<,<=,>,>=} a
922 // / \
923 // proj = {True,False} / \ otherproj = {False,True}
924 // /
925 // this_bool = x {<,<=} b
926 // / \
927 // fail = {True,False} / \ success = {False,True}
928 // /
929 //
930 // (Second test guaranteed canonicalized, first one may not have
931 // been canonicalized yet)
932 //
933 // into:
934 //
935 // cond = (x - lo) {<u,<=u,>u,>=u} adjusted_lim
936 // / \
937 // fail / \ success
938 // /
939 //
940
941 // Figure out which of the two tests sets the upper bound and which
942 // sets the lower bound if any.
943 Node* adjusted_lim = nullptr;
944 if (lo_type != nullptr && hi_type != nullptr && hi_type->_lo > lo_type->_hi &&
945 hi_type->_hi == max_jint && lo_type->_lo == min_jint && lo_test != BoolTest::ne) {
946 assert((dom_bool->_test.is_less() && !proj->_con) ||
947 (dom_bool->_test.is_greater() && proj->_con), "incorrect test");
948
949 // this_bool = <
950 // dom_bool = >= (proj = True) or dom_bool = < (proj = False)
951 // x in [a, b[ on the fail (= True) projection, b > a-1 (because of hi_type->_lo > lo_type->_hi test above):
952 // lo = a, hi = b, adjusted_lim = b-a, cond = <u
953 // dom_bool = > (proj = True) or dom_bool = <= (proj = False)
954 // x in ]a, b[ on the fail (= True) projection, b > a:
955 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <u
956 // this_bool = <=
957 // dom_bool = >= (proj = True) or dom_bool = < (proj = False)
958 // x in [a, b] on the fail (= True) projection, b+1 > a-1:
959 // lo = a, hi = b, adjusted_lim = b-a+1, cond = <u
960 // lo = a, hi = b, adjusted_lim = b-a, cond = <=u doesn't work because b = a - 1 is possible, then b-a = -1
961 // dom_bool = > (proj = True) or dom_bool = <= (proj = False)
962 // x in ]a, b] on the fail (= True) projection b+1 > a:
963 // lo = a+1, hi = b, adjusted_lim = b-a, cond = <u
964 // 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
965
966 if (hi_test == BoolTest::lt) {
967 if (lo_test == BoolTest::gt || lo_test == BoolTest::le) {
968 lo = igvn->transform(new AddINode(lo, igvn->intcon(1)));
969 }
970 } else if (hi_test == BoolTest::le) {
971 if (lo_test == BoolTest::ge || lo_test == BoolTest::lt) {
972 adjusted_lim = igvn->transform(new SubINode(hi, lo));
973 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1)));
974 cond = BoolTest::lt;
975 } else if (lo_test == BoolTest::gt || lo_test == BoolTest::le) {
976 adjusted_lim = igvn->transform(new SubINode(hi, lo));
977 lo = igvn->transform(new AddINode(lo, igvn->intcon(1)));
978 cond = BoolTest::lt;
979 } else {
980 assert(false, "unhandled lo_test: %d", lo_test);
981 return false;
982 }
983 } else {
984 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled hi_test: %d", hi_test);
985 return false;
986 }
987 // this test was canonicalized
988 assert(this_bool->_test.is_less() && fail->_con, "incorrect test");
989 } else if (lo_type != nullptr && hi_type != nullptr && lo_type->_lo > hi_type->_hi &&
990 lo_type->_hi == max_jint && hi_type->_lo == min_jint && lo_test != BoolTest::ne) {
991
992 // this_bool = <
993 // dom_bool = < (proj = True) or dom_bool = >= (proj = False)
994 // x in [b, a[ on the fail (= False) projection, a > b-1 (because of lo_type->_lo > hi_type->_hi above):
995 // lo = b, hi = a, adjusted_lim = a-b, cond = >=u
996 // dom_bool = <= (proj = True) or dom_bool = > (proj = False)
997 // x in [b, a] on the fail (= False) projection, a+1 > b-1:
998 // lo = b, hi = a, adjusted_lim = a-b+1, cond = >=u
999 // lo = b, hi = a, adjusted_lim = a-b, cond = >u doesn't work because a = b - 1 is possible, then b-a = -1
1000 // this_bool = <=
1001 // dom_bool = < (proj = True) or dom_bool = >= (proj = False)
1002 // x in ]b, a[ on the fail (= False) projection, a > b:
1003 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >=u
1004 // dom_bool = <= (proj = True) or dom_bool = > (proj = False)
1005 // x in ]b, a] on the fail (= False) projection, a+1 > b:
1006 // lo = b+1, hi = a, adjusted_lim = a-b, cond = >=u
1007 // 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
1008
1009 swap(lo, hi);
1010 swap(lo_type, hi_type);
1011 swap(lo_test, hi_test);
1012
1013 assert((dom_bool->_test.is_less() && proj->_con) ||
1014 (dom_bool->_test.is_greater() && !proj->_con), "incorrect test");
1015
1016 cond = (hi_test == BoolTest::le || hi_test == BoolTest::gt) ? BoolTest::gt : BoolTest::ge;
1017
1018 if (lo_test == BoolTest::lt) {
1019 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) {
1020 cond = BoolTest::ge;
1021 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) {
1022 adjusted_lim = igvn->transform(new SubINode(hi, lo));
1023 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1)));
1024 cond = BoolTest::ge;
1025 } else {
1026 assert(false, "unhandled hi_test: %d", hi_test);
1027 return false;
1028 }
1029 } else if (lo_test == BoolTest::le) {
1030 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) {
1031 lo = igvn->transform(new AddINode(lo, igvn->intcon(1)));
1032 cond = BoolTest::ge;
1033 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) {
1034 adjusted_lim = igvn->transform(new SubINode(hi, lo));
1035 lo = igvn->transform(new AddINode(lo, igvn->intcon(1)));
1036 cond = BoolTest::ge;
1037 } else {
1038 assert(false, "unhandled hi_test: %d", hi_test);
1039 return false;
1040 }
1041 } else {
1042 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled lo_test: %d", lo_test);
1043 return false;
1044 }
1045 // this test was canonicalized
1046 assert(this_bool->_test.is_less() && !fail->_con, "incorrect test");
1047 } else {
1048 const TypeInt* failtype = filtered_int_type(igvn, n, proj);
1049 if (failtype != nullptr) {
1050 const TypeInt* type2 = filtered_int_type(igvn, n, fail);
1051 if (type2 != nullptr) {
1052 failtype = failtype->join(type2)->is_int();
1053 if (failtype->empty()) {
1054 // previous if determines the result of this if so
1055 // replace Bool with constant
1056 igvn->replace_input_of(this, 1, igvn->intcon(success->_con));
1057 return true;
1058 }
1059 }
1060 }
1061 return false;
1062 }
1063
1064 assert(lo != nullptr && hi != nullptr, "sanity");
1065 Node* hook = new Node(lo); // Add a use to lo to prevent him from dying
1066 // Merge the two compares into a single unsigned compare by building (CmpU (n - lo) (hi - lo))
1067 Node* adjusted_val = igvn->transform(new SubINode(n, lo));
1068 if (adjusted_lim == nullptr) {
1069 adjusted_lim = igvn->transform(new SubINode(hi, lo));
1070 }
1071 hook->destruct(igvn);
1072
1073 if (adjusted_val->is_top() || adjusted_lim->is_top()) {
1074 return false;
1075 }
1076
1077 if (igvn->type(adjusted_lim)->is_int()->_lo < 0 &&
1078 !igvn->C->post_loop_opts_phase()) {
1079 // If range check elimination applies to this comparison, it includes code to protect from overflows that may
1080 // cause the main loop to be skipped entirely. Delay this transformation.
