1 /*
2 * Copyright (c) 1997, 2025, 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 "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "oops/objArrayKlass.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/inlinetypenode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/machnode.hpp"
38 #include "opto/movenode.hpp"
39 #include "opto/mulnode.hpp"
40 #include "opto/narrowptrnode.hpp"
41 #include "opto/phaseX.hpp"
42 #include "opto/regalloc.hpp"
43 #include "opto/regmask.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subnode.hpp"
46 #include "opto/vectornode.hpp"
47 #include "utilities/vmError.hpp"
48
49 // Portions of code courtesy of Clifford Click
50
51 // Optimization - Graph Style
52
53 //=============================================================================
54 //------------------------------Value------------------------------------------
55 // Compute the type of the RegionNode.
56 const Type* RegionNode::Value(PhaseGVN* phase) const {
57 for( uint i=1; i<req(); ++i ) { // For all paths in
58 Node *n = in(i); // Get Control source
59 if( !n ) continue; // Missing inputs are TOP
60 if( phase->type(n) == Type::CONTROL )
61 return Type::CONTROL;
62 }
63 return Type::TOP; // All paths dead? Then so are we
64 }
65
66 //------------------------------Identity---------------------------------------
67 // Check for Region being Identity.
68 Node* RegionNode::Identity(PhaseGVN* phase) {
69 // Cannot have Region be an identity, even if it has only 1 input.
70 // Phi users cannot have their Region input folded away for them,
71 // since they need to select the proper data input
72 return this;
73 }
74
75 //------------------------------merge_region-----------------------------------
76 // If a Region flows into a Region, merge into one big happy merge. This is
77 // hard to do if there is stuff that has to happen
78 static Node *merge_region(RegionNode *region, PhaseGVN *phase) {
79 if( region->Opcode() != Op_Region ) // Do not do to LoopNodes
80 return nullptr;
81 Node *progress = nullptr; // Progress flag
82 PhaseIterGVN *igvn = phase->is_IterGVN();
83
84 uint rreq = region->req();
85 for( uint i = 1; i < rreq; i++ ) {
86 Node *r = region->in(i);
87 if( r && r->Opcode() == Op_Region && // Found a region?
88 r->in(0) == r && // Not already collapsed?
89 r != region && // Avoid stupid situations
90 r->outcnt() == 2 ) { // Self user and 'region' user only?
91 assert(!r->as_Region()->has_phi(), "no phi users");
92 if( !progress ) { // No progress
93 if (region->has_phi()) {
94 return nullptr; // Only flatten if no Phi users
95 // igvn->hash_delete( phi );
96 }
97 igvn->hash_delete( region );
98 progress = region; // Making progress
99 }
100 igvn->hash_delete( r );
101
102 // Append inputs to 'r' onto 'region'
103 for( uint j = 1; j < r->req(); j++ ) {
104 // Move an input from 'r' to 'region'
105 region->add_req(r->in(j));
106 r->set_req(j, phase->C->top());
107 // Update phis of 'region'
108 //for( uint k = 0; k < max; k++ ) {
109 // Node *phi = region->out(k);
110 // if( phi->is_Phi() ) {
111 // phi->add_req(phi->in(i));
112 // }
113 //}
114
115 rreq++; // One more input to Region
116 } // Found a region to merge into Region
117 igvn->_worklist.push(r);
118 // Clobber pointer to the now dead 'r'
119 region->set_req(i, phase->C->top());
120 }
121 }
122
123 return progress;
124 }
125
126
127
128 //--------------------------------has_phi--------------------------------------
129 // Helper function: Return any PhiNode that uses this region or null
130 PhiNode* RegionNode::has_phi() const {
131 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
132 Node* phi = fast_out(i);
133 if (phi->is_Phi()) { // Check for Phi users
134 assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)");
135 return phi->as_Phi(); // this one is good enough
136 }
137 }
138
139 return nullptr;
140 }
141
142
143 //-----------------------------has_unique_phi----------------------------------
144 // Helper function: Return the only PhiNode that uses this region or null
145 PhiNode* RegionNode::has_unique_phi() const {
146 // Check that only one use is a Phi
147 PhiNode* only_phi = nullptr;
148 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
149 Node* phi = fast_out(i);
150 if (phi->is_Phi()) { // Check for Phi users
151 assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)");
152 if (only_phi == nullptr) {
153 only_phi = phi->as_Phi();
154 } else {
155 return nullptr; // multiple phis
156 }
157 }
158 }
159
160 return only_phi;
161 }
162
163
164 //------------------------------check_phi_clipping-----------------------------
165 // Helper function for RegionNode's identification of FP clipping
166 // Check inputs to the Phi
167 static bool check_phi_clipping( PhiNode *phi, ConNode * &min, uint &min_idx, ConNode * &max, uint &max_idx, Node * &val, uint &val_idx ) {
168 min = nullptr;
169 max = nullptr;
170 val = nullptr;
171 min_idx = 0;
172 max_idx = 0;
173 val_idx = 0;
174 uint phi_max = phi->req();
175 if( phi_max == 4 ) {
176 for( uint j = 1; j < phi_max; ++j ) {
177 Node *n = phi->in(j);
178 int opcode = n->Opcode();
179 switch( opcode ) {
180 case Op_ConI:
181 {
182 if( min == nullptr ) {
183 min = n->Opcode() == Op_ConI ? (ConNode*)n : nullptr;
184 min_idx = j;
185 } else {
186 max = n->Opcode() == Op_ConI ? (ConNode*)n : nullptr;
187 max_idx = j;
188 if( min->get_int() > max->get_int() ) {
189 // Swap min and max
190 ConNode *temp;
191 uint temp_idx;
192 temp = min; min = max; max = temp;
193 temp_idx = min_idx; min_idx = max_idx; max_idx = temp_idx;
194 }
195 }
196 }
197 break;
198 default:
199 {
200 val = n;
201 val_idx = j;
202 }
203 break;
204 }
205 }
206 }
207 return ( min && max && val && (min->get_int() <= 0) && (max->get_int() >=0) );
208 }
209
210
211 //------------------------------check_if_clipping------------------------------
212 // Helper function for RegionNode's identification of FP clipping
213 // Check that inputs to Region come from two IfNodes,
214 //
215 // If
216 // False True
217 // If |
218 // False True |
219 // | | |
220 // RegionNode_inputs
221 //
222 static bool check_if_clipping( const RegionNode *region, IfNode * &bot_if, IfNode * &top_if ) {
223 top_if = nullptr;
224 bot_if = nullptr;
225
226 // Check control structure above RegionNode for (if ( if ) )
227 Node *in1 = region->in(1);
228 Node *in2 = region->in(2);
229 Node *in3 = region->in(3);
230 // Check that all inputs are projections
231 if( in1->is_Proj() && in2->is_Proj() && in3->is_Proj() ) {
232 Node *in10 = in1->in(0);
233 Node *in20 = in2->in(0);
234 Node *in30 = in3->in(0);
235 // Check that #1 and #2 are ifTrue and ifFalse from same If
236 if( in10 != nullptr && in10->is_If() &&
237 in20 != nullptr && in20->is_If() &&
238 in30 != nullptr && in30->is_If() && in10 == in20 &&
239 (in1->Opcode() != in2->Opcode()) ) {
240 Node *in100 = in10->in(0);
241 Node *in1000 = (in100 != nullptr && in100->is_Proj()) ? in100->in(0) : nullptr;
242 // Check that control for in10 comes from other branch of IF from in3
243 if( in1000 != nullptr && in1000->is_If() &&
244 in30 == in1000 && (in3->Opcode() != in100->Opcode()) ) {
245 // Control pattern checks
246 top_if = (IfNode*)in1000;
247 bot_if = (IfNode*)in10;
248 }
249 }
250 }
251
252 return (top_if != nullptr);
253 }
254
255
256 //------------------------------check_convf2i_clipping-------------------------
257 // Helper function for RegionNode's identification of FP clipping
258 // Verify that the value input to the phi comes from "ConvF2I; LShift; RShift"
259 static bool check_convf2i_clipping( PhiNode *phi, uint idx, ConvF2INode * &convf2i, Node *min, Node *max) {
260 convf2i = nullptr;
261
262 // Check for the RShiftNode
263 Node *rshift = phi->in(idx);
264 assert( rshift, "Previous checks ensure phi input is present");
265 if( rshift->Opcode() != Op_RShiftI ) { return false; }
266
267 // Check for the LShiftNode
268 Node *lshift = rshift->in(1);
269 assert( lshift, "Previous checks ensure phi input is present");
270 if( lshift->Opcode() != Op_LShiftI ) { return false; }
271
272 // Check for the ConvF2INode
273 Node *conv = lshift->in(1);
274 if( conv->Opcode() != Op_ConvF2I ) { return false; }
275
276 // Check that shift amounts are only to get sign bits set after F2I
277 jint max_cutoff = max->get_int();
278 jint min_cutoff = min->get_int();
279 jint left_shift = lshift->in(2)->get_int();
280 jint right_shift = rshift->in(2)->get_int();
281 jint max_post_shift = nth_bit(BitsPerJavaInteger - left_shift - 1);
282 if( left_shift != right_shift ||
283 0 > left_shift || left_shift >= BitsPerJavaInteger ||
284 max_post_shift < max_cutoff ||
285 max_post_shift < -min_cutoff ) {
286 // Shifts are necessary but current transformation eliminates them
287 return false;
288 }
289
290 // OK to return the result of ConvF2I without shifting
291 convf2i = (ConvF2INode*)conv;
292 return true;
293 }
294
295
296 //------------------------------check_compare_clipping-------------------------
297 // Helper function for RegionNode's identification of FP clipping
298 static bool check_compare_clipping( bool less_than, IfNode *iff, ConNode *limit, Node * & input ) {
299 Node *i1 = iff->in(1);
300 if ( !i1->is_Bool() ) { return false; }
301 BoolNode *bool1 = i1->as_Bool();
302 if( less_than && bool1->_test._test != BoolTest::le ) { return false; }
303 else if( !less_than && bool1->_test._test != BoolTest::lt ) { return false; }
304 const Node *cmpF = bool1->in(1);
305 if( cmpF->Opcode() != Op_CmpF ) { return false; }
306 // Test that the float value being compared against
307 // is equivalent to the int value used as a limit
308 Node *nodef = cmpF->in(2);
309 if( nodef->Opcode() != Op_ConF ) { return false; }
310 jfloat conf = nodef->getf();
311 jint coni = limit->get_int();
312 if( ((int)conf) != coni ) { return false; }
313 input = cmpF->in(1);
314 return true;
315 }
316
317 //------------------------------is_unreachable_region--------------------------
318 // Check if the RegionNode is part of an unsafe loop and unreachable from root.
319 bool RegionNode::is_unreachable_region(const PhaseGVN* phase) {
320 Node* top = phase->C->top();
321 assert(req() == 2 || (req() == 3 && in(1) != nullptr && in(2) == top), "sanity check arguments");
322 if (_is_unreachable_region) {
323 // Return cached result from previous evaluation which should still be valid
324 assert(is_unreachable_from_root(phase), "walk the graph again and check if its indeed unreachable");
325 return true;
326 }
327
328 // First, cut the simple case of fallthrough region when NONE of
329 // region's phis references itself directly or through a data node.
330 if (is_possible_unsafe_loop(phase)) {
331 // If we have a possible unsafe loop, check if the region node is actually unreachable from root.
332 if (is_unreachable_from_root(phase)) {
333 _is_unreachable_region = true;
334 return true;
335 }
336 }
337 return false;
338 }
339
340 bool RegionNode::is_possible_unsafe_loop(const PhaseGVN* phase) const {
341 uint max = outcnt();
342 uint i;
343 for (i = 0; i < max; i++) {
344 Node* n = raw_out(i);
345 if (n != nullptr && n->is_Phi()) {
346 PhiNode* phi = n->as_Phi();
347 assert(phi->in(0) == this, "sanity check phi");
348 if (phi->outcnt() == 0) {
349 continue; // Safe case - no loops
350 }
351 if (phi->outcnt() == 1) {
352 Node* u = phi->raw_out(0);
353 // Skip if only one use is an other Phi or Call or Uncommon trap.
354 // It is safe to consider this case as fallthrough.
355 if (u != nullptr && (u->is_Phi() || u->is_CFG())) {
356 continue;
357 }
358 }
359 // Check when phi references itself directly or through an other node.
360 if (phi->as_Phi()->simple_data_loop_check(phi->in(1)) >= PhiNode::Unsafe) {
361 break; // Found possible unsafe data loop.
362 }
363 }
364 }
365 if (i >= max) {
366 return false; // An unsafe case was NOT found - don't need graph walk.
367 }
368 return true;
369 }
370
371 bool RegionNode::is_unreachable_from_root(const PhaseGVN* phase) const {
372 ResourceMark rm;
373 Node_List nstack;
374 VectorSet visited;
375
376 // Mark all control nodes reachable from root outputs
377 Node* n = (Node*)phase->C->root();
378 nstack.push(n);
379 visited.set(n->_idx);
380 while (nstack.size() != 0) {
381 n = nstack.pop();
382 uint max = n->outcnt();
383 for (uint i = 0; i < max; i++) {
384 Node* m = n->raw_out(i);
385 if (m != nullptr && m->is_CFG()) {
386 if (m == this) {
387 return false; // We reached the Region node - it is not dead.
388 }
389 if (!visited.test_set(m->_idx))
390 nstack.push(m);
391 }
392 }
393 }
394 return true; // The Region node is unreachable - it is dead.
395 }
396
397 #ifdef ASSERT
398 // Is this region in an infinite subgraph?
399 // (no path to root except through false NeverBranch exit)
400 bool RegionNode::is_in_infinite_subgraph() {
401 ResourceMark rm;
402 Unique_Node_List worklist;
403 worklist.push(this);
404 return RegionNode::are_all_nodes_in_infinite_subgraph(worklist);
405 }
406
407 // Are all nodes in worklist in infinite subgraph?
408 // (no path to root except through false NeverBranch exit)
409 // worklist is directly used for the traversal
410 bool RegionNode::are_all_nodes_in_infinite_subgraph(Unique_Node_List& worklist) {
411 // BFS traversal down the CFG, except through NeverBranch exits
412 for (uint i = 0; i < worklist.size(); ++i) {
413 Node* n = worklist.at(i);
414 assert(n->is_CFG(), "only traverse CFG");
415 if (n->is_Root()) {
416 // Found root -> there was an exit!
417 return false;
418 } else if (n->is_NeverBranch()) {
419 // Only follow the loop-internal projection, not the NeverBranch exit
420 ProjNode* proj = n->as_NeverBranch()->proj_out_or_null(0);
421 assert(proj != nullptr, "must find loop-internal projection of NeverBranch");
422 worklist.push(proj);
423 } else {
424 // Traverse all CFG outputs
425 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
426 Node* use = n->fast_out(i);
427 if (use->is_CFG()) {
428 worklist.push(use);
429 }
430 }
431 }
432 }
433 // No exit found for any loop -> all are infinite
434 return true;
435 }
436 #endif //ASSERT
437
438 void RegionNode::set_loop_status(RegionNode::LoopStatus status) {
439 assert(loop_status() == RegionNode::LoopStatus::NeverIrreducibleEntry, "why set our status again?");
440 assert(status != RegionNode::LoopStatus::MaybeIrreducibleEntry || !is_Loop(), "LoopNode is never irreducible entry.");
441 _loop_status = status;
442 }
443
444 // A Region can only be an irreducible entry if:
445 // - It is marked as "maybe irreducible entry". Any other loop status would guarantee
446 // that it is never an irreducible loop entry.
447 // - And it is not a LoopNode, those are guaranteed to be reducible loop entries.
448 bool RegionNode::can_be_irreducible_entry() const {
449 return loop_status() == RegionNode::LoopStatus::MaybeIrreducibleEntry &&
450 !is_Loop();
451 }
452
453 void RegionNode::try_clean_mem_phis(PhaseIterGVN* igvn) {
454 // Incremental inlining + PhaseStringOpts sometimes produce:
455 //
456 // cmpP with 1 top input
457 // |
458 // If
459 // / \
460 // IfFalse IfTrue /- Some Node
461 // \ / / /
462 // Region / /-MergeMem
463 // \---Phi
464 //
465 //
466 // It's expected by PhaseStringOpts that the Region goes away and is
467 // replaced by If's control input but because there's still a Phi,
468 // the Region stays in the graph. The top input from the cmpP is
469 // propagated forward and a subgraph that is useful goes away. The
470 // code in PhiNode::try_clean_memory_phi() replaces the Phi with the
471 // MergeMem in order to remove the Region if its last phi dies.
472
473 if (!is_diamond()) {
474 return;
475 }
476
477 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
478 Node* phi = fast_out(i);
479 if (phi->is_Phi() && phi->as_Phi()->try_clean_memory_phi(igvn)) {
480 --i;
481 --imax;
482 }
483 }
484 }
485
486 // Does this region merge a simple diamond formed by a proper IfNode?