1081 // Example:
1082 // for (int i = 0; i < limit; i++) {
1083 // if (i < max_jint && i > min_jint) {...
1084 // }
1085 // Comparisons folded as:
1086 // i - min_jint - 1 <u -2
1087 // when RC applies, main loop limit becomes:
1088 // min(limit, max(-2 + min_jint + 1, min_jint))
1089 // = min(limit, min_jint)
1090 // = min_jint
1091 if (adjusted_val->outcnt() == 0) {
1092 igvn->remove_dead_node(adjusted_val);
1093 }
1094 if (adjusted_lim->outcnt() == 0) {
1095 igvn->remove_dead_node(adjusted_lim);
1096 }
1097 igvn->C->record_for_post_loop_opts_igvn(this);
1098 return false;
1099 }
1100
1101 Node* newcmp = igvn->transform(new CmpUNode(adjusted_val, adjusted_lim));
1102 Node* newbool = igvn->transform(new BoolNode(newcmp, cond));
1103
1104 igvn->replace_input_of(dom_iff, 1, igvn->intcon(proj->_con));
1105 igvn->replace_input_of(this, 1, newbool);
1106
1107 return true;
1108 }
1109
1110 // Merge the branches that trap for this If and the dominating If into
1111 // a single region that branches to the uncommon trap for the
1112 // dominating If
1113 Node* IfNode::merge_uncommon_traps(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) {
1114 Node* res = this;
1115 assert(success->in(0) == this, "bad projection");
1116
1117 ProjNode* otherproj = proj->other_if_proj();
1118
1119 CallStaticJavaNode* unc = success->is_uncommon_trap_proj();
1120 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj();
1121
1122 if (unc != dom_unc) {
1123 Node* r = new RegionNode(3);
1124
1125 r->set_req(1, otherproj);
1126 r->set_req(2, success);
1127 r = igvn->transform(r);
1128 assert(r->is_Region(), "can't go away");
1129
1130 // Make both If trap at the state of the first If: once the CmpI
1131 // nodes are merged, if we trap we don't know which of the CmpI
1132 // nodes would have caused the trap so we have to restart
1133 // execution at the first one
1134 igvn->replace_input_of(dom_unc, 0, r);
1135 igvn->replace_input_of(unc, 0, igvn->C->top());
1136 }
1137 int trap_request = dom_unc->uncommon_trap_request();
1138 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1139 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1140
1141 int flip_test = 0;
1142 Node* l = nullptr;
1143 Node* r = nullptr;
1144
1145 if (success->in(0)->as_If()->range_check_trap_proj(flip_test, l, r) != nullptr) {
1146 // If this looks like a range check, change the trap to
1147 // Reason_range_check so the compiler recognizes it as a range
1148 // check and applies the corresponding optimizations
1149 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_range_check, action);
1150
1151 improve_address_types(l, r, fail, igvn);
1152
1153 res = igvn->transform(new RangeCheckNode(in(0), in(1), _prob, _fcnt));
1154 } else if (unc != dom_unc) {
1155 // If we trap we won't know what CmpI would have caused the trap
1156 // so use a special trap reason to mark this pair of CmpI nodes as
1157 // bad candidate for folding. On recompilation we won't fold them
1158 // and we may trap again but this time we'll know what branch
1159 // traps
1160 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_unstable_fused_if, action);
1161 }
1162 igvn->replace_input_of(dom_unc, TypeFunc::Parms, igvn->intcon(trap_request));
1163 return res;
1164 }
1165
1166 // If we are turning 2 CmpI nodes into a CmpU that follows the pattern
1167 // of a rangecheck on index i, on 64 bit the compares may be followed
1168 // by memory accesses using i as index. In that case, the CmpU tells
1169 // us something about the values taken by i that can help the compiler
1170 // (see Compile::conv_I2X_index())
1171 void IfNode::improve_address_types(Node* l, Node* r, ProjNode* fail, PhaseIterGVN* igvn) {
1172 #ifdef _LP64
1173 ResourceMark rm;
1174 Node_Stack stack(2);
1175
1176 assert(r->Opcode() == Op_LoadRange, "unexpected range check");
1177 const TypeInt* array_size = igvn->type(r)->is_int();
1178
1179 stack.push(l, 0);
1180
1181 while(stack.size() > 0) {
1182 Node* n = stack.node();
1183 uint start = stack.index();
1184
1185 uint i = start;
1186 for (; i < n->outcnt(); i++) {
1187 Node* use = n->raw_out(i);
1188 if (stack.size() == 1) {
1189 if (use->Opcode() == Op_ConvI2L) {
1190 const TypeLong* bounds = use->as_Type()->type()->is_long();
1191 if (bounds->_lo <= array_size->_lo && bounds->_hi >= array_size->_hi &&
1192 (bounds->_lo != array_size->_lo || bounds->_hi != array_size->_hi)) {
1193 stack.set_index(i+1);
1194 stack.push(use, 0);
1195 break;
1196 }
1197 }
1198 } else if (use->is_Mem()) {
1199 Node* ctrl = use->in(0);
1200 for (int i = 0; i < 10 && ctrl != nullptr && ctrl != fail; i++) {
1201 ctrl = up_one_dom(ctrl);
1202 }
1203 if (ctrl == fail) {
1204 Node* init_n = stack.node_at(1);
1205 assert(init_n->Opcode() == Op_ConvI2L, "unexpected first node");
1206 // Create a new narrow ConvI2L node that is dependent on the range check
1207 Node* new_n = igvn->C->conv_I2X_index(igvn, l, array_size, fail);
1208
1209 // The type of the ConvI2L may be widen and so the new
1210 // ConvI2L may not be better than an existing ConvI2L
1211 if (new_n != init_n) {
1212 for (uint j = 2; j < stack.size(); j++) {
1213 Node* n = stack.node_at(j);
1214 Node* clone = n->clone();
1215 int rep = clone->replace_edge(init_n, new_n, igvn);
1216 assert(rep > 0, "can't find expected node?");
1217 clone = igvn->transform(clone);
1218 init_n = n;
1219 new_n = clone;
1220 }
1221 igvn->hash_delete(use);
1222 int rep = use->replace_edge(init_n, new_n, igvn);
1223 assert(rep > 0, "can't find expected node?");
1224 igvn->transform(use);
1225 if (init_n->outcnt() == 0) {
1226 igvn->_worklist.push(init_n);
1227 }
1228 }
1229 }
1230 } else if (use->in(0) == nullptr && (igvn->type(use)->isa_long() ||
1231 igvn->type(use)->isa_ptr())) {
1232 stack.set_index(i+1);
1233 stack.push(use, 0);
1234 break;
1235 }
1236 }
1237 if (i == n->outcnt()) {
1238 stack.pop();
1239 }
1240 }
1241 #endif
1242 }
1243
1244 bool IfNode::is_cmp_with_loadrange(ProjNode* proj) {
1245 if (in(1) != nullptr &&
1246 in(1)->in(1) != nullptr &&
1247 in(1)->in(1)->in(2) != nullptr) {
1248 Node* other = in(1)->in(1)->in(2);
1249 if (other->Opcode() == Op_LoadRange &&
1250 ((other->in(0) != nullptr && other->in(0) == proj) ||
1251 (other->in(0) == nullptr &&
1252 other->in(2) != nullptr &&
1253 other->in(2)->is_AddP() &&
1254 other->in(2)->in(1) != nullptr &&
1255 other->in(2)->in(1)->Opcode() == Op_CastPP &&
1256 other->in(2)->in(1)->in(0) == proj))) {
1257 return true;
1258 }
1259 }
1260 return false;
1261 }
1262
1263 bool IfNode::is_null_check(ProjNode* proj, PhaseIterGVN* igvn) {
1264 Node* other = in(1)->in(1)->in(2);
1265 if (other->in(MemNode::Address) != nullptr &&
1266 proj->in(0)->in(1) != nullptr &&
1267 proj->in(0)->in(1)->is_Bool() &&
1268 proj->in(0)->in(1)->in(1) != nullptr &&
1269 proj->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1270 proj->in(0)->in(1)->in(1)->in(2) != nullptr &&
1271 proj->in(0)->in(1)->in(1)->in(1) == other->in(MemNode::Address)->in(AddPNode::Address)->uncast() &&
1272 igvn->type(proj->in(0)->in(1)->in(1)->in(2)) == TypePtr::NULL_PTR) {
1273 return true;
1274 }
1275 return false;
1276 }
1277
1278 // Check that the If that is in between the 2 integer comparisons has
1279 // no side effect
1280 bool IfNode::is_side_effect_free_test(ProjNode* proj, PhaseIterGVN* igvn) {
1281 if (proj == nullptr) {
1282 return false;
1283 }
1284 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern();
1285 if (unc != nullptr && proj->outcnt() <= 2) {
1286 if (proj->outcnt() == 1 ||
1287 // Allow simple null check from LoadRange
1288 (is_cmp_with_loadrange(proj) && is_null_check(proj, igvn))) {
1289 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern();
1290 CallStaticJavaNode* dom_unc = proj->in(0)->in(0)->as_Proj()->is_uncommon_trap_if_pattern();
1291 assert(dom_unc != nullptr, "is_uncommon_trap_if_pattern returned null");
1292
1293 // reroute_side_effect_free_unc changes the state of this
1294 // uncommon trap to restart execution at the previous
1295 // CmpI. Check that this change in a previous compilation didn't
1296 // cause too many traps.
1297 int trap_request = unc->uncommon_trap_request();
1298 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1299
1300 if (igvn->C->too_many_traps(dom_unc->jvms()->method(), dom_unc->jvms()->bci(), reason)) {
1301 return false;
1302 }
1303
1304 if (!is_dominator_unc(dom_unc, unc)) {
1305 return false;
1306 }
1307
1308 return true;
1309 }
1310 }
1311 return false;
1312 }
1313
1314 // Make the If between the 2 integer comparisons trap at the state of
1315 // the first If: the last CmpI is the one replaced by a CmpU and the
1316 // first CmpI is eliminated, so the test between the 2 CmpI nodes
1317 // won't be guarded by the first CmpI anymore. It can trap in cases
1318 // where the first CmpI would have prevented it from executing: on a
1319 // trap, we need to restart execution at the state of the first CmpI
1320 void IfNode::reroute_side_effect_free_unc(ProjNode* proj, ProjNode* dom_proj, PhaseIterGVN* igvn) {
1321 CallStaticJavaNode* dom_unc = dom_proj->is_uncommon_trap_if_pattern();
1322 ProjNode* otherproj = proj->other_if_proj();
1323 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern();
1324 Node* call_proj = dom_unc->unique_ctrl_out();
1325 Node* halt = call_proj->unique_ctrl_out();
1326
1327 Node* new_unc = dom_unc->clone();
1328 call_proj = call_proj->clone();
1329 halt = halt->clone();
1330 Node* c = otherproj->clone();
1331
1332 c = igvn->transform(c);
1333 new_unc->set_req(TypeFunc::Parms, unc->in(TypeFunc::Parms));
1334 new_unc->set_req(0, c);
1335 new_unc = igvn->transform(new_unc);
1336 call_proj->set_req(0, new_unc);
1337 call_proj = igvn->transform(call_proj);
1338 halt->set_req(0, call_proj);
1339 halt = igvn->transform(halt);
1340
1341 igvn->replace_node(otherproj, igvn->C->top());
1342 igvn->C->root()->add_req(halt);
1343 }
1344
1345 Node* IfNode::fold_compares(PhaseIterGVN* igvn) {
1346 if (Opcode() != Op_If) return nullptr;
1347
1348 if (cmpi_folds(igvn)) {
1349 Node* ctrl = in(0);
1350 if (is_ctrl_folds(ctrl, igvn)) {
1351 // A integer comparison immediately dominated by another integer
1352 // comparison
1353 ProjNode* success = nullptr;
1354 ProjNode* fail = nullptr;
1355 ProjNode* dom_cmp = ctrl->as_Proj();
1356 if (has_shared_region(dom_cmp, success, fail) &&
1357 // Next call modifies graph so must be last
1358 fold_compares_helper(dom_cmp, success, fail, igvn)) {
1359 return this;
1360 }
1361 if (has_only_uncommon_traps(dom_cmp, success, fail, igvn) &&
1362 // Next call modifies graph so must be last
1363 fold_compares_helper(dom_cmp, success, fail, igvn)) {
1364 return merge_uncommon_traps(dom_cmp, success, fail, igvn);
1365 }
1366 return nullptr;
1367 } else if (ctrl->in(0) != nullptr &&
1368 ctrl->in(0)->in(0) != nullptr) {
1369 ProjNode* success = nullptr;
1370 ProjNode* fail = nullptr;
1371 Node* dom = ctrl->in(0)->in(0);
1372 ProjNode* dom_cmp = dom->isa_Proj();
1373 ProjNode* other_cmp = ctrl->isa_Proj();
1374
1375 // Check if it's an integer comparison dominated by another
1376 // integer comparison with another test in between
1377 if (is_ctrl_folds(dom, igvn) &&
1378 has_only_uncommon_traps(dom_cmp, success, fail, igvn) &&
1379 is_side_effect_free_test(other_cmp, igvn) &&
1380 // Next call modifies graph so must be last
1381 fold_compares_helper(dom_cmp, success, fail, igvn)) {
1382 reroute_side_effect_free_unc(other_cmp, dom_cmp, igvn);
1383 return merge_uncommon_traps(dom_cmp, success, fail, igvn);
1384 }
1385 }
1386 }
1387 return nullptr;
1388 }
1389
1390 //------------------------------remove_useless_bool----------------------------
1391 // Check for people making a useless boolean: things like
1392 // if( (x < y ? true : false) ) { ... }
1393 // Replace with if( x < y ) { ... }
1394 static Node *remove_useless_bool(IfNode *iff, PhaseGVN *phase) {
1395 Node *i1 = iff->in(1);
1396 if( !i1->is_Bool() ) return nullptr;
1397 BoolNode *bol = i1->as_Bool();
1398
1399 Node *cmp = bol->in(1);
1400 if( cmp->Opcode() != Op_CmpI ) return nullptr;
1401
1402 // Must be comparing against a bool
1403 const Type *cmp2_t = phase->type( cmp->in(2) );
1404 if( cmp2_t != TypeInt::ZERO &&
1405 cmp2_t != TypeInt::ONE )
1406 return nullptr;
1407
1408 // Find a prior merge point merging the boolean
1409 i1 = cmp->in(1);
1410 if( !i1->is_Phi() ) return nullptr;
1411 PhiNode *phi = i1->as_Phi();
1412 if( phase->type( phi ) != TypeInt::BOOL )
1413 return nullptr;
1414
1415 // Check for diamond pattern
1416 int true_path = phi->is_diamond_phi();
1417 if( true_path == 0 ) return nullptr;
1418
1419 // Make sure that iff and the control of the phi are different. This
1420 // should really only happen for dead control flow since it requires
1421 // an illegal cycle.