487 //
488 // Cmp
489 // /
490 // ctrl Bool
491 // \ /
492 // IfNode
493 // / \
494 // IfFalse IfTrue
495 // \ /
496 // Region
497 bool RegionNode::is_diamond() const {
498 if (req() != 3) {
499 return false;
500 }
501
502 Node* left_path = in(1);
503 Node* right_path = in(2);
504 if (left_path == nullptr || right_path == nullptr) {
505 return false;
506 }
507 Node* diamond_if = left_path->in(0);
508 if (diamond_if == nullptr || !diamond_if->is_If() || diamond_if != right_path->in(0)) {
509 // Not an IfNode merging a diamond or TOP.
510 return false;
511 }
512
513 // Check for a proper bool/cmp
514 const Node* bol = diamond_if->in(1);
515 if (!bol->is_Bool()) {
516 return false;
517 }
518 const Node* cmp = bol->in(1);
519 if (!cmp->is_Cmp()) {
520 return false;
521 }
522 return true;
523 }
524
525 //------------------------------Ideal------------------------------------------
526 // Return a node which is more "ideal" than the current node. Must preserve
527 // the CFG, but we can still strip out dead paths.
528 Node *RegionNode::Ideal(PhaseGVN *phase, bool can_reshape) {
529 if( !can_reshape && !in(0) ) return nullptr; // Already degraded to a Copy
530 assert(!in(0) || !in(0)->is_Root(), "not a specially hidden merge");
531
532 // Check for RegionNode with no Phi users and both inputs come from either
533 // arm of the same IF. If found, then the control-flow split is useless.
534 bool has_phis = false;
535 if (can_reshape) { // Need DU info to check for Phi users
536 try_clean_mem_phis(phase->is_IterGVN());
537 has_phis = (has_phi() != nullptr); // Cache result
538
539 if (!has_phis) { // No Phi users? Nothing merging?
540 for (uint i = 1; i < req()-1; i++) {
541 Node *if1 = in(i);
542 if( !if1 ) continue;
543 Node *iff = if1->in(0);
544 if( !iff || !iff->is_If() ) continue;
545 for( uint j=i+1; j<req(); j++ ) {
546 if( in(j) && in(j)->in(0) == iff &&
547 if1->Opcode() != in(j)->Opcode() ) {
548 // Add the IF Projections to the worklist. They (and the IF itself)
549 // will be eliminated if dead.
550 phase->is_IterGVN()->add_users_to_worklist(iff);
551 set_req(i, iff->in(0));// Skip around the useless IF diamond
552 set_req(j, nullptr);
553 return this; // Record progress
554 }
555 }
556 }
557 }
558 }
559
560 // Remove TOP or null input paths. If only 1 input path remains, this Region
561 // degrades to a copy.
562 bool add_to_worklist = true;
563 bool modified = false;
564 int cnt = 0; // Count of values merging
565 DEBUG_ONLY( int cnt_orig = req(); ) // Save original inputs count
566 DEBUG_ONLY( uint outcnt_orig = outcnt(); )
567 int del_it = 0; // The last input path we delete
568 bool found_top = false; // irreducible loops need to check reachability if we find TOP
569 // For all inputs...
570 for( uint i=1; i<req(); ++i ){// For all paths in
571 Node *n = in(i); // Get the input
572 if( n != nullptr ) {
573 // Remove useless control copy inputs
574 if( n->is_Region() && n->as_Region()->is_copy() ) {
575 set_req(i, n->nonnull_req());
576 modified = true;
577 i--;
578 continue;
579 }
580 if( n->is_Proj() ) { // Remove useless rethrows
581 Node *call = n->in(0);
582 if (call->is_Call() && call->as_Call()->entry_point() == OptoRuntime::rethrow_stub()) {
583 set_req(i, call->in(0));
584 modified = true;
585 i--;
586 continue;
587 }
588 }
589 if( phase->type(n) == Type::TOP ) {
590 set_req_X(i, nullptr, phase); // Ignore TOP inputs
591 modified = true;
592 found_top = true;
593 i--;
594 continue;
595 }
596 cnt++; // One more value merging
597 } else if (can_reshape) { // Else found dead path with DU info
598 PhaseIterGVN *igvn = phase->is_IterGVN();
599 del_req(i); // Yank path from self
600 del_it = i;
601
602 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
603 Node* use = fast_out(j);
604
605 if (use->req() != req() && use->is_Phi()) {
606 assert(use->in(0) == this, "unexpected control input");
607 igvn->hash_delete(use); // Yank from hash before hacking edges
608 use->set_req_X(i, nullptr, igvn);// Correct DU info
609 use->del_req(i); // Yank path from Phis
610 }
611 }
612
613 if (add_to_worklist) {
614 igvn->add_users_to_worklist(this);
615 add_to_worklist = false;
616 }
617
618 i--;
619 }
620 }
621
622 assert(outcnt() == outcnt_orig, "not expect to remove any use");
623
624 if (can_reshape && found_top && loop_status() == RegionNode::LoopStatus::MaybeIrreducibleEntry) {
625 // Is it a dead irreducible loop?
626 // If an irreducible loop loses one of the multiple entries
627 // that went into the loop head, or any secondary entries,
628 // we need to verify if the irreducible loop is still reachable,
629 // as the special logic in is_unreachable_region only works
630 // for reducible loops.
631 if (is_unreachable_from_root(phase)) {
632 // The irreducible loop is dead - must remove it
633 PhaseIterGVN* igvn = phase->is_IterGVN();
634 remove_unreachable_subgraph(igvn);
635 return nullptr;
636 }
637 } else if (can_reshape && cnt == 1) {
638 // Is it dead loop?
639 // If it is LoopNopde it had 2 (+1 itself) inputs and
640 // one of them was cut. The loop is dead if it was EntryContol.
641 // Loop node may have only one input because entry path
642 // is removed in PhaseIdealLoop::Dominators().
643 assert(!this->is_Loop() || cnt_orig <= 3, "Loop node should have 3 or less inputs");
644 if ((this->is_Loop() && (del_it == LoopNode::EntryControl ||
645 (del_it == 0 && is_unreachable_region(phase)))) ||
646 (!this->is_Loop() && has_phis && is_unreachable_region(phase))) {
647 PhaseIterGVN* igvn = phase->is_IterGVN();
648 remove_unreachable_subgraph(igvn);
649 return nullptr;
650 }
651 }
652
653 if( cnt <= 1 ) { // Only 1 path in?
654 set_req(0, nullptr); // Null control input for region copy
655 if( cnt == 0 && !can_reshape) { // Parse phase - leave the node as it is.
656 // No inputs or all inputs are null.
657 return nullptr;
658 } else if (can_reshape) { // Optimization phase - remove the node
659 PhaseIterGVN *igvn = phase->is_IterGVN();
660 // Strip mined (inner) loop is going away, remove outer loop.
661 if (is_CountedLoop() &&
662 as_Loop()->is_strip_mined()) {
663 Node* outer_sfpt = as_CountedLoop()->outer_safepoint();
664 Node* outer_out = as_CountedLoop()->outer_loop_exit();
665 if (outer_sfpt != nullptr && outer_out != nullptr) {
666 Node* in = outer_sfpt->in(0);
667 igvn->replace_node(outer_out, in);
668 LoopNode* outer = as_CountedLoop()->outer_loop();
669 igvn->replace_input_of(outer, LoopNode::LoopBackControl, igvn->C->top());
670 }
671 }
672 if (is_CountedLoop()) {
673 Node* opaq = as_CountedLoop()->is_canonical_loop_entry();
674 if (opaq != nullptr) {
675 // This is not a loop anymore. No need to keep the Opaque1 node on the test that guards the loop as it won't be
676 // subject to further loop opts.
677 assert(opaq->Opcode() == Op_OpaqueZeroTripGuard, "");
678 igvn->replace_node(opaq, opaq->in(1));
679 }
680 }
681 Node *parent_ctrl;
682 if( cnt == 0 ) {
683 assert( req() == 1, "no inputs expected" );
684 // During IGVN phase such region will be subsumed by TOP node
685 // so region's phis will have TOP as control node.
686 // Kill phis here to avoid it.
687 // Also set other user's input to top.
688 parent_ctrl = phase->C->top();
689 } else {
690 // The fallthrough case since we already checked dead loops above.
691 parent_ctrl = in(1);
692 assert(parent_ctrl != nullptr, "Region is a copy of some non-null control");
693 assert(parent_ctrl != this, "Close dead loop");
694 }
695 if (add_to_worklist) {
696 igvn->add_users_to_worklist(this); // Check for further allowed opts
697 }
698 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) {
699 Node* n = last_out(i);
700 igvn->hash_delete(n); // Remove from worklist before modifying edges
701 if (n->outcnt() == 0) {
702 int uses_found = n->replace_edge(this, phase->C->top(), igvn);
703 if (uses_found > 1) { // (--i) done at the end of the loop.
704 i -= (uses_found - 1);
705 }
706 continue;
707 }
708 if( n->is_Phi() ) { // Collapse all Phis
709 // Eagerly replace phis to avoid regionless phis.
710 Node* in;
711 if( cnt == 0 ) {
712 assert( n->req() == 1, "No data inputs expected" );
713 in = parent_ctrl; // replaced by top
714 } else {
715 assert( n->req() == 2 && n->in(1) != nullptr, "Only one data input expected" );
716 in = n->in(1); // replaced by unique input
717 if( n->as_Phi()->is_unsafe_data_reference(in) )
718 in = phase->C->top(); // replaced by top
719 }
720 igvn->replace_node(n, in);
721 }
722 else if( n->is_Region() ) { // Update all incoming edges
723 assert(n != this, "Must be removed from DefUse edges");
724 int uses_found = n->replace_edge(this, parent_ctrl, igvn);
725 if (uses_found > 1) { // (--i) done at the end of the loop.
726 i -= (uses_found - 1);
727 }
728 }
729 else {
730 assert(n->in(0) == this, "Expect RegionNode to be control parent");
731 n->set_req(0, parent_ctrl);
732 }
733 #ifdef ASSERT
734 for( uint k=0; k < n->req(); k++ ) {
735 assert(n->in(k) != this, "All uses of RegionNode should be gone");
736 }
737 #endif
738 }
739 // Remove the RegionNode itself from DefUse info
740 igvn->remove_dead_node(this);
741 return nullptr;
742 }
743 return this; // Record progress
744 }
745
746
747 // If a Region flows into a Region, merge into one big happy merge.
748 if (can_reshape) {
749 Node *m = merge_region(this, phase);
750 if (m != nullptr) return m;
751 }
752
753 // Check if this region is the root of a clipping idiom on floats
754 if( ConvertFloat2IntClipping && can_reshape && req() == 4 ) {
755 // Check that only one use is a Phi and that it simplifies to two constants +
756 PhiNode* phi = has_unique_phi();
757 if (phi != nullptr) { // One Phi user
758 // Check inputs to the Phi
759 ConNode *min;
760 ConNode *max;
761 Node *val;
762 uint min_idx;
763 uint max_idx;
764 uint val_idx;
765 if( check_phi_clipping( phi, min, min_idx, max, max_idx, val, val_idx ) ) {
766 IfNode *top_if;
767 IfNode *bot_if;
768 if( check_if_clipping( this, bot_if, top_if ) ) {
769 // Control pattern checks, now verify compares
770 Node *top_in = nullptr; // value being compared against
771 Node *bot_in = nullptr;
772 if( check_compare_clipping( true, bot_if, min, bot_in ) &&
773 check_compare_clipping( false, top_if, max, top_in ) ) {
774 if( bot_in == top_in ) {
775 PhaseIterGVN *gvn = phase->is_IterGVN();
776 assert( gvn != nullptr, "Only had DefUse info in IterGVN");
777 // Only remaining check is that bot_in == top_in == (Phi's val + mods)
778
779 // Check for the ConvF2INode
780 ConvF2INode *convf2i;
781 if( check_convf2i_clipping( phi, val_idx, convf2i, min, max ) &&
782 convf2i->in(1) == bot_in ) {
783 // Matched pattern, including LShiftI; RShiftI, replace with integer compares
784 // max test
785 Node *cmp = gvn->register_new_node_with_optimizer(new CmpINode( convf2i, min ));
786 Node *boo = gvn->register_new_node_with_optimizer(new BoolNode( cmp, BoolTest::lt ));
787 IfNode *iff = (IfNode*)gvn->register_new_node_with_optimizer(new IfNode( top_if->in(0), boo, PROB_UNLIKELY_MAG(5), top_if->_fcnt ));
788 Node *if_min= gvn->register_new_node_with_optimizer(new IfTrueNode (iff));
789 Node *ifF = gvn->register_new_node_with_optimizer(new IfFalseNode(iff));
790 // min test
791 cmp = gvn->register_new_node_with_optimizer(new CmpINode( convf2i, max ));
792 boo = gvn->register_new_node_with_optimizer(new BoolNode( cmp, BoolTest::gt ));
793 iff = (IfNode*)gvn->register_new_node_with_optimizer(new IfNode( ifF, boo, PROB_UNLIKELY_MAG(5), bot_if->_fcnt ));
794 Node *if_max= gvn->register_new_node_with_optimizer(new IfTrueNode (iff));
795 ifF = gvn->register_new_node_with_optimizer(new IfFalseNode(iff));
796 // update input edges to region node
797 set_req_X( min_idx, if_min, gvn );
798 set_req_X( max_idx, if_max, gvn );
799 set_req_X( val_idx, ifF, gvn );
800 // remove unnecessary 'LShiftI; RShiftI' idiom
801 gvn->hash_delete(phi);
802 phi->set_req_X( val_idx, convf2i, gvn );
803 gvn->hash_find_insert(phi);
804 // Return transformed region node
805 return this;
806 }
807 }
808 }
809 }
810 }
811 }
812 }
813
814 if (can_reshape) {
815 modified |= optimize_trichotomy(phase->is_IterGVN());
816 }
817
818 return modified ? this : nullptr;
819 }
820
821 //--------------------------remove_unreachable_subgraph----------------------
822 // This region and therefore all nodes on the input control path(s) are unreachable
823 // from root. To avoid incomplete removal of unreachable subgraphs, walk up the CFG
824 // and aggressively replace all nodes by top.
825 // If a control node "def" with a single control output "use" has its single output
826 // "use" replaced with top, then "use" removes itself. This has the consequence that
827 // when we visit "use", it already has all inputs removed. They are lost and we cannot
828 // traverse them. This is why we fist find all unreachable nodes, and then remove
829 // them in a second step.
830 void RegionNode::remove_unreachable_subgraph(PhaseIterGVN* igvn) {
831 Node* top = igvn->C->top();
832 ResourceMark rm;
833 Unique_Node_List unreachable; // visit each only once
834 unreachable.push(this);
835 // Recursively find all control inputs.
836 for (uint i = 0; i < unreachable.size(); i++) {
837 Node* n = unreachable.at(i);
838 for (uint i = 0; i < n->req(); ++i) {
839 Node* m = n->in(i);
840 assert(m == nullptr || !m->is_Root(), "Should be unreachable from root");
841 if (m != nullptr && m->is_CFG()) {
842 unreachable.push(m);
843 }
844 }
845 }
846 // Remove all unreachable nodes.
847 for (uint i = 0; i < unreachable.size(); i++) {
848 Node* n = unreachable.at(i);
849 if (n->is_Region()) {
850 // Eagerly replace phis with top to avoid regionless phis.
851 n->set_req(0, nullptr);
852 bool progress = true;
853 uint max = n->outcnt();
854 DUIterator j;
855 while (progress) {
856 progress = false;
857 for (j = n->outs(); n->has_out(j); j++) {
858 Node* u = n->out(j);
859 if (u->is_Phi()) {
860 igvn->replace_node(u, top);
861 if (max != n->outcnt()) {
862 progress = true;
863 j = n->refresh_out_pos(j);
864 max = n->outcnt();
865 }
866 }
867 }
868 }
869 }
870 igvn->replace_node(n, top);
871 }
872 }
873
874 //------------------------------optimize_trichotomy--------------------------
875 // Optimize nested comparisons of the following kind:
876 //
877 // int compare(int a, int b) {
878 // return (a < b) ? -1 : (a == b) ? 0 : 1;
879 // }
880 //
881 // Shape 1:
882 // if (compare(a, b) == 1) { ... } -> if (a > b) { ... }
883 //
884 // Shape 2:
885 // if (compare(a, b) == 0) { ... } -> if (a == b) { ... }
886 //
887 // Above code leads to the following IR shapes where both Ifs compare the
888 // same value and two out of three region inputs idx1 and idx2 map to
889 // the same value and control flow.
890 //
891 // (1) If (2) If
892 // / \ / \
893 // Proj Proj Proj Proj
894 // | \ | \
895 // | If | If If
896 // | / \ | / \ / \
897 // | Proj Proj | Proj Proj ==> Proj Proj
898 // | / / \ | / | /
899 // Region / \ | / | /
900 // \ / \ | / | /
901 // Region Region Region
902 //
903 // The method returns true if 'this' is modified and false otherwise.
904 bool RegionNode::optimize_trichotomy(PhaseIterGVN* igvn) {
905 int idx1 = 1, idx2 = 2;
906 Node* region = nullptr;
907 if (req() == 3 && in(1) != nullptr && in(2) != nullptr) {
908 // Shape 1: Check if one of the inputs is a region that merges two control
909 // inputs and has no other users (especially no Phi users).