1422 if (phi->in(0)->in(1)->in(0) == iff) return nullptr;
1423
1424 // phi->region->if_proj->ifnode->bool->cmp
1425 BoolNode *bol2 = phi->in(0)->in(1)->in(0)->in(1)->as_Bool();
1426
1427 // Now get the 'sense' of the test correct so we can plug in
1428 // either iff2->in(1) or its complement.
1429 int flip = 0;
1430 if( bol->_test._test == BoolTest::ne ) flip = 1-flip;
1431 else if( bol->_test._test != BoolTest::eq ) return nullptr;
1432 if( cmp2_t == TypeInt::ZERO ) flip = 1-flip;
1433
1434 const Type *phi1_t = phase->type( phi->in(1) );
1435 const Type *phi2_t = phase->type( phi->in(2) );
1436 // Check for Phi(0,1) and flip
1437 if( phi1_t == TypeInt::ZERO ) {
1438 if( phi2_t != TypeInt::ONE ) return nullptr;
1439 flip = 1-flip;
1440 } else {
1441 // Check for Phi(1,0)
1442 if( phi1_t != TypeInt::ONE ) return nullptr;
1443 if( phi2_t != TypeInt::ZERO ) return nullptr;
1444 }
1445 if( true_path == 2 ) {
1446 flip = 1-flip;
1447 }
1448
1449 Node* new_bol = (flip ? phase->transform( bol2->negate(phase) ) : bol2);
1450 assert(new_bol != iff->in(1), "must make progress");
1451 iff->set_req_X(1, new_bol, phase);
1452 // Intervening diamond probably goes dead
1453 phase->C->set_major_progress();
1454 return iff;
1455 }
1456
1457 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff);
1458
1459 struct RangeCheck {
1460 IfProjNode* ctl;
1461 jint off;
1462 };
1463
1464 Node* IfNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
1465 if (remove_dead_region(phase, can_reshape)) return this;
1466 // No Def-Use info?
1467 if (!can_reshape) return nullptr;
1468
1469 // Don't bother trying to transform a dead if
1470 if (in(0)->is_top()) return nullptr;
1471 // Don't bother trying to transform an if with a dead test
1472 if (in(1)->is_top()) return nullptr;
1473 // Another variation of a dead test
1474 if (in(1)->is_Con()) return nullptr;
1475 // Another variation of a dead if
1476 if (outcnt() < 2) return nullptr;
1477
1478 // Canonicalize the test.
1479 Node* idt_if = idealize_test(phase, this);
1480 if (idt_if != nullptr) return idt_if;
1481
1482 // Try to split the IF
1483 PhaseIterGVN *igvn = phase->is_IterGVN();
1484 Node *s = split_if(this, igvn);
1485 if (s != nullptr) return s;
1486
1487 return NodeSentinel;
1488 }
1489
1490 //------------------------------Ideal------------------------------------------
1491 // Return a node which is more "ideal" than the current node. Strip out
1492 // control copies
1493 Node* IfNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1494 Node* res = Ideal_common(phase, can_reshape);
1495 if (res != NodeSentinel) {
1496 return res;
1497 }
1498
1499 // Check for people making a useless boolean: things like
1500 // if( (x < y ? true : false) ) { ... }
1501 // Replace with if( x < y ) { ... }
1502 Node* bol2 = remove_useless_bool(this, phase);
1503 if (bol2) return bol2;
1504
1505 if (in(0) == nullptr) return nullptr; // Dead loop?
1506
1507 PhaseIterGVN* igvn = phase->is_IterGVN();
1508 Node* result = fold_compares(igvn);
1509 if (result != nullptr) {
1510 return result;
1511 }
1512
1513 // Scan for an equivalent test
1514 int dist = 4; // Cutoff limit for search
1515 if (is_If() && in(1)->is_Bool()) {
1516 Node* cmp = in(1)->in(1);
1517 if (cmp->Opcode() == Op_CmpP &&
1518 cmp->in(2) != nullptr && // make sure cmp is not already dead
1519 cmp->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1520 dist = 64; // Limit for null-pointer scans
1521 }
1522 }
1523
1524 Node* prev_dom = search_identical(dist, igvn);
1525
1526 if (prev_dom != nullptr) {
1527 // Replace dominated IfNode
1528 return dominated_by(prev_dom, igvn, false);
1529 }
1530
1531 return simple_subsuming(igvn);
1532 }
1533
1534 //------------------------------dominated_by-----------------------------------
1535 Node* IfNode::dominated_by(Node* prev_dom, PhaseIterGVN* igvn, bool pin_array_access_nodes) {
1536 #ifndef PRODUCT
1537 if (TraceIterativeGVN) {
1538 tty->print(" Removing IfNode: "); this->dump();
1539 }
1540 #endif
1541
1542 igvn->hash_delete(this); // Remove self to prevent spurious V-N
1543 Node *idom = in(0);
1544 // Need opcode to decide which way 'this' test goes
1545 int prev_op = prev_dom->Opcode();
1546 Node *top = igvn->C->top(); // Shortcut to top
1547
1548 // Now walk the current IfNode's projections.
1549 // Loop ends when 'this' has no more uses.
1550 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) {
1551 Node *ifp = last_out(i); // Get IfTrue/IfFalse
1552 igvn->add_users_to_worklist(ifp);
1553 // Check which projection it is and set target.
1554 // Data-target is either the dominating projection of the same type
1555 // or TOP if the dominating projection is of opposite type.
1556 // Data-target will be used as the new control edge for the non-CFG
1557 // nodes like Casts and Loads.
1558 Node *data_target = (ifp->Opcode() == prev_op) ? prev_dom : top;
1559 // Control-target is just the If's immediate dominator or TOP.
1560 Node *ctrl_target = (ifp->Opcode() == prev_op) ? idom : top;
1561
1562 // For each child of an IfTrue/IfFalse projection, reroute.
1563 // Loop ends when projection has no more uses.
1564 for (DUIterator_Last jmin, j = ifp->last_outs(jmin); j >= jmin; --j) {
1565 Node* s = ifp->last_out(j); // Get child of IfTrue/IfFalse
1566 if (s->depends_only_on_test() && igvn->no_dependent_zero_check(s)) {
1567 // For control producers.