910 region = in(1)->isa_Region() ? in(1) : in(2)->isa_Region();
911 if (region == nullptr || region->outcnt() != 2 || region->req() != 3) {
912 return false; // No suitable region input found
913 }
914 } else if (req() == 4) {
915 // Shape 2: Check if two control inputs map to the same value of the unique phi
916 // user and treat these as if they would come from another region (shape (1)).
917 PhiNode* phi = has_unique_phi();
918 if (phi == nullptr) {
919 return false; // No unique phi user
920 }
921 if (phi->in(idx1) != phi->in(idx2)) {
922 idx2 = 3;
923 if (phi->in(idx1) != phi->in(idx2)) {
924 idx1 = 2;
925 if (phi->in(idx1) != phi->in(idx2)) {
926 return false; // No equal phi inputs found
927 }
928 }
929 }
930 assert(phi->in(idx1) == phi->in(idx2), "must be"); // Region is merging same value
931 region = this;
932 }
933 if (region == nullptr || region->in(idx1) == nullptr || region->in(idx2) == nullptr) {
934 return false; // Region does not merge two control inputs
935 }
936 // At this point we know that region->in(idx1) and region->(idx2) map to the same
937 // value and control flow. Now search for ifs that feed into these region inputs.
938 ProjNode* proj1 = region->in(idx1)->isa_Proj();
939 ProjNode* proj2 = region->in(idx2)->isa_Proj();
940 if (proj1 == nullptr || proj1->outcnt() != 1 ||
941 proj2 == nullptr || proj2->outcnt() != 1) {
942 return false; // No projection inputs with region as unique user found
943 }
944 assert(proj1 != proj2, "should be different projections");
945 IfNode* iff1 = proj1->in(0)->isa_If();
946 IfNode* iff2 = proj2->in(0)->isa_If();
947 if (iff1 == nullptr || iff1->outcnt() != 2 ||
948 iff2 == nullptr || iff2->outcnt() != 2) {
949 return false; // No ifs found
950 }
951 if (iff1 == iff2) {
952 igvn->add_users_to_worklist(iff1); // Make sure dead if is eliminated
953 igvn->replace_input_of(region, idx1, iff1->in(0));
954 igvn->replace_input_of(region, idx2, igvn->C->top());
955 return (region == this); // Remove useless if (both projections map to the same control/value)
956 }
957 BoolNode* bol1 = iff1->in(1)->isa_Bool();
958 BoolNode* bol2 = iff2->in(1)->isa_Bool();
959 if (bol1 == nullptr || bol2 == nullptr) {
960 return false; // No bool inputs found
961 }
962 Node* cmp1 = bol1->in(1);
963 Node* cmp2 = bol2->in(1);
964 bool commute = false;
965 if (!cmp1->is_Cmp() || !cmp2->is_Cmp()) {
966 return false; // No comparison
967 } else if (cmp1->Opcode() == Op_CmpF || cmp1->Opcode() == Op_CmpD ||
968 cmp2->Opcode() == Op_CmpF || cmp2->Opcode() == Op_CmpD ||
969 cmp1->Opcode() == Op_CmpP || cmp1->Opcode() == Op_CmpN ||
970 cmp2->Opcode() == Op_CmpP || cmp2->Opcode() == Op_CmpN ||
971 cmp1->is_SubTypeCheck() || cmp2->is_SubTypeCheck() ||
972 cmp1->is_FlatArrayCheck() || cmp2->is_FlatArrayCheck()) {
973 // Floats and pointers don't exactly obey trichotomy. To be on the safe side, don't transform their tests.
974 // SubTypeCheck is not commutative
975 return false;
976 } else if (cmp1 != cmp2) {
977 if (cmp1->in(1) == cmp2->in(2) &&
978 cmp1->in(2) == cmp2->in(1)) {
979 commute = true; // Same but swapped inputs, commute the test
980 } else {
981 return false; // Ifs are not comparing the same values
982 }
983 }
984 proj1 = proj1->other_if_proj();
985 proj2 = proj2->other_if_proj();
986 if (!((proj1->unique_ctrl_out_or_null() == iff2 &&
987 proj2->unique_ctrl_out_or_null() == this) ||
988 (proj2->unique_ctrl_out_or_null() == iff1 &&
989 proj1->unique_ctrl_out_or_null() == this))) {
990 return false; // Ifs are not connected through other projs
991 }
992 // Found 'iff -> proj -> iff -> proj -> this' shape where all other projs are merged
993 // through 'region' and map to the same value. Merge the boolean tests and replace
994 // the ifs by a single comparison.
995 BoolTest test1 = (proj1->_con == 1) ? bol1->_test : bol1->_test.negate();
996 BoolTest test2 = (proj2->_con == 1) ? bol2->_test : bol2->_test.negate();
997 test1 = commute ? test1.commute() : test1;
998 // After possibly commuting test1, if we can merge test1 & test2, then proj2/iff2/bol2 are the nodes to refine.
999 BoolTest::mask res = test1.merge(test2);
1000 if (res == BoolTest::illegal) {
1001 return false; // Unable to merge tests
1002 }
1003 // Adjust iff1 to always pass (only iff2 will remain)
1004 igvn->replace_input_of(iff1, 1, igvn->intcon(proj1->_con));
1005 if (res == BoolTest::never) {
1006 // Merged test is always false, adjust iff2 to always fail
1007 igvn->replace_input_of(iff2, 1, igvn->intcon(1 - proj2->_con));
1008 } else {
1009 // Replace bool input of iff2 with merged test
1010 BoolNode* new_bol = new BoolNode(bol2->in(1), res);
1011 igvn->replace_input_of(iff2, 1, igvn->transform((proj2->_con == 1) ? new_bol : new_bol->negate(igvn)));
1012 if (new_bol->outcnt() == 0) {
1013 igvn->remove_dead_node(new_bol);
1014 }
1015 }
1016 return false;
1017 }
1018
1019 const RegMask &RegionNode::out_RegMask() const {
1020 return RegMask::Empty;
1021 }
1022
1023 #ifndef PRODUCT
1024 void RegionNode::dump_spec(outputStream* st) const {
1025 Node::dump_spec(st);
1026 switch (loop_status()) {
1027 case RegionNode::LoopStatus::MaybeIrreducibleEntry:
1028 st->print("#irreducible ");
1029 break;
1030 case RegionNode::LoopStatus::Reducible:
1031 st->print("#reducible ");
1032 break;
1033 case RegionNode::LoopStatus::NeverIrreducibleEntry:
1034 break; // nothing
1035 }
1036 }
1037 #endif
1038
1039 // Find the one non-null required input. RegionNode only
1040 Node *Node::nonnull_req() const {
1041 assert( is_Region(), "" );
1042 for( uint i = 1; i < _cnt; i++ )
1043 if( in(i) )
1044 return in(i);
1045 ShouldNotReachHere();
1046 return nullptr;
1047 }
1048
1049
1050 //=============================================================================
1051 // note that these functions assume that the _adr_type field is flat
1052 uint PhiNode::hash() const {
1053 const Type* at = _adr_type;
1054 return TypeNode::hash() + (at ? at->hash() : 0);
1055 }
1056 bool PhiNode::cmp( const Node &n ) const {
1057 return TypeNode::cmp(n) && _adr_type == ((PhiNode&)n)._adr_type;
1058 }
1059 static inline
1060 const TypePtr* flatten_phi_adr_type(const TypePtr* at) {
1061 if (at == nullptr || at == TypePtr::BOTTOM) return at;
1062 return Compile::current()->alias_type(at)->adr_type();
1063 }
1064
1065 //----------------------------make---------------------------------------------
1066 // create a new phi with edges matching r and set (initially) to x
1067 PhiNode* PhiNode::make(Node* r, Node* x, const Type *t, const TypePtr* at) {
1068 uint preds = r->req(); // Number of predecessor paths
1069 assert(t != Type::MEMORY || at == flatten_phi_adr_type(at) || (flatten_phi_adr_type(at) == TypeAryPtr::INLINES && Compile::current()->flat_accesses_share_alias()), "flatten at");
1070 PhiNode* p = new PhiNode(r, t, at);
1071 for (uint j = 1; j < preds; j++) {
1072 // Fill in all inputs, except those which the region does not yet have
1073 if (r->in(j) != nullptr)
1074 p->init_req(j, x);
1075 }
1076 return p;
1077 }
1078 PhiNode* PhiNode::make(Node* r, Node* x) {
1079 const Type* t = x->bottom_type();
1080 const TypePtr* at = nullptr;
1081 if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type());
1082 return make(r, x, t, at);
1083 }
1084 PhiNode* PhiNode::make_blank(Node* r, Node* x) {
1085 const Type* t = x->bottom_type();
1086 const TypePtr* at = nullptr;
1087 if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type());
1088 return new PhiNode(r, t, at);
1089 }
1090
1091
1092 //------------------------slice_memory-----------------------------------------
1093 // create a new phi with narrowed memory type
1094 PhiNode* PhiNode::slice_memory(const TypePtr* adr_type) const {
1095 PhiNode* mem = (PhiNode*) clone();
1096 *(const TypePtr**)&mem->_adr_type = adr_type;
1097 // convert self-loops, or else we get a bad graph
1098 for (uint i = 1; i < req(); i++) {
1099 if ((const Node*)in(i) == this) mem->set_req(i, mem);
1100 }
1101 mem->verify_adr_type();
1102 return mem;
1103 }
1104
1105 //------------------------split_out_instance-----------------------------------
1106 // Split out an instance type from a bottom phi.
1107 PhiNode* PhiNode::split_out_instance(const TypePtr* at, PhaseIterGVN *igvn) const {
1108 const TypeOopPtr *t_oop = at->isa_oopptr();
1109 assert(t_oop != nullptr && t_oop->is_known_instance(), "expecting instance oopptr");
1110
1111 // Check if an appropriate node already exists.
1112 Node *region = in(0);
1113 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
1114 Node* use = region->fast_out(k);
1115 if( use->is_Phi()) {
1116 PhiNode *phi2 = use->as_Phi();
1117 if (phi2->type() == Type::MEMORY && phi2->adr_type() == at) {
1118 return phi2;
1119 }
1120 }
1121 }
1122 Compile *C = igvn->C;
1123 ResourceMark rm;
1124 Node_Array node_map;
1125 Node_Stack stack(C->live_nodes() >> 4);
1126 PhiNode *nphi = slice_memory(at);
1127 igvn->register_new_node_with_optimizer( nphi );
1128 node_map.map(_idx, nphi);
1129 stack.push((Node *)this, 1);
1130 while(!stack.is_empty()) {
1131 PhiNode *ophi = stack.node()->as_Phi();
1132 uint i = stack.index();
1133 assert(i >= 1, "not control edge");
1134 stack.pop();
1135 nphi = node_map[ophi->_idx]->as_Phi();
1136 for (; i < ophi->req(); i++) {
1137 Node *in = ophi->in(i);
1138 if (in == nullptr || igvn->type(in) == Type::TOP)
1139 continue;
1140 Node *opt = MemNode::optimize_simple_memory_chain(in, t_oop, nullptr, igvn);
1141 PhiNode *optphi = opt->is_Phi() ? opt->as_Phi() : nullptr;
1142 if (optphi != nullptr && optphi->adr_type() == TypePtr::BOTTOM) {
1143 opt = node_map[optphi->_idx];
1144 if (opt == nullptr) {
1145 stack.push(ophi, i);
1146 nphi = optphi->slice_memory(at);
1147 igvn->register_new_node_with_optimizer( nphi );
1148 node_map.map(optphi->_idx, nphi);
1149 ophi = optphi;
1150 i = 0; // will get incremented at top of loop
1151 continue;
1152 }
1153 }
1154 nphi->set_req(i, opt);
1155 }
1156 }
1157 return nphi;
1158 }
1159
1160 //------------------------verify_adr_type--------------------------------------
1161 #ifdef ASSERT
1162 void PhiNode::verify_adr_type(VectorSet& visited, const TypePtr* at) const {
1163 if (visited.test_set(_idx)) return; //already visited
1164
1165 // recheck constructor invariants:
1166 verify_adr_type(false);
1167
1168 // recheck local phi/phi consistency:
1169 assert(_adr_type == at || _adr_type == TypePtr::BOTTOM,
1170 "adr_type must be consistent across phi nest");
1171
1172 // walk around
1173 for (uint i = 1; i < req(); i++) {
1174 Node* n = in(i);
1175 if (n == nullptr) continue;
1176 const Node* np = in(i);
1177 if (np->is_Phi()) {
1178 np->as_Phi()->verify_adr_type(visited, at);
1179 } else if (n->bottom_type() == Type::TOP
1180 || (n->is_Mem() && n->in(MemNode::Address)->bottom_type() == Type::TOP)) {
1181 // ignore top inputs
1182 } else {
1183 const TypePtr* nat = flatten_phi_adr_type(n->adr_type());
1184 // recheck phi/non-phi consistency at leaves:
1185 assert((nat != nullptr) == (at != nullptr), "");
1186 assert(nat == at || nat == TypePtr::BOTTOM,
1187 "adr_type must be consistent at leaves of phi nest");
1188 }
1189 }
1190 }
1191
1192 // Verify a whole nest of phis rooted at this one.
1193 void PhiNode::verify_adr_type(bool recursive) const {
1194 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
1195 if (Node::in_dump()) return; // muzzle asserts when printing
1196
1197 assert((_type == Type::MEMORY) == (_adr_type != nullptr), "adr_type for memory phis only");
1198 // Flat array element shouldn't get their own memory slice until flat_accesses_share_alias is cleared.
1199 // It could be the graph has no loads/stores and flat_accesses_share_alias is never cleared. EA could still
1200 // creates per element Phis but that wouldn't be a problem as there are no memory accesses for that array.
1201 assert(_adr_type == nullptr || _adr_type->isa_aryptr() == nullptr ||
1202 _adr_type->is_aryptr()->is_known_instance() ||
1203 !_adr_type->is_aryptr()->is_flat() ||
1204 !Compile::current()->flat_accesses_share_alias() ||
1205 _adr_type == TypeAryPtr::INLINES, "flat array element shouldn't get its own slice yet");
1206
1207 if (!VerifyAliases) return; // verify thoroughly only if requested
1208
1209 assert(_adr_type == flatten_phi_adr_type(_adr_type),
1210 "Phi::adr_type must be pre-normalized");
1211
1212 if (recursive) {
1213 VectorSet visited;
1214 verify_adr_type(visited, _adr_type);
1215 }
1216 }
1217 #endif
1218
1219
1220 //------------------------------Value------------------------------------------
1221 // Compute the type of the PhiNode
1222 const Type* PhiNode::Value(PhaseGVN* phase) const {
1223 Node *r = in(0); // RegionNode
1224 if( !r ) // Copy or dead
1225 return in(1) ? phase->type(in(1)) : Type::TOP;
1226
1227 // Note: During parsing, phis are often transformed before their regions.
1228 // This means we have to use type_or_null to defend against untyped regions.
1229 if( phase->type_or_null(r) == Type::TOP ) // Dead code?
1230 return Type::TOP;
1231
1232 // Check for trip-counted loop. If so, be smarter.
1233 BaseCountedLoopNode* l = r->is_BaseCountedLoop() ? r->as_BaseCountedLoop() : nullptr;
1234 if (l && ((const Node*)l->phi() == this)) { // Trip counted loop!
1235 // protect against init_trip() or limit() returning null
1236 if (l->can_be_counted_loop(phase)) {
1237 const Node* init = l->init_trip();
1238 const Node* limit = l->limit();
1239 const Node* stride = l->stride();
1240 if (init != nullptr && limit != nullptr && stride != nullptr) {
1241 const TypeInteger* lo = phase->type(init)->isa_integer(l->bt());
1242 const TypeInteger* hi = phase->type(limit)->isa_integer(l->bt());
1243 const TypeInteger* stride_t = phase->type(stride)->isa_integer(l->bt());
1244 if (lo != nullptr && hi != nullptr && stride_t != nullptr) { // Dying loops might have TOP here
1245 assert(stride_t->is_con(), "bad stride type");
1246 BoolTest::mask bt = l->loopexit()->test_trip();
1247 // If the loop exit condition is "not equal", the condition
1248 // would not trigger if init > limit (if stride > 0) or if
1249 // init < limit if (stride > 0) so we can't deduce bounds
1250 // for the iv from the exit condition.
1251 if (bt != BoolTest::ne) {
1252 jlong stride_con = stride_t->get_con_as_long(l->bt());
1253 if (stride_con < 0) { // Down-counter loop
1254 swap(lo, hi);
1255 jlong iv_range_lower_limit = lo->lo_as_long();
1256 // Prevent overflow when adding one below
1257 if (iv_range_lower_limit < max_signed_integer(l->bt())) {
1258 // The loop exit condition is: iv + stride > limit (iv is this Phi). So the loop iterates until
1259 // iv + stride <= limit
1260 // We know that: limit >= lo->lo_as_long() and stride <= -1
1261 // So when the loop exits, iv has to be at most lo->lo_as_long() + 1
1262 iv_range_lower_limit += 1; // lo is after decrement
1263 // Exact bounds for the phi can be computed when ABS(stride) greater than 1 if bounds are constant.