1568 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check.
1569 igvn->replace_input_of(s, 0, data_target); // Move child to data-target
1570 if (pin_array_access_nodes && data_target != top) {
1571 // As a result of range check smearing, Loads and range check Cast nodes that are control dependent on this
1572 // range check (that is about to be removed) now depend on multiple dominating range checks. After the removal
1573 // of this range check, these control dependent nodes end up at the lowest/nearest dominating check in the
1574 // graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the
1575 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the
1576 // lowest dominating check.
1577 Node* clone = s->pin_array_access_node();
1578 if (clone != nullptr) {
1579 clone = igvn->transform(clone);
1580 igvn->replace_node(s, clone);
1581 }
1582 }
1583 } else {
1584 // Find the control input matching this def-use edge.
1585 // For Regions it may not be in slot 0.
1586 uint l;
1587 for (l = 0; s->in(l) != ifp; l++) { }
1588 igvn->replace_input_of(s, l, ctrl_target);
1589 }
1590 } // End for each child of a projection
1591
1592 igvn->remove_dead_node(ifp);
1593 } // End for each IfTrue/IfFalse child of If
1594
1595 // Kill the IfNode
1596 igvn->remove_dead_node(this);
1597
1598 // Must return either the original node (now dead) or a new node
1599 // (Do not return a top here, since that would break the uniqueness of top.)
1600 return new ConINode(TypeInt::ZERO);
1601 }
1602
1603 Node* IfNode::search_identical(int dist, PhaseIterGVN* igvn) {
1604 // Setup to scan up the CFG looking for a dominating test
1605 Node* dom = in(0);
1606 Node* prev_dom = this;
1607 int op = Opcode();
1608 // Search up the dominator tree for an If with an identical test
1609 while (dom->Opcode() != op || // Not same opcode?
1610 !same_condition(dom, igvn) || // Not same input 1?
1611 prev_dom->in(0) != dom) { // One path of test does not dominate?
1612 if (dist < 0) return nullptr;
1613
1614 dist--;
1615 prev_dom = dom;
1616 dom = up_one_dom(dom);
1617 if (!dom) return nullptr;
1618 }
1619
1620 // Check that we did not follow a loop back to ourselves
1621 if (this == dom) {
1622 return nullptr;
1623 }
1624
1625 #ifndef PRODUCT
1626 if (dist > 2) { // Add to count of null checks elided
1627 explicit_null_checks_elided++;
1628 }
1629 #endif
1630
1631 return prev_dom;
1632 }
1633
1634 bool IfNode::same_condition(const Node* dom, PhaseIterGVN* igvn) const {
1635 Node* dom_bool = dom->in(1);
1636 Node* this_bool = in(1);
1637 if (dom_bool == this_bool) {
1638 return true;
1639 }
1640
1641 if (dom_bool == nullptr || !dom_bool->is_Bool() ||
1642 this_bool == nullptr || !this_bool->is_Bool()) {
1643 return false;
1644 }
1645 Node* dom_cmp = dom_bool->in(1);
1646 Node* this_cmp = this_bool->in(1);
1647
1648 // If the comparison is a subtype check, then SubTypeCheck nodes may have profile data attached to them and may be
1649 // different nodes even-though they perform the same subtype check
1650 if (dom_cmp == nullptr || !dom_cmp->is_SubTypeCheck() ||
1651 this_cmp == nullptr || !this_cmp->is_SubTypeCheck()) {
1652 return false;
1653 }
1654
1655 if (dom_cmp->in(1) != this_cmp->in(1) ||
1656 dom_cmp->in(2) != this_cmp->in(2) ||
1657 dom_bool->as_Bool()->_test._test != this_bool->as_Bool()->_test._test) {
1658 return false;
1659 }
1660
1661 return true;
1662 }
1663
1664
1665 static int subsuming_bool_test_encode(Node*);
1666
1667 // Check if dominating test is subsuming 'this' one.
1668 //
1669 // cmp
1670 // / \
1671 // (r1) bool \
1672 // / bool (r2)
1673 // (dom) if \
1674 // \ )
1675 // (pre) if[TF] /
1676 // \ /
1677 // if (this)
1678 // \r1
1679 // r2\ eqT eqF neT neF ltT ltF leT leF gtT gtF geT geF
1680 // eq t f f t f - - f f - - f
1681 // ne f t t f t - - t t - - t
1682 // lt f - - f t f - f f - f t
1683 // le t - - t t - t f f t - t
1684 // gt f - - f f - f t t f - f
1685 // ge t - - t f t - t t - t f
1686 //
1687 Node* IfNode::simple_subsuming(PhaseIterGVN* igvn) {
1688 // Table encoding: N/A (na), True-branch (tb), False-branch (fb).
1689 static enum { na, tb, fb } s_short_circuit_map[6][12] = {
1690 /*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*/
1691 /*eq*/{ tb, fb, fb, tb, fb, na, na, fb, fb, na, na, fb },
1692 /*ne*/{ fb, tb, tb, fb, tb, na, na, tb, tb, na, na, tb },
1693 /*lt*/{ fb, na, na, fb, tb, fb, na, fb, fb, na, fb, tb },
1694 /*le*/{ tb, na, na, tb, tb, na, tb, fb, fb, tb, na, tb },
1695 /*gt*/{ fb, na, na, fb, fb, na, fb, tb, tb, fb, na, fb },
1696 /*ge*/{ tb, na, na, tb, fb, tb, na, tb, tb, na, tb, fb }};
1697
1698 Node* pre = in(0);
1699 if (!pre->is_IfTrue() && !pre->is_IfFalse()) {
1700 return nullptr;
1701 }
1702 Node* dom = pre->in(0);
1703 if (!dom->is_If()) {
1704 return nullptr;
1705 }
1706 Node* bol = in(1);
1707 if (!bol->is_Bool()) {
1708 return nullptr;
1709 }
1710 Node* cmp = in(1)->in(1);
1711 if (!cmp->is_Cmp()) {
1712 return nullptr;
1713 }
1714
1715 if (!dom->in(1)->is_Bool()) {
1716 return nullptr;
1717 }
1718 if (dom->in(1)->in(1) != cmp) { // Not same cond?
1719 return nullptr;
1720 }
1721
1722 int drel = subsuming_bool_test_encode(dom->in(1));
1723 int trel = subsuming_bool_test_encode(bol);
1724 int bout = pre->is_IfFalse() ? 1 : 0;
1725
1726 if (drel < 0 || trel < 0) {
1727 return nullptr;
1728 }
1729 int br = s_short_circuit_map[trel][2*drel+bout];
1730 if (br == na) {
1731 return nullptr;
1732 }
1733 #ifndef PRODUCT
1734 if (TraceIterativeGVN) {
1735 tty->print(" Subsumed IfNode: "); dump();
1736 }
1737 #endif
1738 // Replace condition with constant True(1)/False(0).
1739 bool is_always_true = br == tb;
1740 set_req(1, igvn->intcon(is_always_true ? 1 : 0));
1741
1742 // Update any data dependencies to the directly dominating test. This subsumed test is not immediately removed by igvn
1743 // and therefore subsequent optimizations might miss these data dependencies otherwise. There might be a dead loop
1744 // ('always_taken_proj' == 'pre') that is cleaned up later. Skip this case to make the iterator work properly.
1745 Node* always_taken_proj = proj_out(is_always_true);
1746 if (always_taken_proj != pre) {
1747 for (DUIterator_Fast imax, i = always_taken_proj->fast_outs(imax); i < imax; i++) {
1748 Node* u = always_taken_proj->fast_out(i);
1749 if (!u->is_CFG()) {
1750 igvn->replace_input_of(u, 0, pre);
1751 --i;
1752 --imax;
1753 }
1754 }
1755 }
1756
1757 if (bol->outcnt() == 0) {
1758 igvn->remove_dead_node(bol); // Kill the BoolNode.
1759 }
1760 return this;
1761 }
1762
1763 // Map BoolTest to local table encoding. The BoolTest (e)numerals
1764 // { eq = 0, ne = 4, le = 5, ge = 7, lt = 3, gt = 1 }
1765 // are mapped to table indices, while the remaining (e)numerals in BoolTest
1766 // { overflow = 2, no_overflow = 6, never = 8, illegal = 9 }
1767 // are ignored (these are not modeled in the table).
1768 //
1769 static int subsuming_bool_test_encode(Node* node) {
1770 precond(node->is_Bool());
1771 BoolTest::mask x = node->as_Bool()->_test._test;
1772 switch (x) {
1773 case BoolTest::eq: return 0;
1774 case BoolTest::ne: return 1;
1775 case BoolTest::lt: return 2;
1776 case BoolTest::le: return 3;
1777 case BoolTest::gt: return 4;
1778 case BoolTest::ge: return 5;
1779 case BoolTest::overflow:
1780 case BoolTest::no_overflow:
1781 case BoolTest::never:
1782 case BoolTest::illegal:
1783 default:
1784 return -1;
1785 }
1786 }
1787
1788 //------------------------------Identity---------------------------------------
1789 // If the test is constant & we match, then we are the input Control
1790 Node* IfProjNode::Identity(PhaseGVN* phase) {
1791 // Can only optimize if cannot go the other way
1792 const TypeTuple *t = phase->type(in(0))->is_tuple();
1793 if (t == TypeTuple::IFNEITHER || (always_taken(t) &&
1794 // During parsing (GVN) we don't remove dead code aggressively.
1795 // Cut off dead branch and let PhaseRemoveUseless take care of it.
1796 (!phase->is_IterGVN() ||
1797 // During IGVN, first wait for the dead branch to be killed.
1798 // Otherwise, the IfNode's control will have two control uses (the IfNode
1799 // that doesn't go away because it still has uses and this branch of the
1800 // If) which breaks other optimizations. Node::has_special_unique_user()
1801 // will cause this node to be reprocessed once the dead branch is killed.
1802 in(0)->outcnt() == 1))) {
1803 // IfNode control
1804 if (in(0)->is_BaseCountedLoopEnd()) {
1805 // CountedLoopEndNode may be eliminated by if subsuming, replace CountedLoopNode with LoopNode to
1806 // avoid mismatching between CountedLoopNode and CountedLoopEndNode in the following optimization.
1807 Node* head = unique_ctrl_out_or_null();
1808 if (head != nullptr && head->is_BaseCountedLoop() && head->in(LoopNode::LoopBackControl) == this) {
1809 Node* new_head = new LoopNode(head->in(LoopNode::EntryControl), this);
1810 phase->is_IterGVN()->register_new_node_with_optimizer(new_head);
1811 phase->is_IterGVN()->replace_node(head, new_head);
1812 }
1813 }
1814 return in(0)->in(0);
1815 }
1816 // no progress
1817 return this;
1818 }
1819
1820 bool IfNode::is_zero_trip_guard() const {
1821 if (in(1)->is_Bool() && in(1)->in(1)->is_Cmp()) {
1822 return in(1)->in(1)->in(1)->Opcode() == Op_OpaqueZeroTripGuard;
1823 }
1824 return false;
1825 }
1826
1827 void IfProjNode::pin_array_access_nodes(PhaseIterGVN* igvn) {
1828 for (DUIterator i = outs(); has_out(i); i++) {
1829 Node* u = out(i);
1830 if (!u->depends_only_on_test()) {
1831 continue;
1832 }
1833 Node* clone = u->pin_array_access_node();
1834 if (clone != nullptr) {
1835 clone = igvn->transform(clone);
1836 assert(clone != u, "shouldn't common");
1837 igvn->replace_node(u, clone);
1838 --i;
1839 }
1840 }
1841 }
1842
1843 #ifndef PRODUCT
1844 void IfNode::dump_spec(outputStream* st) const {
1845 switch (_assertion_predicate_type) {
1846 case AssertionPredicateType::InitValue:
1847 st->print("#Init Value Assertion Predicate ");
1848 break;
1849 case AssertionPredicateType::LastValue:
1850 st->print("#Last Value Assertion Predicate ");
1851 break;
1852 case AssertionPredicateType::FinalIv:
1853 st->print("#Final IV Assertion Predicate ");
1854 break;
1855 case AssertionPredicateType::None:
1856 // No Assertion Predicate
1857 break;
1858 default:
1859 fatal("Unknown Assertion Predicate type");
1860 }
1861 st->print("P=%f, C=%f", _prob, _fcnt);
1862 }
1863 #endif // NOT PRODUCT
1864
1865 //------------------------------idealize_test----------------------------------
1866 // Try to canonicalize tests better. Peek at the Cmp/Bool/If sequence and
1867 // come up with a canonical sequence. Bools getting 'eq', 'gt' and 'ge' forms
1868 // converted to 'ne', 'le' and 'lt' forms. IfTrue/IfFalse get swapped as
1869 // needed.
1870 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff) {
1871 assert(iff->in(0) != nullptr, "If must be live");
1872
1873 if (iff->outcnt() != 2) return nullptr; // Malformed projections.
1874 Node* old_if_f = iff->proj_out(false);
1875 Node* old_if_t = iff->proj_out(true);
1876
1877 // CountedLoopEnds want the back-control test to be TRUE, regardless of
1878 // whether they are testing a 'gt' or 'lt' condition. The 'gt' condition
1879 // happens in count-down loops
1880 if (iff->is_BaseCountedLoopEnd()) return nullptr;
1881 if (!iff->in(1)->is_Bool()) return nullptr; // Happens for partially optimized IF tests
1882 BoolNode *b = iff->in(1)->as_Bool();
1883 BoolTest bt = b->_test;
1884 // Test already in good order?
1885 if( bt.is_canonical() )
1886 return nullptr;
1887
1888 // Flip test to be canonical. Requires flipping the IfFalse/IfTrue and
1889 // cloning the IfNode.
1890 Node* new_b = phase->transform( new BoolNode(b->in(1), bt.negate()) );
1891 if( !new_b->is_Bool() ) return nullptr;
1892 b = new_b->as_Bool();
1893
1894 PhaseIterGVN *igvn = phase->is_IterGVN();
1895 assert( igvn, "Test is not canonical in parser?" );
1896
1897 // The IF node never really changes, but it needs to be cloned
1898 iff = iff->clone()->as_If();
1899 iff->set_req(1, b);
1900 iff->_prob = 1.0-iff->_prob;
1901
1902 Node *prior = igvn->hash_find_insert(iff);
1903 if( prior ) {
1904 igvn->remove_dead_node(iff);
1905 iff = (IfNode*)prior;
1906 } else {
1907 // Cannot call transform on it just yet
1908 igvn->set_type_bottom(iff);
1909 }
1910 igvn->_worklist.push(iff);
1911
1912 // Now handle projections. Cloning not required.