1264 if (lo->is_con() && hi->is_con() && hi->lo_as_long() > lo->hi_as_long() && stride_con != -1) {
1265 julong uhi = static_cast<julong>(hi->lo_as_long());
1266 julong ulo = static_cast<julong>(lo->hi_as_long());
1267 julong diff = ((uhi - ulo - 1) / (-stride_con)) * (-stride_con);
1268 julong ufirst = hi->lo_as_long() - diff;
1269 iv_range_lower_limit = reinterpret_cast<jlong &>(ufirst);
1270 assert(iv_range_lower_limit >= lo->lo_as_long() + 1, "should end up with narrower range");
1271 }
1272 }
1273 return TypeInteger::make(MIN2(iv_range_lower_limit, hi->lo_as_long()), hi->hi_as_long(), 3, l->bt())->filter_speculative(_type);
1274 } else if (stride_con >= 0) {
1275 jlong iv_range_upper_limit = hi->hi_as_long();
1276 // Prevent overflow when subtracting one below
1277 if (iv_range_upper_limit > min_signed_integer(l->bt())) {
1278 // The loop exit condition is: iv + stride < limit (iv is this Phi). So the loop iterates until
1279 // iv + stride >= limit
1280 // We know that: limit <= hi->hi_as_long() and stride >= 1
1281 // So when the loop exits, iv has to be at most hi->hi_as_long() - 1
1282 iv_range_upper_limit -= 1;
1283 // Exact bounds for the phi can be computed when ABS(stride) greater than 1 if bounds are constant.
1284 if (lo->is_con() && hi->is_con() && hi->lo_as_long() > lo->hi_as_long() && stride_con != 1) {
1285 julong uhi = static_cast<julong>(hi->lo_as_long());
1286 julong ulo = static_cast<julong>(lo->hi_as_long());
1287 julong diff = ((uhi - ulo - 1) / stride_con) * stride_con;
1288 julong ulast = lo->hi_as_long() + diff;
1289 iv_range_upper_limit = reinterpret_cast<jlong &>(ulast);
1290 assert(iv_range_upper_limit <= hi->hi_as_long() - 1, "should end up with narrower range");
1291 }
1292 }
1293 return TypeInteger::make(lo->lo_as_long(), MAX2(lo->hi_as_long(), iv_range_upper_limit), 3, l->bt())->filter_speculative(_type);
1294 }
1295 }
1296 }
1297 }
1298 } else if (l->in(LoopNode::LoopBackControl) != nullptr &&
1299 in(LoopNode::EntryControl) != nullptr &&
1300 phase->type(l->in(LoopNode::LoopBackControl)) == Type::TOP) {
1301 // During CCP, if we saturate the type of a counted loop's Phi
1302 // before the special code for counted loop above has a chance
1303 // to run (that is as long as the type of the backedge's control
1304 // is top), we might end up with non monotonic types
1305 return phase->type(in(LoopNode::EntryControl))->filter_speculative(_type);
1306 }
1307 }
1308
1309 // Default case: merge all inputs
1310 const Type *t = Type::TOP; // Merged type starting value
1311 for (uint i = 1; i < req(); ++i) {// For all paths in
1312 // Reachable control path?
1313 if (r->in(i) && phase->type(r->in(i)) == Type::CONTROL) {
1314 const Type* ti = phase->type(in(i));
1315 t = t->meet_speculative(ti);
1316 }
1317 }
1318
1319 // The worst-case type (from ciTypeFlow) should be consistent with "t".
1320 // That is, we expect that "t->higher_equal(_type)" holds true.
1321 // There are various exceptions:
1322 // - Inputs which are phis might in fact be widened unnecessarily.
1323 // For example, an input might be a widened int while the phi is a short.
1324 // - Inputs might be BotPtrs but this phi is dependent on a null check,
1325 // and postCCP has removed the cast which encodes the result of the check.
1326 // - The type of this phi is an interface, and the inputs are classes.
1327 // - Value calls on inputs might produce fuzzy results.
1328 // (Occurrences of this case suggest improvements to Value methods.)
1329 //
1330 // It is not possible to see Type::BOTTOM values as phi inputs,
1331 // because the ciTypeFlow pre-pass produces verifier-quality types.
1332 const Type* ft = t->filter_speculative(_type); // Worst case type
1333
1334 #ifdef ASSERT
1335 // The following logic has been moved into TypeOopPtr::filter.
1336 const Type* jt = t->join_speculative(_type);
1337 if (jt->empty()) { // Emptied out???
1338 // Otherwise it's something stupid like non-overlapping int ranges
1339 // found on dying counted loops.
1340 assert(ft == Type::TOP, ""); // Canonical empty value
1341 }
1342
1343 else {
1344
1345 if (jt != ft && jt->base() == ft->base()) {
1346 if (jt->isa_int() &&
1347 jt->is_int()->_lo == ft->is_int()->_lo &&
1348 jt->is_int()->_hi == ft->is_int()->_hi)
1349 jt = ft;
1350 if (jt->isa_long() &&
1351 jt->is_long()->_lo == ft->is_long()->_lo &&
1352 jt->is_long()->_hi == ft->is_long()->_hi)
1353 jt = ft;
1354 }
1355 if (jt != ft) {
1356 tty->print("merge type: "); t->dump(); tty->cr();
1357 tty->print("kill type: "); _type->dump(); tty->cr();
1358 tty->print("join type: "); jt->dump(); tty->cr();
1359 tty->print("filter type: "); ft->dump(); tty->cr();
1360 }
1361 assert(jt == ft, "");
1362 }
1363 #endif //ASSERT
1364
1365 // Deal with conversion problems found in data loops.
1366 ft = phase->saturate_and_maybe_push_to_igvn_worklist(this, ft);
1367 return ft;
1368 }
1369
1370 // Does this Phi represent a simple well-shaped diamond merge? Return the
1371 // index of the true path or 0 otherwise.
1372 int PhiNode::is_diamond_phi() const {
1373 Node* region = in(0);
1374 assert(region != nullptr && region->is_Region(), "phi must have region");
1375 if (!region->as_Region()->is_diamond()) {
1376 return 0;
1377 }
1378
1379 if (region->in(1)->is_IfTrue()) {
1380 assert(region->in(2)->is_IfFalse(), "bad If");
1381 return 1;
1382 } else {
1383 // Flipped projections.
1384 assert(region->in(2)->is_IfTrue(), "bad If");
1385 return 2;
1386 }
1387 }
1388
1389 // Do the following transformation if we find the corresponding graph shape, remove the involved memory phi and return
1390 // true. Otherwise, return false if the transformation cannot be applied.
1391 //
1392 // If If
1393 // / \ / \
1394 // IfFalse IfTrue /- Some Node IfFalse IfTrue
1395 // \ / / / \ / Some Node
1396 // Region / /-MergeMem ===> Region |
1397 // / \---Phi | MergeMem
1398 // [other phis] \ [other phis] |
1399 // use use
1400 bool PhiNode::try_clean_memory_phi(PhaseIterGVN* igvn) {
1401 if (_type != Type::MEMORY) {
1402 return false;
1403 }
1404 assert(is_diamond_phi() > 0, "sanity");
1405 assert(req() == 3, "same as region");
1406 const Node* region = in(0);
1407 for (uint i = 1; i < 3; i++) {
1408 Node* phi_input = in(i);
1409 if (phi_input != nullptr && phi_input->is_MergeMem() && region->in(i)->outcnt() == 1) {
1410 // Nothing is control-dependent on path #i except the region itself.
1411 MergeMemNode* merge_mem = phi_input->as_MergeMem();
1412 uint j = 3 - i;
1413 Node* other_phi_input = in(j);
1414 if (other_phi_input != nullptr && other_phi_input == merge_mem->base_memory()) {
1415 // merge_mem is a successor memory to other_phi_input, and is not pinned inside the diamond, so push it out.
1416 // This will allow the diamond to collapse completely if there are no other phis left.
1417 igvn->replace_node(this, merge_mem);
1418 return true;
1419 }
1420 }
1421 }
1422 return false;
1423 }
1424
1425 //----------------------------check_cmove_id-----------------------------------
1426 // Check for CMove'ing a constant after comparing against the constant.
1427 // Happens all the time now, since if we compare equality vs a constant in
1428 // the parser, we "know" the variable is constant on one path and we force
1429 // it. Thus code like "if( x==0 ) {/*EMPTY*/}" ends up inserting a
1430 // conditional move: "x = (x==0)?0:x;". Yucko. This fix is slightly more
1431 // general in that we don't need constants. Since CMove's are only inserted
1432 // in very special circumstances, we do it here on generic Phi's.
1433 Node* PhiNode::is_cmove_id(PhaseTransform* phase, int true_path) {
1434 assert(true_path !=0, "only diamond shape graph expected");
1435
1436 // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1437 // phi->region->if_proj->ifnode->bool->cmp
1438 Node* region = in(0);
1439 Node* iff = region->in(1)->in(0);
1440 BoolNode* b = iff->in(1)->as_Bool();
1441 Node* cmp = b->in(1);
1442 Node* tval = in(true_path);
1443 Node* fval = in(3-true_path);
1444 Node* id = CMoveNode::is_cmove_id(phase, cmp, tval, fval, b);
1445 if (id == nullptr)
1446 return nullptr;
1447
1448 // Either value might be a cast that depends on a branch of 'iff'.
1449 // Since the 'id' value will float free of the diamond, either
1450 // decast or return failure.
1451 Node* ctl = id->in(0);
1452 if (ctl != nullptr && ctl->in(0) == iff) {
1453 if (id->is_ConstraintCast()) {
1454 return id->in(1);
1455 } else {
1456 // Don't know how to disentangle this value.
1457 return nullptr;
1458 }
1459 }
1460
1461 return id;
1462 }
1463
1464 //------------------------------Identity---------------------------------------
1465 // Check for Region being Identity.
1466 Node* PhiNode::Identity(PhaseGVN* phase) {
1467 if (must_wait_for_region_in_irreducible_loop(phase)) {
1468 return this;
1469 }
1470 // Check for no merging going on
1471 // (There used to be special-case code here when this->region->is_Loop.
1472 // It would check for a tributary phi on the backedge that the main phi
1473 // trivially, perhaps with a single cast. The unique_input method
1474 // does all this and more, by reducing such tributaries to 'this'.)
1475 Node* uin = unique_input(phase, false);
1476 if (uin != nullptr) {
1477 return uin;
1478 }
1479 uin = unique_constant_input_recursive(phase);
1480 if (uin != nullptr) {
1481 return uin;
1482 }
1483
1484 int true_path = is_diamond_phi();
1485 // Delay CMove'ing identity if Ideal has not had the chance to handle unsafe cases, yet.
1486 if (true_path != 0 && !(phase->is_IterGVN() && wait_for_region_igvn(phase))) {
1487 Node* id = is_cmove_id(phase, true_path);
1488 if (id != nullptr) {
1489 return id;
1490 }
1491 }
1492
1493 // Looking for phis with identical inputs. If we find one that has
1494 // type TypePtr::BOTTOM, replace the current phi with the bottom phi.
1495 if (phase->is_IterGVN() && type() == Type::MEMORY && adr_type() !=
1496 TypePtr::BOTTOM && !adr_type()->is_known_instance()) {
1497 uint phi_len = req();
1498 Node* phi_reg = region();
1499 for (DUIterator_Fast imax, i = phi_reg->fast_outs(imax); i < imax; i++) {
1500 Node* u = phi_reg->fast_out(i);
1501 if (u->is_Phi() && u->as_Phi()->type() == Type::MEMORY &&
1502 u->adr_type() == TypePtr::BOTTOM && u->in(0) == phi_reg &&
1503 u->req() == phi_len) {
1504 for (uint j = 1; j < phi_len; j++) {
1505 if (in(j) != u->in(j)) {
1506 u = nullptr;
1507 break;
1508 }
1509 }
1510 if (u != nullptr) {
1511 return u;
1512 }
1513 }
1514 }
1515 }
1516
1517 return this; // No identity
1518 }
1519
1520 //-----------------------------unique_input------------------------------------
1521 // Find the unique value, discounting top, self-loops, and casts.
1522 // Return top if there are no inputs, and self if there are multiple.
1523 Node* PhiNode::unique_input(PhaseValues* phase, bool uncast) {
1524 // 1) One unique direct input,
1525 // or if uncast is true:
1526 // 2) some of the inputs have an intervening ConstraintCast
1527 // 3) an input is a self loop
1528 //
1529 // 1) input or 2) input or 3) input __
1530 // / \ / \ \ / \
1531 // \ / | cast phi cast
1532 // phi \ / / \ /
1533 // phi / --
1534
1535 Node* r = in(0); // RegionNode
1536 Node* input = nullptr; // The unique direct input (maybe uncasted = ConstraintCasts removed)
1537
1538 for (uint i = 1, cnt = req(); i < cnt; ++i) {
1539 Node* rc = r->in(i);
1540 if (rc == nullptr || phase->type(rc) == Type::TOP)
1541 continue; // ignore unreachable control path
1542 Node* n = in(i);
1543 if (n == nullptr)
1544 continue;
1545 Node* un = n;
1546 if (uncast) {
1547 #ifdef ASSERT
1548 Node* m = un->uncast();
1549 #endif
1550 while (un != nullptr && un->req() == 2 && un->is_ConstraintCast()) {
1551 Node* next = un->in(1);
1552 if (phase->type(next)->isa_rawptr() && phase->type(un)->isa_oopptr()) {
1553 // risk exposing raw ptr at safepoint
1554 break;
1555 }
1556 un = next;
1557 }
1558 assert(m == un || un->in(1) == m, "Only expected at CheckCastPP from allocation");
1559 }
1560 if (un == nullptr || un == this || phase->type(un) == Type::TOP) {
1561 continue; // ignore if top, or in(i) and "this" are in a data cycle
1562 }
1563 // Check for a unique input (maybe uncasted)
1564 if (input == nullptr) {
1565 input = un;
1566 } else if (input != un) {
1567 input = NodeSentinel; // no unique input
1568 }
1569 }
1570 if (input == nullptr) {
1571 return phase->C->top(); // no inputs
1572 }
1573
1574 if (input != NodeSentinel) {
1575 return input; // one unique direct input
1576 }
1577
1578 // Nothing.
1579 return nullptr;
1580 }
1581
1582 // Find the unique input, try to look recursively through input Phis
1583 Node* PhiNode::unique_constant_input_recursive(PhaseGVN* phase) {
1584 if (!phase->is_IterGVN()) {
1585 return nullptr;
1586 }
1587
1588 ResourceMark rm;
1589 Node* unique = nullptr;
1590 Unique_Node_List visited;
1591 visited.push(this);
1592
1593 for (uint visited_idx = 0; visited_idx < visited.size(); visited_idx++) {
1594 Node* current = visited.at(visited_idx);
1595 for (uint i = 1; i < current->req(); i++) {
1596 Node* phi_in = current->in(i);
1597 if (phi_in == nullptr) {
1598 continue;
1599 }
1600
1601 if (phi_in->is_Phi()) {
1602 visited.push(phi_in);
1603 } else {
1604 if (unique == nullptr) {
1605 if (!phi_in->is_Con()) {
1606 return nullptr;
1607 }
1608 unique = phi_in;
1609 } else if (unique != phi_in) {
1610 return nullptr;
1611 }
1612 }
1613 }
1614 }
1615 return unique;
1616 }
1617
1618 //------------------------------is_x2logic-------------------------------------
1619 // Check for simple convert-to-boolean pattern
1620 // If:(C Bool) Region:(IfF IfT) Phi:(Region 0 1)
1621 // Convert Phi to an ConvIB.
1622 static Node *is_x2logic( PhaseGVN *phase, PhiNode *phi, int true_path ) {
1623 assert(true_path !=0, "only diamond shape graph expected");
1624
1625 // If we're late in the optimization process, we may have already expanded Conv2B nodes
1626 if (phase->C->post_loop_opts_phase() && !Matcher::match_rule_supported(Op_Conv2B)) {
1627 return nullptr;
1628 }
1629
1630 // Convert the true/false index into an expected 0/1 return.
1631 // Map 2->0 and 1->1.
1632 int flipped = 2-true_path;
1633
1634 // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1635 // phi->region->if_proj->ifnode->bool->cmp
1636 Node *region = phi->in(0);
1637 Node *iff = region->in(1)->in(0);
1638 BoolNode *b = (BoolNode*)iff->in(1);
1639 const CmpNode *cmp = (CmpNode*)b->in(1);
1640
1641 Node *zero = phi->in(1);
1642 Node *one = phi->in(2);
1643 const Type *tzero = phase->type( zero );
1644 const Type *tone = phase->type( one );
1645
1646 // Check for compare vs 0
1647 const Type *tcmp = phase->type(cmp->in(2));
1648 if( tcmp != TypeInt::ZERO && tcmp != TypePtr::NULL_PTR ) {
1649 // Allow cmp-vs-1 if the other input is bounded by 0-1
1650 if( !(tcmp == TypeInt::ONE && phase->type(cmp->in(1)) == TypeInt::BOOL) )
1651 return nullptr;
1652 flipped = 1-flipped; // Test is vs 1 instead of 0!