1913 Node* new_if_f = (Node*)(new IfFalseNode( iff ));
1914 Node* new_if_t = (Node*)(new IfTrueNode ( iff ));
1915
1916 igvn->register_new_node_with_optimizer(new_if_f);
1917 igvn->register_new_node_with_optimizer(new_if_t);
1918 // Flip test, so flip trailing control
1919 igvn->replace_node(old_if_f, new_if_t);
1920 igvn->replace_node(old_if_t, new_if_f);
1921
1922 // Progress
1923 return iff;
1924 }
1925
1926 Node* RangeCheckNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1927 Node* res = Ideal_common(phase, can_reshape);
1928 if (res != NodeSentinel) {
1929 return res;
1930 }
1931
1932 PhaseIterGVN *igvn = phase->is_IterGVN();
1933 // Setup to scan up the CFG looking for a dominating test
1934 Node* prev_dom = this;
1935
1936 // Check for range-check vs other kinds of tests
1937 Node* index1;
1938 Node* range1;
1939 jint offset1;
1940 int flip1 = is_range_check(range1, index1, offset1);
1941 if (flip1) {
1942 Node* dom = in(0);
1943 // Try to remove extra range checks. All 'up_one_dom' gives up at merges
1944 // so all checks we inspect post-dominate the top-most check we find.
1945 // If we are going to fail the current check and we reach the top check
1946 // then we are guaranteed to fail, so just start interpreting there.
1947 // We 'expand' the top 3 range checks to include all post-dominating
1948 // checks.
1949 //
1950 // Example:
1951 // a[i+x] // (1) 1 < x < 6
1952 // a[i+3] // (2)
1953 // a[i+4] // (3)
1954 // a[i+6] // max = max of all constants
1955 // a[i+2]
1956 // a[i+1] // min = min of all constants
1957 //
1958 // If x < 3:
1959 // (1) a[i+x]: Leave unchanged
1960 // (2) a[i+3]: Replace with a[i+max] = a[i+6]: i+x < i+3 <= i+6 -> (2) is covered
1961 // (3) a[i+4]: Replace with a[i+min] = a[i+1]: i+1 < i+4 <= i+6 -> (3) and all following checks are covered
1962 // Remove all other a[i+c] checks
1963 //
1964 // If x >= 3:
1965 // (1) a[i+x]: Leave unchanged
1966 // (2) a[i+3]: Replace with a[i+min] = a[i+1]: i+1 < i+3 <= i+x -> (2) is covered
1967 // (3) a[i+4]: Replace with a[i+max] = a[i+6]: i+1 < i+4 <= i+6 -> (3) and all following checks are covered
1968 // Remove all other a[i+c] checks
1969 //
1970 // We only need the top 2 range checks if x is the min or max of all constants.
1971 //
1972 // This, however, only works if the interval [i+min,i+max] is not larger than max_int (i.e. abs(max - min) < max_int):
1973 // The theoretical max size of an array is max_int with:
1974 // - Valid index space: [0,max_int-1]
1975 // - Invalid index space: [max_int,-1] // max_int, min_int, min_int - 1 ..., -1
1976 //
1977 // The size of the consecutive valid index space is smaller than the size of the consecutive invalid index space.
1978 // If we choose min and max in such a way that:
1979 // - abs(max - min) < max_int
1980 // - i+max and i+min are inside the valid index space
1981 // then all indices [i+min,i+max] must be in the valid index space. Otherwise, the invalid index space must be
1982 // smaller than the valid index space which is never the case for any array size.
1983 //
1984 // Choosing a smaller array size only makes the valid index space smaller and the invalid index space larger and
1985 // the argument above still holds.
1986 //
1987 // Note that the same optimization with the same maximal accepted interval size can also be found in C1.
1988 const jlong maximum_number_of_min_max_interval_indices = (jlong)max_jint;
1989
1990 // The top 3 range checks seen
1991 const int NRC = 3;
1992 RangeCheck prev_checks[NRC];
1993 int nb_checks = 0;
1994
1995 // Low and high offsets seen so far
1996 jint off_lo = offset1;
1997 jint off_hi = offset1;
1998
1999 bool found_immediate_dominator = false;
2000
2001 // Scan for the top checks and collect range of offsets
2002 for (int dist = 0; dist < 999; dist++) { // Range-Check scan limit
2003 if (dom->Opcode() == Op_RangeCheck && // Not same opcode?
2004 prev_dom->in(0) == dom) { // One path of test does dominate?
2005 if (dom == this) return nullptr; // dead loop
2006 // See if this is a range check
2007 Node* index2;
2008 Node* range2;
2009 jint offset2;
2010 int flip2 = dom->as_RangeCheck()->is_range_check(range2, index2, offset2);
2011 // See if this is a _matching_ range check, checking against
2012 // the same array bounds.
2013 if (flip2 == flip1 && range2 == range1 && index2 == index1 &&
2014 dom->outcnt() == 2) {
2015 if (nb_checks == 0 && dom->in(1) == in(1)) {
2016 // Found an immediately dominating test at the same offset.
2017 // This kind of back-to-back test can be eliminated locally,
2018 // and there is no need to search further for dominating tests.
2019 assert(offset2 == offset1, "Same test but different offsets");
2020 found_immediate_dominator = true;
2021 break;
2022 }
2023
2024 // "x - y" -> must add one to the difference for number of elements in [x,y]
2025 const jlong diff = (jlong)MIN2(offset2, off_lo) - (jlong)MAX2(offset2, off_hi);
2026 if (ABS(diff) < maximum_number_of_min_max_interval_indices) {
2027 // Gather expanded bounds
2028 off_lo = MIN2(off_lo, offset2);
2029 off_hi = MAX2(off_hi, offset2);
2030 // Record top NRC range checks
2031 prev_checks[nb_checks % NRC].ctl = prev_dom->as_IfProj();
2032 prev_checks[nb_checks % NRC].off = offset2;
2033 nb_checks++;
2034 }
2035 }
2036 }
2037 prev_dom = dom;
2038 dom = up_one_dom(dom);
2039 if (!dom) break;
2040 }
2041
2042 if (!found_immediate_dominator) {
2043 // Attempt to widen the dominating range check to cover some later
2044 // ones. Since range checks "fail" by uncommon-trapping to the
2045 // interpreter, widening a check can make us speculatively enter
2046 // the interpreter. If we see range-check deopt's, do not widen!
2047 if (!phase->C->allow_range_check_smearing()) return nullptr;
2048
2049 if (can_reshape && !phase->C->post_loop_opts_phase()) {
2050 // We are about to perform range check smearing (i.e. remove this RangeCheck if it is dominated by
2051 // a series of RangeChecks which have a range that covers this RangeCheck). This can cause array access nodes to
2052 // be pinned. We want to avoid that and first allow range check elimination a chance to remove the RangeChecks
2053 // from loops. Hence, we delay range check smearing until after loop opts.
2054 phase->C->record_for_post_loop_opts_igvn(this);
2055 return nullptr;
2056 }
2057
2058 // Didn't find prior covering check, so cannot remove anything.
2059 if (nb_checks == 0) {
2060 return nullptr;
2061 }
2062 // Constant indices only need to check the upper bound.
2063 // Non-constant indices must check both low and high.
2064 int chk0 = (nb_checks - 1) % NRC;
2065 if (index1) {
2066 if (nb_checks == 1) {
2067 return nullptr;
2068 } else {
2069 // If the top range check's constant is the min or max of
2070 // all constants we widen the next one to cover the whole
2071 // range of constants.
2072 RangeCheck rc0 = prev_checks[chk0];
2073 int chk1 = (nb_checks - 2) % NRC;
2074 RangeCheck rc1 = prev_checks[chk1];
2075 if (rc0.off == off_lo) {
2076 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn);
2077 prev_dom = rc1.ctl;
2078 } else if (rc0.off == off_hi) {
2079 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn);
2080 prev_dom = rc1.ctl;
2081 } else {
2082 // If the top test's constant is not the min or max of all
2083 // constants, we need 3 range checks. We must leave the
2084 // top test unchanged because widening it would allow the
2085 // accesses it protects to successfully read/write out of
2086 // bounds.
2087 if (nb_checks == 2) {
2088 return nullptr;
2089 }
2090 int chk2 = (nb_checks - 3) % NRC;
2091 RangeCheck rc2 = prev_checks[chk2];
2092 // The top range check a+i covers interval: -a <= i < length-a
2093 // The second range check b+i covers interval: -b <= i < length-b
2094 if (rc1.off <= rc0.off) {
2095 // if b <= a, we change the second range check to:
2096 // -min_of_all_constants <= i < length-min_of_all_constants
2097 // Together top and second range checks now cover:
2098 // -min_of_all_constants <= i < length-a
2099 // which is more restrictive than -b <= i < length-b:
2100 // -b <= -min_of_all_constants <= i < length-a <= length-b
2101 // The third check is then changed to:
2102 // -max_of_all_constants <= i < length-max_of_all_constants
2103 // so 2nd and 3rd checks restrict allowed values of i to:
2104 // -min_of_all_constants <= i < length-max_of_all_constants
2105 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn);
2106 adjust_check(rc2.ctl, range1, index1, flip1, off_hi, igvn);
2107 } else {
2108 // if b > a, we change the second range check to:
2109 // -max_of_all_constants <= i < length-max_of_all_constants
2110 // Together top and second range checks now cover:
2111 // -a <= i < length-max_of_all_constants
2112 // which is more restrictive than -b <= i < length-b:
2113 // -b < -a <= i < length-max_of_all_constants <= length-b
2114 // The third check is then changed to:
2115 // -max_of_all_constants <= i < length-max_of_all_constants
2116 // so 2nd and 3rd checks restrict allowed values of i to:
2117 // -min_of_all_constants <= i < length-max_of_all_constants
2118 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn);
2119 adjust_check(rc2.ctl, range1, index1, flip1, off_lo, igvn);
2120 }
2121 prev_dom = rc2.ctl;
2122 }
2123 }
2124 } else {
2125 RangeCheck rc0 = prev_checks[chk0];
2126 // 'Widen' the offset of the 1st and only covering check
2127 adjust_check(rc0.ctl, range1, index1, flip1, off_hi, igvn);
2128 // Test is now covered by prior checks, dominate it out
2129 prev_dom = rc0.ctl;
2130 }
2131 // The last RangeCheck is found to be redundant with a sequence of n (n >= 2) preceding RangeChecks.
2132 // If an array load is control dependent on the eliminated range check, the array load nodes (CastII and Load)
2133 // become control dependent on the last range check of the sequence, but they are really dependent on the entire
2134 // sequence of RangeChecks. If RangeCheck#n is later replaced by a dominating identical check, the array load
2135 // nodes must not float above the n-1 other RangeCheck in the sequence. We pin the array load nodes here to
2136 // guarantee it doesn't happen.
2137 //
2138 // RangeCheck#1 RangeCheck#1
2139 // | \ | \
2140 // | uncommon trap | uncommon trap
2141 // .. ..
2142 // RangeCheck#n -> RangeCheck#n
2143 // | \ | \
2144 // | uncommon trap CastII uncommon trap
2145 // RangeCheck Load
2146 // | \
2147 // CastII uncommon trap
2148 // Load
2149
2150 return dominated_by(prev_dom, igvn, true);
2151 }
2152 } else {
2153 prev_dom = search_identical(4, igvn);
2154
2155 if (prev_dom == nullptr) {
2156 return nullptr;
2157 }
2158 }
2159
2160 // Replace dominated IfNode
2161 return dominated_by(prev_dom, igvn, false);
2162 }
2163
2164 ParsePredicateNode::ParsePredicateNode(Node* control, Deoptimization::DeoptReason deopt_reason, PhaseGVN* gvn)
2165 : IfNode(control, gvn->intcon(1), PROB_MAX, COUNT_UNKNOWN),
2166 _deopt_reason(deopt_reason),
2167 _useless(false) {
2168 init_class_id(Class_ParsePredicate);
2169 gvn->C->add_parse_predicate(this);
2170 gvn->C->record_for_post_loop_opts_igvn(this);
2171 #ifdef ASSERT
2172 switch (deopt_reason) {
2173 case Deoptimization::Reason_predicate:
2174 case Deoptimization::Reason_profile_predicate:
2175 case Deoptimization::Reason_loop_limit_check:
2176 break;
2177 default:
2178 assert(false, "unsupported deoptimization reason for Parse Predicate");
2179 }
2180 #endif // ASSERT
2181 }
2182
2183 Node* ParsePredicateNode::uncommon_trap() const {
2184 ParsePredicateUncommonProj* uncommon_proj = proj_out(0)->as_IfFalse();
2185 Node* uct_region_or_call = uncommon_proj->unique_ctrl_out();
2186 assert(uct_region_or_call->is_Region() || uct_region_or_call->is_Call(), "must be a region or call uct");
2187 return uct_region_or_call;
2188 }
2189
2190 // Fold this node away once it becomes useless or at latest in post loop opts IGVN.
2191 const Type* ParsePredicateNode::Value(PhaseGVN* phase) const {
2192 if (phase->type(in(0)) == Type::TOP) {
2193 return Type::TOP;
2194 }
2195 if (_useless || phase->C->post_loop_opts_phase()) {
2196 return TypeTuple::IFTRUE;
2197 } else {
2198 return bottom_type();
2199 }
2200 }
2201
2202 #ifndef PRODUCT
2203 void ParsePredicateNode::dump_spec(outputStream* st) const {
2204 st->print(" #");
2205 switch (_deopt_reason) {
2206 case Deoptimization::DeoptReason::Reason_predicate:
2207 st->print("Loop ");
2208 break;
2209 case Deoptimization::DeoptReason::Reason_profile_predicate:
2210 st->print("Profiled_Loop ");
2211 break;
2212 case Deoptimization::DeoptReason::Reason_loop_limit_check:
2213 st->print("Loop_Limit_Check ");
2214 break;
2215 default:
2216 fatal("unknown kind");
2217 }
2218 if (_useless) {
2219 st->print("#useless ");
2220 }
2221 }
2222 #endif // NOT PRODUCT