1653 }
1654
1655 // Check for setting zero/one opposite expected
1656 if( tzero == TypeInt::ZERO ) {
1657 if( tone == TypeInt::ONE ) {
1658 } else return nullptr;
1659 } else if( tzero == TypeInt::ONE ) {
1660 if( tone == TypeInt::ZERO ) {
1661 flipped = 1-flipped;
1662 } else return nullptr;
1663 } else return nullptr;
1664
1665 // Check for boolean test backwards
1666 if( b->_test._test == BoolTest::ne ) {
1667 } else if( b->_test._test == BoolTest::eq ) {
1668 flipped = 1-flipped;
1669 } else return nullptr;
1670
1671 // Build int->bool conversion
1672 Node* n = new Conv2BNode(cmp->in(1));
1673 if (flipped) {
1674 n = new XorINode(phase->transform(n), phase->intcon(1));
1675 }
1676
1677 return n;
1678 }
1679
1680 //------------------------------is_cond_add------------------------------------
1681 // Check for simple conditional add pattern: "(P < Q) ? X+Y : X;"
1682 // To be profitable the control flow has to disappear; there can be no other
1683 // values merging here. We replace the test-and-branch with:
1684 // "(sgn(P-Q))&Y) + X". Basically, convert "(P < Q)" into 0 or -1 by
1685 // moving the carry bit from (P-Q) into a register with 'sbb EAX,EAX'.
1686 // Then convert Y to 0-or-Y and finally add.
1687 // This is a key transform for SpecJava _201_compress.
1688 static Node* is_cond_add(PhaseGVN *phase, PhiNode *phi, int true_path) {
1689 assert(true_path !=0, "only diamond shape graph expected");
1690
1691 // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1692 // phi->region->if_proj->ifnode->bool->cmp
1693 RegionNode *region = (RegionNode*)phi->in(0);
1694 Node *iff = region->in(1)->in(0);
1695 BoolNode* b = iff->in(1)->as_Bool();
1696 const CmpNode *cmp = (CmpNode*)b->in(1);
1697
1698 // Make sure only merging this one phi here
1699 if (region->has_unique_phi() != phi) return nullptr;
1700
1701 // Make sure each arm of the diamond has exactly one output, which we assume
1702 // is the region. Otherwise, the control flow won't disappear.
1703 if (region->in(1)->outcnt() != 1) return nullptr;
1704 if (region->in(2)->outcnt() != 1) return nullptr;
1705
1706 // Check for "(P < Q)" of type signed int
1707 if (b->_test._test != BoolTest::lt) return nullptr;
1708 if (cmp->Opcode() != Op_CmpI) return nullptr;
1709
1710 Node *p = cmp->in(1);
1711 Node *q = cmp->in(2);
1712 Node *n1 = phi->in( true_path);
1713 Node *n2 = phi->in(3-true_path);
1714
1715 int op = n1->Opcode();
1716 if( op != Op_AddI // Need zero as additive identity
1717 /*&&op != Op_SubI &&
1718 op != Op_AddP &&
1719 op != Op_XorI &&
1720 op != Op_OrI*/ )
1721 return nullptr;
1722
1723 Node *x = n2;
1724 Node *y = nullptr;
1725 if( x == n1->in(1) ) {
1726 y = n1->in(2);
1727 } else if( x == n1->in(2) ) {
1728 y = n1->in(1);
1729 } else return nullptr;
1730
1731 // Not so profitable if compare and add are constants
1732 if( q->is_Con() && phase->type(q) != TypeInt::ZERO && y->is_Con() )
1733 return nullptr;
1734
1735 Node *cmplt = phase->transform( new CmpLTMaskNode(p,q) );
1736 Node *j_and = phase->transform( new AndINode(cmplt,y) );
1737 return new AddINode(j_and,x);
1738 }
1739
1740 //------------------------------is_absolute------------------------------------
1741 // Check for absolute value.
1742 static Node* is_absolute( PhaseGVN *phase, PhiNode *phi_root, int true_path) {
1743 assert(true_path !=0, "only diamond shape graph expected");
1744
1745 int cmp_zero_idx = 0; // Index of compare input where to look for zero
1746 int phi_x_idx = 0; // Index of phi input where to find naked x
1747
1748 // ABS ends with the merge of 2 control flow paths.
1749 // Find the false path from the true path. With only 2 inputs, 3 - x works nicely.
1750 int false_path = 3 - true_path;
1751
1752 // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1753 // phi->region->if_proj->ifnode->bool->cmp
1754 BoolNode *bol = phi_root->in(0)->in(1)->in(0)->in(1)->as_Bool();
1755 Node *cmp = bol->in(1);
1756
1757 // Check bool sense
1758 if (cmp->Opcode() == Op_CmpF || cmp->Opcode() == Op_CmpD) {
1759 switch (bol->_test._test) {
1760 case BoolTest::lt: cmp_zero_idx = 1; phi_x_idx = true_path; break;
1761 case BoolTest::le: cmp_zero_idx = 2; phi_x_idx = false_path; break;
1762 case BoolTest::gt: cmp_zero_idx = 2; phi_x_idx = true_path; break;
1763 case BoolTest::ge: cmp_zero_idx = 1; phi_x_idx = false_path; break;
1764 default: return nullptr; break;
1765 }
1766 } else if (cmp->Opcode() == Op_CmpI || cmp->Opcode() == Op_CmpL) {
1767 switch (bol->_test._test) {
1768 case BoolTest::lt:
1769 case BoolTest::le: cmp_zero_idx = 2; phi_x_idx = false_path; break;
1770 case BoolTest::gt:
1771 case BoolTest::ge: cmp_zero_idx = 2; phi_x_idx = true_path; break;
1772 default: return nullptr; break;
1773 }
1774 }
1775
1776 // Test is next
1777 const Type *tzero = nullptr;
1778 switch (cmp->Opcode()) {
1779 case Op_CmpI: tzero = TypeInt::ZERO; break; // Integer ABS
1780 case Op_CmpL: tzero = TypeLong::ZERO; break; // Long ABS
1781 case Op_CmpF: tzero = TypeF::ZERO; break; // Float ABS
1782 case Op_CmpD: tzero = TypeD::ZERO; break; // Double ABS
1783 default: return nullptr;
1784 }
1785
1786 // Find zero input of compare; the other input is being abs'd
1787 Node *x = nullptr;
1788 bool flip = false;
1789 if( phase->type(cmp->in(cmp_zero_idx)) == tzero ) {
1790 x = cmp->in(3 - cmp_zero_idx);
1791 } else if( phase->type(cmp->in(3 - cmp_zero_idx)) == tzero ) {
1792 // The test is inverted, we should invert the result...
1793 x = cmp->in(cmp_zero_idx);
1794 flip = true;
1795 } else {
1796 return nullptr;
1797 }
1798
1799 // Next get the 2 pieces being selected, one is the original value
1800 // and the other is the negated value.
1801 if( phi_root->in(phi_x_idx) != x ) return nullptr;
1802
1803 // Check other phi input for subtract node
1804 Node *sub = phi_root->in(3 - phi_x_idx);
1805
1806 bool is_sub = sub->Opcode() == Op_SubF || sub->Opcode() == Op_SubD ||
1807 sub->Opcode() == Op_SubI || sub->Opcode() == Op_SubL;
1808
1809 // Allow only Sub(0,X) and fail out for all others; Neg is not OK
1810 if (!is_sub || phase->type(sub->in(1)) != tzero || sub->in(2) != x) return nullptr;
1811
1812 if (tzero == TypeF::ZERO) {
1813 x = new AbsFNode(x);
1814 if (flip) {
1815 x = new SubFNode(sub->in(1), phase->transform(x));
1816 }
1817 } else if (tzero == TypeD::ZERO) {
1818 x = new AbsDNode(x);
1819 if (flip) {
1820 x = new SubDNode(sub->in(1), phase->transform(x));
1821 }
1822 } else if (tzero == TypeInt::ZERO && Matcher::match_rule_supported(Op_AbsI)) {
1823 x = new AbsINode(x);
1824 if (flip) {
1825 x = new SubINode(sub->in(1), phase->transform(x));
1826 }
1827 } else if (tzero == TypeLong::ZERO && Matcher::match_rule_supported(Op_AbsL)) {
1828 x = new AbsLNode(x);
1829 if (flip) {
1830 x = new SubLNode(sub->in(1), phase->transform(x));
1831 }
1832 } else return nullptr;
1833
1834 return x;
1835 }
1836
1837 //------------------------------split_once-------------------------------------
1838 // Helper for split_flow_path
1839 static void split_once(PhaseIterGVN *igvn, Node *phi, Node *val, Node *n, Node *newn) {
1840 igvn->hash_delete(n); // Remove from hash before hacking edges
1841
1842 uint j = 1;
1843 for (uint i = phi->req()-1; i > 0; i--) {
1844 if (phi->in(i) == val) { // Found a path with val?
1845 // Add to NEW Region/Phi, no DU info
1846 newn->set_req( j++, n->in(i) );
1847 // Remove from OLD Region/Phi
1848 n->del_req(i);
1849 }
1850 }
1851
1852 // Register the new node but do not transform it. Cannot transform until the
1853 // entire Region/Phi conglomerate has been hacked as a single huge transform.
1854 igvn->register_new_node_with_optimizer( newn );
1855
1856 // Now I can point to the new node.
1857 n->add_req(newn);
1858 igvn->_worklist.push(n);
1859 }
1860
1861 //------------------------------split_flow_path--------------------------------
1862 // Check for merging identical values and split flow paths
1863 static Node* split_flow_path(PhaseGVN *phase, PhiNode *phi) {
1864 // This optimization tries to find two or more inputs of phi with the same constant value
1865 // It then splits them into a separate Phi, and according Region. If this is a loop-entry,
1866 // and the loop entry has multiple fall-in edges, and some of those fall-in edges have that
1867 // constant, and others not, we may split the fall-in edges into separate Phi's, and create
1868 // an irreducible loop. For reducible loops, this never seems to happen, as the multiple
1869 // fall-in edges are already merged before the loop head during parsing. But with irreducible
1870 // loops present the order or merging during parsing can sometimes prevent this.
1871 if (phase->C->has_irreducible_loop()) {
1872 // Avoid this optimization if any irreducible loops are present. Else we may create
1873 // an irreducible loop that we do not detect.
1874 return nullptr;
1875 }
1876 BasicType bt = phi->type()->basic_type();
1877 if( bt == T_ILLEGAL || type2size[bt] <= 0 )
1878 return nullptr; // Bail out on funny non-value stuff
1879 if( phi->req() <= 3 ) // Need at least 2 matched inputs and a
1880 return nullptr; // third unequal input to be worth doing
1881
1882 // Scan for a constant
1883 uint i;
1884 for( i = 1; i < phi->req()-1; i++ ) {
1885 Node *n = phi->in(i);
1886 if( !n ) return nullptr;
1887 if( phase->type(n) == Type::TOP ) return nullptr;
1888 if( n->Opcode() == Op_ConP || n->Opcode() == Op_ConN || n->Opcode() == Op_ConNKlass )
1889 break;
1890 }
1891 if( i >= phi->req() ) // Only split for constants
1892 return nullptr;
1893
1894 Node *val = phi->in(i); // Constant to split for
1895 uint hit = 0; // Number of times it occurs
1896 Node *r = phi->region();
1897
1898 for( ; i < phi->req(); i++ ){ // Count occurrences of constant
1899 Node *n = phi->in(i);
1900 if( !n ) return nullptr;
1901 if( phase->type(n) == Type::TOP ) return nullptr;
1902 if( phi->in(i) == val ) {
1903 hit++;
1904 if (Node::may_be_loop_entry(r->in(i))) {
1905 return nullptr; // don't split loop entry path
1906 }
1907 }
1908 }
1909
1910 if( hit <= 1 || // Make sure we find 2 or more
1911 hit == phi->req()-1 ) // and not ALL the same value
1912 return nullptr;
1913
1914 // Now start splitting out the flow paths that merge the same value.
1915 // Split first the RegionNode.
1916 PhaseIterGVN *igvn = phase->is_IterGVN();
1917 RegionNode *newr = new RegionNode(hit+1);
1918 split_once(igvn, phi, val, r, newr);
1919
1920 // Now split all other Phis than this one
1921 for (DUIterator_Fast kmax, k = r->fast_outs(kmax); k < kmax; k++) {
1922 Node* phi2 = r->fast_out(k);
1923 if( phi2->is_Phi() && phi2->as_Phi() != phi ) {
1924 PhiNode *newphi = PhiNode::make_blank(newr, phi2);
1925 split_once(igvn, phi, val, phi2, newphi);
1926 }
1927 }
1928
1929 // Clean up this guy
1930 igvn->hash_delete(phi);
1931 for( i = phi->req()-1; i > 0; i-- ) {
1932 if( phi->in(i) == val ) {
1933 phi->del_req(i);
1934 }
1935 }
1936 phi->add_req(val);
1937
1938 return phi;
1939 }
1940
1941 // Returns the BasicType of a given convert node and a type, with special handling to ensure that conversions to
1942 // and from half float will return the SHORT basic type, as that wouldn't be returned typically from TypeInt.
1943 static BasicType get_convert_type(Node* convert, const Type* type) {
1944 int convert_op = convert->Opcode();
1945 if (type->isa_int() && (convert_op == Op_ConvHF2F || convert_op == Op_ConvF2HF)) {
1946 return T_SHORT;
1947 }
1948
1949 return type->basic_type();
1950 }
1951
1952 //=============================================================================
1953 //------------------------------simple_data_loop_check-------------------------
1954 // Try to determining if the phi node in a simple safe/unsafe data loop.
1955 // Returns:
1956 // enum LoopSafety { Safe = 0, Unsafe, UnsafeLoop };
1957 // Safe - safe case when the phi and it's inputs reference only safe data
1958 // nodes;
1959 // Unsafe - the phi and it's inputs reference unsafe data nodes but there
1960 // is no reference back to the phi - need a graph walk
1961 // to determine if it is in a loop;
1962 // UnsafeLoop - unsafe case when the phi references itself directly or through
1963 // unsafe data node.
1964 // Note: a safe data node is a node which could/never reference itself during
1965 // GVN transformations. For now it is Con, Proj, Phi, CastPP, CheckCastPP.
1966 // I mark Phi nodes as safe node not only because they can reference itself
1967 // but also to prevent mistaking the fallthrough case inside an outer loop
1968 // as dead loop when the phi references itself through an other phi.
1969 PhiNode::LoopSafety PhiNode::simple_data_loop_check(Node *in) const {
1970 // It is unsafe loop if the phi node references itself directly.
1971 if (in == (Node*)this)
1972 return UnsafeLoop; // Unsafe loop
1973 // Unsafe loop if the phi node references itself through an unsafe data node.
1974 // Exclude cases with null inputs or data nodes which could reference
1975 // itself (safe for dead loops).
1976 if (in != nullptr && !in->is_dead_loop_safe()) {
1977 // Check inputs of phi's inputs also.
1978 // It is much less expensive then full graph walk.
1979 uint cnt = in->req();
1980 uint i = (in->is_Proj() && !in->is_CFG()) ? 0 : 1;
1981 for (; i < cnt; ++i) {
1982 Node* m = in->in(i);
1983 if (m == (Node*)this)
1984 return UnsafeLoop; // Unsafe loop
1985 if (m != nullptr && !m->is_dead_loop_safe()) {
1986 // Check the most common case (about 30% of all cases):
1987 // phi->Load/Store->AddP->(ConP ConP Con)/(Parm Parm Con).
1988 Node *m1 = (m->is_AddP() && m->req() > 3) ? m->in(1) : nullptr;
1989 if (m1 == (Node*)this)
1990 return UnsafeLoop; // Unsafe loop
1991 if (m1 != nullptr && m1 == m->in(2) &&
1992 m1->is_dead_loop_safe() && m->in(3)->is_Con()) {
1993 continue; // Safe case
1994 }
1995 // The phi references an unsafe node - need full analysis.
1996 return Unsafe;
1997 }
1998 }
1999 }
2000 return Safe; // Safe case - we can optimize the phi node.
2001 }
2002
2003 //------------------------------is_unsafe_data_reference-----------------------
2004 // If phi can be reached through the data input - it is data loop.
2005 bool PhiNode::is_unsafe_data_reference(Node *in) const {
2006 assert(req() > 1, "");
2007 // First, check simple cases when phi references itself directly or
2008 // through an other node.
2009 LoopSafety safety = simple_data_loop_check(in);
2010 if (safety == UnsafeLoop)
2011 return true; // phi references itself - unsafe loop
2012 else if (safety == Safe)
2013 return false; // Safe case - phi could be replaced with the unique input.
2014
2015 // Unsafe case when we should go through data graph to determine
2016 // if the phi references itself.
2017
2018 ResourceMark rm;
2019
2020 Node_List nstack;
2021 VectorSet visited;
2022
2023 nstack.push(in); // Start with unique input.
2024 visited.set(in->_idx);
2025 while (nstack.size() != 0) {
2026 Node* n = nstack.pop();
2027 uint cnt = n->req();
2028 uint i = (n->is_Proj() && !n->is_CFG()) ? 0 : 1;
2029 for (; i < cnt; i++) {
2030 Node* m = n->in(i);
2031 if (m == (Node*)this) {
2032 return true; // Data loop
2033 }
2034 if (m != nullptr && !m->is_dead_loop_safe()) { // Only look for unsafe cases.
2035 if (!visited.test_set(m->_idx))
2036 nstack.push(m);
2037 }
2038 }
2039 }
2040 return false; // The phi is not reachable from its inputs
2041 }
2042
2043 // Is this Phi's region or some inputs to the region enqueued for IGVN
2044 // and so could cause the region to be optimized out?
2045 bool PhiNode::wait_for_region_igvn(PhaseGVN* phase) {
2046 PhaseIterGVN* igvn = phase->is_IterGVN();
2047 Unique_Node_List& worklist = igvn->_worklist;
2048 bool delay = false;
2049 Node* r = in(0);
2050 for (uint j = 1; j < req(); j++) {
2051 Node* rc = r->in(j);
2052 Node* n = in(j);
2053
2054 if (rc == nullptr || !rc->is_Proj()) { continue; }
2055 if (worklist.member(rc)) {
2056 delay = true;
2057 break;
2058 }
2059
2060 if (rc->in(0) == nullptr || !rc->in(0)->is_If()) { continue; }
2061 if (worklist.member(rc->in(0))) {
2062 delay = true;
2063 break;
2064 }
2065
2066 if (rc->in(0)->in(1) == nullptr || !rc->in(0)->in(1)->is_Bool()) { continue; }
2067 if (worklist.member(rc->in(0)->in(1))) {
2068 delay = true;
2069 break;
2070 }
2071
2072 if (rc->in(0)->in(1)->in(1) == nullptr || !rc->in(0)->in(1)->in(1)->is_Cmp()) { continue; }
2073 if (worklist.member(rc->in(0)->in(1)->in(1))) {
2074 delay = true;
2075 break;
2076 }
2077 }
2078
2079 if (delay) {
2080 worklist.push(this);
2081 }
2082 return delay;
2083 }
2084
2085 // Push inline type input nodes (and null) down through the phi recursively (can handle data loops).
2086 InlineTypeNode* PhiNode::push_inline_types_down(PhaseGVN* phase, bool can_reshape, ciInlineKlass* inline_klass) {
2087 assert(inline_klass != nullptr, "must be");
2088 InlineTypeNode* vt = InlineTypeNode::make_null(*phase, inline_klass, /* transform = */ false)->clone_with_phis(phase, in(0), nullptr, !_type->maybe_null(), true);
2089 if (can_reshape) {
2090 // Replace phi right away to be able to use the inline
2091 // type node when reaching the phi again through data loops.
2092 PhaseIterGVN* igvn = phase->is_IterGVN();
2093 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2094 Node* u = fast_out(i);
2095 igvn->rehash_node_delayed(u);
2096 imax -= u->replace_edge(this, vt);
2097 --i;
2098 }
2099 igvn->rehash_node_delayed(this);
2100 assert(outcnt() == 0, "should be dead now");
2101 }
2102 ResourceMark rm;
2103 Node_List casts;
2104 for (uint i = 1; i < req(); ++i) {
2105 Node* n = in(i);
2106 while (n->is_ConstraintCast()) {
2107 casts.push(n);
2108 n = n->in(1);
2109 }
2110 if (phase->type(n)->is_zero_type()) {
2111 n = InlineTypeNode::make_null(*phase, inline_klass);
2112 } else if (n->is_Phi()) {
2113 assert(can_reshape, "can only handle phis during IGVN");
2114 n = phase->transform(n->as_Phi()->push_inline_types_down(phase, can_reshape, inline_klass));
2115 }
2116 while (casts.size() != 0) {
2117 // Push the cast(s) through the InlineTypeNode
2118 // TODO 8302217 Can we avoid cloning? See InlineTypeNode::clone_if_required
2119 Node* cast = casts.pop()->clone();
2120 cast->set_req_X(1, n->as_InlineType()->get_oop(), phase);
2121 n = n->clone();
2122 n->as_InlineType()->set_oop(*phase, phase->transform(cast));
2123 n = phase->transform(n);
2124 if (n->is_top()) {
2125 break;
2126 }
2127 }
2128 bool transform = !can_reshape && (i == (req()-1)); // Transform phis on last merge
2129 assert(n->is_top() || n->is_InlineType(), "Only InlineType or top at this point.");
2130 if (n->is_InlineType()) {
2131 vt->merge_with(phase, n->as_InlineType(), i, transform);
2132 } // else nothing to do: phis above vt created by clone_with_phis are initialized to top already.
2133 }
2134 return vt;
2135 }
2136
2137 // If the Phi's Region is in an irreducible loop, and the Region
2138 // has had an input removed, but not yet transformed, it could be
2139 // that the Region (and this Phi) are not reachable from Root.
2140 // If we allow the Phi to collapse before the Region, this may lead
2141 // to dead-loop data. Wait for the Region to check for reachability,
2142 // and potentially remove the dead code.
2143 bool PhiNode::must_wait_for_region_in_irreducible_loop(PhaseGVN* phase) const {
2144 RegionNode* region = in(0)->as_Region();
2145 if (region->loop_status() == RegionNode::LoopStatus::MaybeIrreducibleEntry) {
2146 Node* top = phase->C->top();
2147 for (uint j = 1; j < req(); j++) {
2148 Node* rc = region->in(j); // for each control input
2149 if (rc == nullptr || phase->type(rc) == Type::TOP) {
2150 // Region is missing a control input
2151 Node* n = in(j);
2152 if (n != nullptr && n != top) {
2153 // Phi still has its input, so region just lost its input
2154 return true;
2155 }
2156 }
2157 }
2158 }
2159 return false;
2160 }
2161
2162 // Check if splitting a bot memory Phi through a parent MergeMem may lead to
2163 // non-termination. For more details, see comments at the call site in
2164 // PhiNode::Ideal.
2165 bool PhiNode::is_split_through_mergemem_terminating() const {
2166 ResourceMark rm;
2167 VectorSet visited;
2168 GrowableArray<const Node*> worklist;
2169 worklist.push(this);
2170 visited.set(this->_idx);
2171 auto maybe_add_to_worklist = [&](Node* input) {
2172 if (input != nullptr &&
2173 (input->is_MergeMem() || input->is_memory_phi()) &&
2174 !visited.test_set(input->_idx)) {
2175 worklist.push(input);
2176 assert(input->adr_type() == TypePtr::BOTTOM,
2177 "should only visit bottom memory");
2178 }
2179 };
2180 while (worklist.length() > 0) {
2181 const Node* n = worklist.pop();
2182 if (n->is_MergeMem()) {
2183 Node* input = n->as_MergeMem()->base_memory();
2184 if (input == this) {
2185 return false;
2186 }
2187 maybe_add_to_worklist(input);
2188 } else {
2189 assert(n->is_memory_phi(), "invariant");
2190 for (uint i = PhiNode::Input; i < n->req(); i++) {
2191 Node* input = n->in(i);
2192 if (input == this) {
2193 return false;
2194 }
2195 maybe_add_to_worklist(input);
2196 }
2197 }
2198 }
2199 return true;
2200 }
2201
2202 //------------------------------Ideal------------------------------------------
2203 // Return a node which is more "ideal" than the current node. Must preserve
2204 // the CFG, but we can still strip out dead paths.
2205 Node *PhiNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2206 Node *r = in(0); // RegionNode
2207 assert(r != nullptr && r->is_Region(), "this phi must have a region");
2208 assert(r->in(0) == nullptr || !r->in(0)->is_Root(), "not a specially hidden merge");
2209
2210 // Note: During parsing, phis are often transformed before their regions.
2211 // This means we have to use type_or_null to defend against untyped regions.
2212 if( phase->type_or_null(r) == Type::TOP ) // Dead code?
2213 return nullptr; // No change
2214
2215 Node *top = phase->C->top();
2216 bool new_phi = (outcnt() == 0); // transforming new Phi
2217 // No change for igvn if new phi is not hooked
2218 if (new_phi && can_reshape)
2219 return nullptr;
2220
2221 if (must_wait_for_region_in_irreducible_loop(phase)) {
2222 return nullptr;
2223 }
2224
2225 // The are 2 situations when only one valid phi's input is left
2226 // (in addition to Region input).
2227 // One: region is not loop - replace phi with this input.
2228 // Two: region is loop - replace phi with top since this data path is dead
2229 // and we need to break the dead data loop.
2230 Node* progress = nullptr; // Record if any progress made
2231 for( uint j = 1; j < req(); ++j ){ // For all paths in
2232 // Check unreachable control paths
2233 Node* rc = r->in(j);
2234 Node* n = in(j); // Get the input
2235 if (rc == nullptr || phase->type(rc) == Type::TOP) {
2236 if (n != top) { // Not already top?
2237 PhaseIterGVN *igvn = phase->is_IterGVN();
2238 if (can_reshape && igvn != nullptr) {
2239 igvn->_worklist.push(r);
2240 }
2241 // Nuke it down
2242 set_req_X(j, top, phase);
2243 progress = this; // Record progress
2244 }
2245 }
2246 }
2247
2248 if (can_reshape && outcnt() == 0) {
2249 // set_req() above may kill outputs if Phi is referenced
2250 // only by itself on the dead (top) control path.
2251 return top;
2252 }
2253
2254 bool uncasted = false;
2255 Node* uin = unique_input(phase, false);
2256 if (uin == nullptr && can_reshape &&
2257 // If there is a chance that the region can be optimized out do
2258 // not add a cast node that we can't remove yet.
2259 !wait_for_region_igvn(phase)) {
2260 uncasted = true;
2261 uin = unique_input(phase, true);
2262 }
2263 if (uin == top) { // Simplest case: no alive inputs.
2264 if (can_reshape) // IGVN transformation
2265 return top;
2266 else
2267 return nullptr; // Identity will return TOP
2268 } else if (uin != nullptr) {
2269 // Only one not-null unique input path is left.
2270 // Determine if this input is backedge of a loop.
2271 // (Skip new phis which have no uses and dead regions).
2272 if (outcnt() > 0 && r->in(0) != nullptr) {
2273 if (is_data_loop(r->as_Region(), uin, phase)) {
2274 // Break this data loop to avoid creation of a dead loop.
2275 if (can_reshape) {
2276 return top;
2277 } else {
2278 // We can't return top if we are in Parse phase - cut inputs only
2279 // let Identity to handle the case.
2280 replace_edge(uin, top, phase);
2281 return nullptr;
2282 }
2283 }
2284 }
2285
2286 if (uncasted) {
2287 // Add cast nodes between the phi to be removed and its unique input.
2288 // Wait until after parsing for the type information to propagate from the casts.
2289 assert(can_reshape, "Invalid during parsing");
2290 const Type* phi_type = bottom_type();
2291 // Add casts to carry the control dependency of the Phi that is
2292 // going away
2293 Node* cast = nullptr;
2294 const TypeTuple* extra_types = collect_types(phase);
2295 if (phi_type->isa_ptr()) {
2296 const Type* uin_type = phase->type(uin);
2297 if (!phi_type->isa_oopptr() && !uin_type->isa_oopptr()) {
2298 cast = new CastPPNode(r, uin, phi_type, ConstraintCastNode::StrongDependency, extra_types);
2299 } else {
2300 // Use a CastPP for a cast to not null and a CheckCastPP for
2301 // a cast to a new klass (and both if both null-ness and
2302 // klass change).
2303
2304 // If the type of phi is not null but the type of uin may be
2305 // null, uin's type must be casted to not null
2306 if (phi_type->join(TypePtr::NOTNULL) == phi_type->remove_speculative() &&
2307 uin_type->join(TypePtr::NOTNULL) != uin_type->remove_speculative()) {
2308 cast = new CastPPNode(r, uin, TypePtr::NOTNULL, ConstraintCastNode::StrongDependency, extra_types);
2309 }
2310
2311 // If the type of phi and uin, both casted to not null,
2312 // differ the klass of uin must be (check)cast'ed to match
2313 // that of phi
2314 if (phi_type->join_speculative(TypePtr::NOTNULL) != uin_type->join_speculative(TypePtr::NOTNULL)) {
2315 Node* n = uin;
2316 if (cast != nullptr) {
2317 cast = phase->transform(cast);
2318 n = cast;
2319 }
2320 cast = new CheckCastPPNode(r, n, phi_type, ConstraintCastNode::StrongDependency, extra_types);
2321 }
2322 if (cast == nullptr) {
2323 cast = new CastPPNode(r, uin, phi_type, ConstraintCastNode::StrongDependency, extra_types);
2324 }
2325 }
2326 } else {
2327 cast = ConstraintCastNode::make_cast_for_type(r, uin, phi_type, ConstraintCastNode::StrongDependency, extra_types);
2328 }
2329 assert(cast != nullptr, "cast should be set");
2330 cast = phase->transform(cast);
2331 // set all inputs to the new cast(s) so the Phi is removed by Identity
2332 PhaseIterGVN* igvn = phase->is_IterGVN();
2333 for (uint i = 1; i < req(); i++) {
2334 set_req_X(i, cast, igvn);
2335 }
2336 uin = cast;
2337 }
2338
2339 // One unique input.
2340 DEBUG_ONLY(Node* ident = Identity(phase));
2341 // The unique input must eventually be detected by the Identity call.
2342 #ifdef ASSERT
2343 if (ident != uin && !ident->is_top() && !must_wait_for_region_in_irreducible_loop(phase)) {
2344 // print this output before failing assert
2345 r->dump(3);
2346 this->dump(3);
2347 ident->dump();
2348 uin->dump();
2349 }
2350 #endif
2351 // Identity may not return the expected uin, if it has to wait for the region, in irreducible case
2352 assert(ident == uin || ident->is_top() || must_wait_for_region_in_irreducible_loop(phase), "Identity must clean this up");
2353 return nullptr;
2354 }
2355
2356 Node* opt = nullptr;
2357 int true_path = is_diamond_phi();
2358 if (true_path != 0 &&
2359 // If one of the diamond's branch is in the process of dying then, the Phi's input for that branch might transform
2360 // to top. If that happens replacing the Phi with an operation that consumes the Phi's inputs will cause the Phi
2361 // to be replaced by top. To prevent that, delay the transformation until the branch has a chance to be removed.
2362 !(can_reshape && wait_for_region_igvn(phase))) {
2363 // Check for CMove'ing identity. If it would be unsafe,
2364 // handle it here. In the safe case, let Identity handle it.
2365 Node* unsafe_id = is_cmove_id(phase, true_path);
2366 if( unsafe_id != nullptr && is_unsafe_data_reference(unsafe_id) )
2367 opt = unsafe_id;
2368
2369 // Check for simple convert-to-boolean pattern
2370 if( opt == nullptr )
2371 opt = is_x2logic(phase, this, true_path);
2372
2373 // Check for absolute value
2374 if( opt == nullptr )
2375 opt = is_absolute(phase, this, true_path);
2376
2377 // Check for conditional add
2378 if( opt == nullptr && can_reshape )
2379 opt = is_cond_add(phase, this, true_path);
2380
2381 // These 4 optimizations could subsume the phi:
2382 // have to check for a dead data loop creation.
2383 if( opt != nullptr ) {
2384 if( opt == unsafe_id || is_unsafe_data_reference(opt) ) {
2385 // Found dead loop.
2386 if( can_reshape )
2387 return top;
2388 // We can't return top if we are in Parse phase - cut inputs only
2389 // to stop further optimizations for this phi. Identity will return TOP.
2390 assert(req() == 3, "only diamond merge phi here");
2391 set_req(1, top);
2392 set_req(2, top);
2393 return nullptr;
2394 } else {
2395 return opt;
2396 }
2397 }
2398 }
2399
2400 // Check for merging identical values and split flow paths
2401 if (can_reshape) {
2402 opt = split_flow_path(phase, this);
2403 // This optimization only modifies phi - don't need to check for dead loop.
2404 assert(opt == nullptr || opt == this, "do not elide phi");
2405 if (opt != nullptr) return opt;
2406 }
2407
2408 if (in(1) != nullptr && in(1)->Opcode() == Op_AddP && can_reshape) {
2409 // Try to undo Phi of AddP:
2410 // (Phi (AddP base address offset) (AddP base2 address2 offset2))
2411 // becomes:
2412 // newbase := (Phi base base2)
2413 // newaddress := (Phi address address2)
2414 // newoffset := (Phi offset offset2)
2415 // (AddP newbase newaddress newoffset)
2416 //
2417 // This occurs as a result of unsuccessful split_thru_phi and
2418 // interferes with taking advantage of addressing modes. See the
2419 // clone_shift_expressions code in matcher.cpp
2420 Node* addp = in(1);
2421 Node* base = addp->in(AddPNode::Base);
2422 Node* address = addp->in(AddPNode::Address);
2423 Node* offset = addp->in(AddPNode::Offset);
2424 if (base != nullptr && address != nullptr && offset != nullptr &&
2425 !base->is_top() && !address->is_top() && !offset->is_top()) {
2426 const Type* base_type = base->bottom_type();
2427 const Type* address_type = address->bottom_type();
2428 // make sure that all the inputs are similar to the first one,
2429 // i.e. AddP with base == address and same offset as first AddP
2430 bool doit = true;
2431 for (uint i = 2; i < req(); i++) {
2432 if (in(i) == nullptr ||
2433 in(i)->Opcode() != Op_AddP ||
2434 in(i)->in(AddPNode::Base) == nullptr ||
2435 in(i)->in(AddPNode::Address) == nullptr ||
2436 in(i)->in(AddPNode::Offset) == nullptr ||
2437 in(i)->in(AddPNode::Base)->is_top() ||
2438 in(i)->in(AddPNode::Address)->is_top() ||
2439 in(i)->in(AddPNode::Offset)->is_top()) {
2440 doit = false;
2441 break;
2442 }
2443 if (in(i)->in(AddPNode::Base) != base) {
2444 base = nullptr;
2445 }
2446 if (in(i)->in(AddPNode::Offset) != offset) {
2447 offset = nullptr;
2448 }
2449 if (in(i)->in(AddPNode::Address) != address) {
2450 address = nullptr;
2451 }
2452 // Accumulate type for resulting Phi
2453 base_type = base_type->meet_speculative(in(i)->in(AddPNode::Base)->bottom_type());
2454 address_type = address_type->meet_speculative(in(i)->in(AddPNode::Address)->bottom_type());
2455 }
2456 if (doit && base == nullptr) {
2457 // Check for neighboring AddP nodes in a tree.
2458 // If they have a base, use that it.
2459 for (DUIterator_Fast kmax, k = this->fast_outs(kmax); k < kmax; k++) {
2460 Node* u = this->fast_out(k);
2461 if (u->is_AddP()) {
2462 Node* base2 = u->in(AddPNode::Base);
2463 if (base2 != nullptr && !base2->is_top()) {
2464 if (base == nullptr)
2465 base = base2;
2466 else if (base != base2)
2467 { doit = false; break; }
2468 }
2469 }
2470 }
2471 }
2472 if (doit) {
2473 if (base == nullptr) {
2474 base = new PhiNode(in(0), base_type, nullptr);
2475 for (uint i = 1; i < req(); i++) {
2476 base->init_req(i, in(i)->in(AddPNode::Base));
2477 }
2478 phase->is_IterGVN()->register_new_node_with_optimizer(base);
2479 }
2480 if (address == nullptr) {
2481 address = new PhiNode(in(0), address_type, nullptr);
2482 for (uint i = 1; i < req(); i++) {
2483 address->init_req(i, in(i)->in(AddPNode::Address));
2484 }
2485 phase->is_IterGVN()->register_new_node_with_optimizer(address);
2486 }
2487 if (offset == nullptr) {
2488 offset = new PhiNode(in(0), TypeX_X, nullptr);
2489 for (uint i = 1; i < req(); i++) {
2490 offset->init_req(i, in(i)->in(AddPNode::Offset));
2491 }
2492 phase->is_IterGVN()->register_new_node_with_optimizer(offset);
2493 }
2494 return new AddPNode(base, address, offset);
2495 }
2496 }
2497 }
2498
2499 // Split phis through memory merges, so that the memory merges will go away.
2500 // Piggy-back this transformation on the search for a unique input....
2501 // It will be as if the merged memory is the unique value of the phi.
2502 // (Do not attempt this optimization unless parsing is complete.
2503 // It would make the parser's memory-merge logic sick.)
2504 // (MergeMemNode is not dead_loop_safe - need to check for dead loop.)
2505 if (progress == nullptr && can_reshape && type() == Type::MEMORY) {
2506
2507 // See if this Phi should be sliced. Determine the merge width of input
2508 // MergeMems and check if there is a direct loop to self, as illustrated
2509 // below.
2510 //
2511 // +-------------+
2512 // | |
2513 // (base_memory) v |
2514 // MergeMem |
2515 // | |
2516 // v |
2517 // Phi (this) |
2518 // | |
2519 // +-----------+
2520 //
2521 // Generally, there are issues with non-termination with such circularity
2522 // (see comment further below). However, if there is a direct loop to self,
2523 // splitting the Phi through the MergeMem will result in the below.
2524 //
2525 // +---+
2526 // | |
2527 // v |
2528 // Phi |
2529 // |\ |
2530 // | +-+
2531 // (base_memory) v
2532 // MergeMem
2533 //
2534 // This split breaks the circularity and consequently does not lead to
2535 // non-termination.
2536 uint merge_width = 0;
2537 // TODO revisit this with JDK-8247216
2538 bool mergemem_only = true;
2539 bool split_always_terminates = false; // Is splitting guaranteed to terminate?
2540 for( uint i=1; i<req(); ++i ) {// For all paths in
2541 Node *ii = in(i);
2542 // TOP inputs should not be counted as safe inputs because if the
2543 // Phi references itself through all other inputs then splitting the
2544 // Phi through memory merges would create dead loop at later stage.
2545 if (ii == top) {
2546 return nullptr; // Delay optimization until graph is cleaned.
2547 }
2548 if (ii->is_MergeMem()) {
2549 MergeMemNode* n = ii->as_MergeMem();
2550 merge_width = MAX2(merge_width, n->req());
2551 if (n->base_memory() == this) {
2552 split_always_terminates = true;
2553 }
2554 } else {
2555 mergemem_only = false;
2556 }
2557 }
2558
2559 // There are cases with circular dependencies between bottom Phis
2560 // and MergeMems. Below is a minimal example.
2561 //
2562 // +------------+
2563 // | |
2564 // (base_memory) v |
2565 // MergeMem |
2566 // | |
2567 // v |
2568 // Phi (this) |
2569 // | |
2570 // v |
2571 // Phi |
2572 // | |
2573 // +----------+
2574 //
2575 // Here, we cannot break the circularity through a self-loop as there
2576 // are two Phis involved. Repeatedly splitting the Phis through the
2577 // MergeMem leads to non-termination. We check for non-termination below.
2578 // Only check for non-termination if necessary.
2579 if (!mergemem_only && !split_always_terminates && adr_type() == TypePtr::BOTTOM &&
2580 merge_width > Compile::AliasIdxRaw) {
2581 split_always_terminates = is_split_through_mergemem_terminating();
2582 }
2583
2584 if (merge_width > Compile::AliasIdxRaw) {
2585 // found at least one non-empty MergeMem
2586 const TypePtr* at = adr_type();
2587 if (at != TypePtr::BOTTOM) {
2588 // Patch the existing phi to select an input from the merge:
2589 // Phi:AT1(...MergeMem(m0, m1, m2)...) into
2590 // Phi:AT1(...m1...)
2591 int alias_idx = phase->C->get_alias_index(at);
2592 for (uint i=1; i<req(); ++i) {
2593 Node *ii = in(i);
2594 if (ii->is_MergeMem()) {
2595 MergeMemNode* n = ii->as_MergeMem();
2596 // compress paths and change unreachable cycles to TOP
2597 // If not, we can update the input infinitely along a MergeMem cycle
2598 // Equivalent code is in MemNode::Ideal_common
2599 Node *m = phase->transform(n);
2600 if (outcnt() == 0) { // Above transform() may kill us!
2601 return top;
2602 }
2603 // If transformed to a MergeMem, get the desired slice
2604 // Otherwise the returned node represents memory for every slice
2605 Node *new_mem = (m->is_MergeMem()) ?
2606 m->as_MergeMem()->memory_at(alias_idx) : m;
2607 // Update input if it is progress over what we have now
2608 if (new_mem != ii) {
2609 set_req_X(i, new_mem, phase->is_IterGVN());
2610 progress = this;
2611 }
2612 }
2613 }
2614 } else if (mergemem_only || split_always_terminates) {
2615 // If all inputs reference this phi (directly or through data nodes) -
2616 // it is a dead loop.
2617 bool saw_safe_input = false;
2618 for (uint j = 1; j < req(); ++j) {
2619 Node* n = in(j);
2620 if (n->is_MergeMem()) {
2621 MergeMemNode* mm = n->as_MergeMem();
2622 if (mm->base_memory() == this || mm->base_memory() == mm->empty_memory()) {
2623 // Skip this input if it references back to this phi or if the memory path is dead
2624 continue;
2625 }
2626 }
2627 if (!is_unsafe_data_reference(n)) {
2628 saw_safe_input = true; // found safe input
2629 break;
2630 }
2631 }
2632 if (!saw_safe_input) {
2633 // There is a dead loop: All inputs are either dead or reference back to this phi
2634 return top;
2635 }
2636
2637 // Phi(...MergeMem(m0, m1:AT1, m2:AT2)...) into
2638 // MergeMem(Phi(...m0...), Phi:AT1(...m1...), Phi:AT2(...m2...))
2639 PhaseIterGVN* igvn = phase->is_IterGVN();
2640 assert(igvn != nullptr, "sanity check");
2641 PhiNode* new_base = (PhiNode*) clone();
2642 // Must eagerly register phis, since they participate in loops.
2643 igvn->register_new_node_with_optimizer(new_base);
2644
2645 MergeMemNode* result = MergeMemNode::make(new_base);
2646 for (uint i = 1; i < req(); ++i) {
2647 Node *ii = in(i);
2648 if (ii->is_MergeMem()) {
2649 MergeMemNode* n = ii->as_MergeMem();
2650 if (igvn) {
2651 // TODO revisit this with JDK-8247216
2652 // Put 'n' on the worklist because it might be modified by MergeMemStream::iteration_setup
2653 igvn->_worklist.push(n);
2654 }
2655 for (MergeMemStream mms(result, n); mms.next_non_empty2(); ) {
2656 // If we have not seen this slice yet, make a phi for it.
2657 bool made_new_phi = false;
2658 if (mms.is_empty()) {
2659 Node* new_phi = new_base->slice_memory(mms.adr_type(phase->C));
2660 made_new_phi = true;
2661 igvn->register_new_node_with_optimizer(new_phi);
2662 mms.set_memory(new_phi);
2663 }
2664 Node* phi = mms.memory();
2665 assert(made_new_phi || phi->in(i) == n, "replace the i-th merge by a slice");
2666 phi->set_req(i, mms.memory2());
2667 }
2668 }
2669 }
2670 // Distribute all self-loops.
2671 { // (Extra braces to hide mms.)
2672 for (MergeMemStream mms(result); mms.next_non_empty(); ) {
2673 Node* phi = mms.memory();
2674 for (uint i = 1; i < req(); ++i) {
2675 if (phi->in(i) == this) phi->set_req(i, phi);
2676 }
2677 }
2678 }
2679
2680 // We could immediately transform the new Phi nodes here, but that can
2681 // result in creating an excessive number of new nodes within a single
2682 // IGVN iteration. We have put the Phi nodes on the IGVN worklist, so
2683 // they are transformed later on in any case.
2684
2685 // Replace self with the result.
2686 return result;
2687 }
2688 }
2689 //
2690 // Other optimizations on the memory chain
2691 //
2692 const TypePtr* at = adr_type();
2693 for( uint i=1; i<req(); ++i ) {// For all paths in
2694 Node *ii = in(i);
2695 Node *new_in = MemNode::optimize_memory_chain(ii, at, nullptr, phase);
2696 if (ii != new_in ) {
2697 set_req(i, new_in);
2698 progress = this;
2699 }
2700 }
2701 }
2702
2703 #ifdef _LP64
2704 // Push DecodeN/DecodeNKlass down through phi.
2705 // The rest of phi graph will transform by split EncodeP node though phis up.
2706 if ((UseCompressedOops || UseCompressedClassPointers) && can_reshape && progress == nullptr) {
2707 bool may_push = true;
2708 bool has_decodeN = false;
2709 bool is_decodeN = false;
2710 for (uint i=1; i<req(); ++i) {// For all paths in
2711 Node *ii = in(i);
2712 if (ii->is_DecodeNarrowPtr() && ii->bottom_type() == bottom_type()) {
2713 // Do optimization if a non dead path exist.
2714 if (ii->in(1)->bottom_type() != Type::TOP) {
2715 has_decodeN = true;
2716 is_decodeN = ii->is_DecodeN();
2717 }
2718 } else if (!ii->is_Phi()) {
2719 may_push = false;
2720 }
2721 }
2722
2723 if (has_decodeN && may_push) {
2724 PhaseIterGVN *igvn = phase->is_IterGVN();
2725 // Make narrow type for new phi.
2726 const Type* narrow_t;
2727 if (is_decodeN) {
2728 narrow_t = TypeNarrowOop::make(this->bottom_type()->is_ptr());
2729 } else {
2730 narrow_t = TypeNarrowKlass::make(this->bottom_type()->is_ptr());
2731 }
2732 PhiNode* new_phi = new PhiNode(r, narrow_t);
2733 uint orig_cnt = req();
2734 for (uint i=1; i<req(); ++i) {// For all paths in
2735 Node *ii = in(i);
2736 Node* new_ii = nullptr;
2737 if (ii->is_DecodeNarrowPtr()) {
2738 assert(ii->bottom_type() == bottom_type(), "sanity");
2739 new_ii = ii->in(1);
2740 } else {
2741 assert(ii->is_Phi(), "sanity");
2742 if (ii->as_Phi() == this) {
2743 new_ii = new_phi;
2744 } else {
2745 if (is_decodeN) {
2746 new_ii = new EncodePNode(ii, narrow_t);
2747 } else {
2748 new_ii = new EncodePKlassNode(ii, narrow_t);
2749 }
2750 igvn->register_new_node_with_optimizer(new_ii);
2751 }
2752 }
2753 new_phi->set_req(i, new_ii);
2754 }
2755 igvn->register_new_node_with_optimizer(new_phi, this);
2756 if (is_decodeN) {
2757 progress = new DecodeNNode(new_phi, bottom_type());
2758 } else {
2759 progress = new DecodeNKlassNode(new_phi, bottom_type());
2760 }
2761 }
2762 }
2763 #endif
2764
2765 Node* inline_type = try_push_inline_types_down(phase, can_reshape);
2766 if (inline_type != this) {
2767 return inline_type;
2768 }
2769
2770 // Try to convert a Phi with two duplicated convert nodes into a phi of the pre-conversion type and the convert node
2771 // proceeding the phi, to de-duplicate the convert node and compact the IR.
2772 if (can_reshape && progress == nullptr) {
2773 ConvertNode* convert = in(1)->isa_Convert();
2774 if (convert != nullptr) {
2775 int conv_op = convert->Opcode();
2776 bool ok = true;
2777
2778 // Check the rest of the inputs
2779 for (uint i = 2; i < req(); i++) {
2780 // Make sure that all inputs are of the same type of convert node
2781 if (in(i)->Opcode() != conv_op) {
2782 ok = false;
2783 break;
2784 }
2785 }
2786
2787 if (ok) {
2788 // Find the local bottom type to set as the type of the phi
2789 const Type* source_type = Type::get_const_basic_type(convert->in_type()->basic_type());
2790 const Type* dest_type = convert->bottom_type();
2791
2792 PhiNode* newphi = new PhiNode(in(0), source_type, nullptr);
2793 // Set inputs to the new phi be the inputs of the convert
2794 for (uint i = 1; i < req(); i++) {
2795 newphi->init_req(i, in(i)->in(1));
2796 }
2797
2798 phase->is_IterGVN()->register_new_node_with_optimizer(newphi, this);
2799
2800 return ConvertNode::create_convert(get_convert_type(convert, source_type), get_convert_type(convert, dest_type), newphi);
2801 }
2802 }
2803 }
2804
2805 // Phi (VB ... VB) => VB (Phi ...) (Phi ...)
2806 if (EnableVectorReboxing && can_reshape && progress == nullptr && type()->isa_oopptr()) {
2807 progress = merge_through_phi(this, phase->is_IterGVN());
2808 }
2809
2810 return progress; // Return any progress
2811 }
2812
2813 // Check recursively if inputs are either an inline type, constant null
2814 // or another Phi (including self references through data loops). If so,
2815 // push the inline types down through the phis to enable folding of loads.
2816 Node* PhiNode::try_push_inline_types_down(PhaseGVN* phase, const bool can_reshape) {
2817 if (!can_be_inline_type()) {
2818 return this;
2819 }
2820
2821 ciInlineKlass* inline_klass;
2822 if (can_push_inline_types_down(phase, can_reshape, inline_klass)) {
2823 assert(inline_klass != nullptr, "must be");
2824 return push_inline_types_down(phase, can_reshape, inline_klass);
2825 }
2826 return this;
2827 }
2828
2829 bool PhiNode::can_push_inline_types_down(PhaseGVN* phase, const bool can_reshape, ciInlineKlass*& inline_klass) {
2830 if (req() <= 2) {
2831 // Dead phi.
2832 return false;
2833 }
2834 inline_klass = nullptr;
2835
2836 // TODO 8302217 We need to prevent endless pushing through
2837 bool only_phi = (outcnt() != 0);
2838 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2839 Node* n = fast_out(i);
2840 if (n->is_InlineType() && n->in(1) == this) {
2841 return false;
2842 }
2843 if (!n->is_Phi()) {
2844 only_phi = false;
2845 }
2846 }
2847 if (only_phi) {
2848 return false;
2849 }
2850
2851 ResourceMark rm;
2852 Unique_Node_List worklist;
2853 worklist.push(this);
2854 Node_List casts;
2855
2856 for (uint next = 0; next < worklist.size(); next++) {
2857 Node* phi = worklist.at(next);
2858 for (uint i = 1; i < phi->req(); i++) {
2859 Node* n = phi->in(i);
2860 if (n == nullptr) {
2861 return false;
2862 }
2863 while (n->is_ConstraintCast()) {
2864 if (n->in(0) != nullptr && n->in(0)->is_top()) {
2865 // Will die, don't optimize
2866 return false;
2867 }
2868 casts.push(n);
2869 n = n->in(1);
2870 }
2871 const Type* type = phase->type(n);
2872 if (n->is_InlineType() && (inline_klass == nullptr || inline_klass == type->inline_klass())) {
2873 inline_klass = type->inline_klass();
2874 } else if (n->is_Phi() && can_reshape && n->bottom_type()->isa_ptr()) {
2875 worklist.push(n);
2876 } else if (!type->is_zero_type()) {
2877 return false;
2878 }
2879 }
2880 }
2881 if (inline_klass == nullptr) {
2882 return false;
2883 }
2884
2885 // Check if cast nodes can be pushed through
2886 const Type* t = Type::get_const_type(inline_klass);
2887 while (casts.size() != 0 && t != nullptr) {
2888 Node* cast = casts.pop();
2889 if (t->filter(cast->bottom_type()) == Type::TOP) {
2890 return false;
2891 }
2892 }
2893
2894 return true;
2895 }
2896
2897 #ifdef ASSERT
2898 bool PhiNode::can_push_inline_types_down(PhaseGVN* phase) {
2899 if (!can_be_inline_type()) {
2900 return false;
2901 }
2902
2903 ciInlineKlass* inline_klass;
2904 return can_push_inline_types_down(phase, true, inline_klass);
2905 }
2906 #endif // ASSERT
2907
2908 static int compare_types(const Type* const& e1, const Type* const& e2) {
2909 return (intptr_t)e1 - (intptr_t)e2;
2910 }
2911
2912 // Collect types at casts that are going to be eliminated at that Phi and store them in a TypeTuple.
2913 // Sort the types using an arbitrary order so a list of some types always hashes to the same TypeTuple (and TypeTuple
2914 // pointer comparison is enough to tell if 2 list of types are the same or not)
2915 const TypeTuple* PhiNode::collect_types(PhaseGVN* phase) const {
2916 const Node* region = in(0);
2917 const Type* phi_type = bottom_type();
2918 ResourceMark rm;
2919 GrowableArray<const Type*> types;
2920 for (uint i = 1; i < req(); i++) {
2921 if (region->in(i) == nullptr || phase->type(region->in(i)) == Type::TOP) {
2922 continue;
2923 }
2924 Node* in = Node::in(i);
2925 const Type* t = phase->type(in);
2926 if (in == nullptr || in == this || t == Type::TOP) {
2927 continue;
2928 }
2929 if (t != phi_type && t->higher_equal_speculative(phi_type)) {
2930 types.insert_sorted<compare_types>(t);
2931 }
2932 while (in != nullptr && in->is_ConstraintCast()) {
2933 Node* next = in->in(1);
2934 if (phase->type(next)->isa_rawptr() && phase->type(in)->isa_oopptr()) {
2935 break;
2936 }
2937 ConstraintCastNode* cast = in->as_ConstraintCast();
2938 for (int j = 0; j < cast->extra_types_count(); ++j) {
2939 const Type* extra_t = cast->extra_type_at(j);
2940 if (extra_t != phi_type && extra_t->higher_equal_speculative(phi_type)) {
2941 types.insert_sorted<compare_types>(extra_t);
2942 }
2943 }
2944 in = next;
2945 }
2946 }
2947 const Type **flds = (const Type **)(phase->C->type_arena()->AmallocWords(types.length()*sizeof(Type*)));
2948 for (int i = 0; i < types.length(); ++i) {
2949 flds[i] = types.at(i);
2950 }
2951 return TypeTuple::make(types.length(), flds);
2952 }
2953
2954 Node* PhiNode::clone_through_phi(Node* root_phi, const Type* t, uint c, PhaseIterGVN* igvn) {
2955 Node_Stack stack(1);
2956 VectorSet visited;
2957 Node_List node_map;
2958
2959 stack.push(root_phi, 1); // ignore control
2960 visited.set(root_phi->_idx);
2961
2962 Node* new_phi = new PhiNode(root_phi->in(0), t);
2963 node_map.map(root_phi->_idx, new_phi);
2964
2965 while (stack.is_nonempty()) {
2966 Node* n = stack.node();
2967 uint idx = stack.index();
2968 assert(n->is_Phi(), "not a phi");
2969 if (idx < n->req()) {
2970 stack.set_index(idx + 1);
2971 Node* def = n->in(idx);
2972 if (def == nullptr) {
2973 continue; // ignore dead path
2974 } else if (def->is_Phi()) { // inner node
2975 Node* new_phi = node_map[n->_idx];
2976 if (!visited.test_set(def->_idx)) { // not visited yet
2977 node_map.map(def->_idx, new PhiNode(def->in(0), t));
2978 stack.push(def, 1); // ignore control
2979 }
2980 Node* new_in = node_map[def->_idx];
2981 new_phi->set_req(idx, new_in);
2982 } else if (def->Opcode() == Op_VectorBox) { // leaf
2983 assert(n->is_Phi(), "not a phi");
2984 Node* new_phi = node_map[n->_idx];
2985 new_phi->set_req(idx, def->in(c));
2986 } else {
2987 assert(false, "not optimizeable");
2988 return nullptr;
2989 }
2990 } else {
2991 Node* new_phi = node_map[n->_idx];
2992 igvn->register_new_node_with_optimizer(new_phi, n);
2993 stack.pop();
2994 }
2995 }
2996 return new_phi;
2997 }
2998
2999 Node* PhiNode::merge_through_phi(Node* root_phi, PhaseIterGVN* igvn) {
3000 Node_Stack stack(1);
3001 VectorSet visited;
3002
3003 stack.push(root_phi, 1); // ignore control
3004 visited.set(root_phi->_idx);
3005
3006 VectorBoxNode* cached_vbox = nullptr;
3007 while (stack.is_nonempty()) {
3008 Node* n = stack.node();
3009 uint idx = stack.index();
3010 if (idx < n->req()) {
3011 stack.set_index(idx + 1);
3012 Node* in = n->in(idx);
3013 if (in == nullptr) {
3014 continue; // ignore dead path
3015 } else if (in->isa_Phi()) {
3016 if (!visited.test_set(in->_idx)) {
3017 stack.push(in, 1); // ignore control
3018 }
3019 } else if (in->Opcode() == Op_VectorBox) {
3020 VectorBoxNode* vbox = static_cast<VectorBoxNode*>(in);
3021 if (cached_vbox == nullptr) {
3022 cached_vbox = vbox;
3023 } else if (vbox->vec_type() != cached_vbox->vec_type()) {
3024 // TODO: vector type mismatch can be handled with additional reinterpret casts
3025 assert(!Type::equals(vbox->vec_type(), cached_vbox->vec_type()), "inconsistent");
3026 return nullptr; // not optimizable: vector type mismatch
3027 } else if (vbox->box_type() != cached_vbox->box_type()) {
3028 assert(!Type::equals(vbox->box_type(), cached_vbox->box_type()), "inconsistent");
3029 return nullptr; // not optimizable: box type mismatch
3030 }
3031 } else {
3032 return nullptr; // not optimizable: neither Phi nor VectorBox
3033 }
3034 } else {
3035 stack.pop();
3036 }
3037 }
3038 if (cached_vbox == nullptr) {
3039 // We have a Phi dead-loop (no data-input). Phi nodes are considered safe,
3040 // so just avoid this optimization.
3041 return nullptr;
3042 }
3043 const TypeInstPtr* btype = cached_vbox->box_type();
3044 const TypeVect* vtype = cached_vbox->vec_type();
3045 Node* new_vbox_phi = clone_through_phi(root_phi, btype, VectorBoxNode::Box, igvn);
3046 Node* new_vect_phi = clone_through_phi(root_phi, vtype, VectorBoxNode::Value, igvn);
3047 return new VectorBoxNode(igvn->C, new_vbox_phi, new_vect_phi, btype, vtype);
3048 }
3049
3050 bool PhiNode::is_data_loop(RegionNode* r, Node* uin, const PhaseGVN* phase) {
3051 // First, take the short cut when we know it is a loop and the EntryControl data path is dead.
3052 // The loop node may only have one input because the entry path was removed in PhaseIdealLoop::Dominators().
3053 // Then, check if there is a data loop when the phi references itself directly or through other data nodes.
3054 assert(!r->is_Loop() || r->req() <= 3, "Loop node should have 3 or less inputs");
3055 const bool is_loop = (r->is_Loop() && r->req() == 3);
3056 const Node* top = phase->C->top();
3057 if (is_loop) {
3058 return !uin->eqv_uncast(in(LoopNode::EntryControl));
3059 } else {
3060 // We have a data loop either with an unsafe data reference or if a region is unreachable.
3061 return is_unsafe_data_reference(uin)
3062 || (r->req() == 3 && (r->in(1) != top && r->in(2) == top && r->is_unreachable_region(phase)));
3063 }
3064 }
3065
3066 //------------------------------is_tripcount-----------------------------------
3067 bool PhiNode::is_tripcount(BasicType bt) const {
3068 return (in(0) != nullptr && in(0)->is_BaseCountedLoop() &&
3069 in(0)->as_BaseCountedLoop()->bt() == bt &&
3070 in(0)->as_BaseCountedLoop()->phi() == this);
3071 }
3072
3073 //------------------------------out_RegMask------------------------------------
3074 const RegMask &PhiNode::in_RegMask(uint i) const {
3075 return i ? out_RegMask() : RegMask::Empty;
3076 }
3077
3078 const RegMask &PhiNode::out_RegMask() const {
3079 uint ideal_reg = _type->ideal_reg();
3080 assert( ideal_reg != Node::NotAMachineReg, "invalid type at Phi" );
3081 if( ideal_reg == 0 ) return RegMask::Empty;
3082 assert(ideal_reg != Op_RegFlags, "flags register is not spillable");
3083 return *(Compile::current()->matcher()->idealreg2spillmask[ideal_reg]);
3084 }
3085
3086 #ifndef PRODUCT
3087 void PhiNode::dump_spec(outputStream *st) const {
3088 TypeNode::dump_spec(st);
3089 if (is_tripcount(T_INT) || is_tripcount(T_LONG)) {
3090 st->print(" #tripcount");
3091 }
3092 }
3093 #endif
3094
3095
3096 //=============================================================================
3097 const Type* GotoNode::Value(PhaseGVN* phase) const {
3098 // If the input is reachable, then we are executed.
3099 // If the input is not reachable, then we are not executed.
3100 return phase->type(in(0));
3101 }
3102
3103 Node* GotoNode::Identity(PhaseGVN* phase) {
3104 return in(0); // Simple copy of incoming control
3105 }
3106
3107 const RegMask &GotoNode::out_RegMask() const {
3108 return RegMask::Empty;
3109 }
3110
3111 //=============================================================================
3112 const RegMask &JumpNode::out_RegMask() const {
3113 return RegMask::Empty;
3114 }
3115
3116 //=============================================================================
3117 const RegMask &JProjNode::out_RegMask() const {
3118 return RegMask::Empty;
3119 }
3120
3121 //=============================================================================
3122 const RegMask &CProjNode::out_RegMask() const {
3123 return RegMask::Empty;
3124 }
3125
3126
3127
3128 //=============================================================================
3129
3130 uint PCTableNode::hash() const { return Node::hash() + _size; }
3131 bool PCTableNode::cmp( const Node &n ) const
3132 { return _size == ((PCTableNode&)n)._size; }
3133
3134 const Type *PCTableNode::bottom_type() const {
3135 const Type** f = TypeTuple::fields(_size);
3136 for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL;
3137 return TypeTuple::make(_size, f);
3138 }
3139
3140 //------------------------------Value------------------------------------------
3141 // Compute the type of the PCTableNode. If reachable it is a tuple of
3142 // Control, otherwise the table targets are not reachable
3143 const Type* PCTableNode::Value(PhaseGVN* phase) const {
3144 if( phase->type(in(0)) == Type::CONTROL )
3145 return bottom_type();
3146 return Type::TOP; // All paths dead? Then so are we
3147 }
3148
3149 //------------------------------Ideal------------------------------------------
3150 // Return a node which is more "ideal" than the current node. Strip out
3151 // control copies
3152 Node *PCTableNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3153 return remove_dead_region(phase, can_reshape) ? this : nullptr;
3154 }
3155
3156 //=============================================================================
3157 uint JumpProjNode::hash() const {
3158 return Node::hash() + _dest_bci;
3159 }
3160
3161 bool JumpProjNode::cmp( const Node &n ) const {
3162 return ProjNode::cmp(n) &&
3163 _dest_bci == ((JumpProjNode&)n)._dest_bci;
3164 }
3165
3166 #ifndef PRODUCT
3167 void JumpProjNode::dump_spec(outputStream *st) const {
3168 ProjNode::dump_spec(st);
3169 st->print("@bci %d ",_dest_bci);
3170 }
3171
3172 void JumpProjNode::dump_compact_spec(outputStream *st) const {
3173 ProjNode::dump_compact_spec(st);
3174 st->print("(%d)%d@%d", _switch_val, _proj_no, _dest_bci);
3175 }
3176 #endif
3177
3178 //=============================================================================
3179 //------------------------------Value------------------------------------------
3180 // Check for being unreachable, or for coming from a Rethrow. Rethrow's cannot
3181 // have the default "fall_through_index" path.
3182 const Type* CatchNode::Value(PhaseGVN* phase) const {
3183 // Unreachable? Then so are all paths from here.
3184 if( phase->type(in(0)) == Type::TOP ) return Type::TOP;
3185 // First assume all paths are reachable
3186 const Type** f = TypeTuple::fields(_size);
3187 for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL;
3188 // Identify cases that will always throw an exception
3189 // () rethrow call
3190 // () virtual or interface call with null receiver
3191 // () call is a check cast with incompatible arguments
3192 if( in(1)->is_Proj() ) {
3193 Node *i10 = in(1)->in(0);
3194 if( i10->is_Call() ) {
3195 CallNode *call = i10->as_Call();
3196 // Rethrows always throw exceptions, never return
3197 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3198 f[CatchProjNode::fall_through_index] = Type::TOP;
3199 } else if (call->is_AllocateArray()) {
3200 Node* klass_node = call->in(AllocateNode::KlassNode);
3201 Node* length = call->in(AllocateNode::ALength);
3202 const Type* length_type = phase->type(length);
3203 const Type* klass_type = phase->type(klass_node);
3204 Node* valid_length_test = call->in(AllocateNode::ValidLengthTest);
3205 const Type* valid_length_test_t = phase->type(valid_length_test);
3206 if (length_type == Type::TOP || klass_type == Type::TOP || valid_length_test_t == Type::TOP ||
3207 valid_length_test_t->is_int()->is_con(0)) {
3208 f[CatchProjNode::fall_through_index] = Type::TOP;
3209 }
3210 } else if( call->req() > TypeFunc::Parms ) {
3211 const Type *arg0 = phase->type( call->in(TypeFunc::Parms) );
3212 // Check for null receiver to virtual or interface calls
3213 if( call->is_CallDynamicJava() &&
3214 arg0->higher_equal(TypePtr::NULL_PTR) ) {
3215 f[CatchProjNode::fall_through_index] = Type::TOP;
3216 }
3217 } // End of if not a runtime stub
3218 } // End of if have call above me
3219 } // End of slot 1 is not a projection
3220 return TypeTuple::make(_size, f);
3221 }
3222
3223 //=============================================================================
3224 uint CatchProjNode::hash() const {
3225 return Node::hash() + _handler_bci;
3226 }
3227
3228
3229 bool CatchProjNode::cmp( const Node &n ) const {
3230 return ProjNode::cmp(n) &&
3231 _handler_bci == ((CatchProjNode&)n)._handler_bci;
3232 }
3233
3234
3235 //------------------------------Identity---------------------------------------
3236 // If only 1 target is possible, choose it if it is the main control
3237 Node* CatchProjNode::Identity(PhaseGVN* phase) {
3238 // If my value is control and no other value is, then treat as ID
3239 const TypeTuple *t = phase->type(in(0))->is_tuple();
3240 if (t->field_at(_con) != Type::CONTROL) return this;
3241 // If we remove the last CatchProj and elide the Catch/CatchProj, then we
3242 // also remove any exception table entry. Thus we must know the call
3243 // feeding the Catch will not really throw an exception. This is ok for
3244 // the main fall-thru control (happens when we know a call can never throw
3245 // an exception) or for "rethrow", because a further optimization will
3246 // yank the rethrow (happens when we inline a function that can throw an
3247 // exception and the caller has no handler). Not legal, e.g., for passing
3248 // a null receiver to a v-call, or passing bad types to a slow-check-cast.
3249 // These cases MUST throw an exception via the runtime system, so the VM
3250 // will be looking for a table entry.
3251 Node *proj = in(0)->in(1); // Expect a proj feeding CatchNode
3252 CallNode *call;
3253 if (_con != TypeFunc::Control && // Bail out if not the main control.
3254 !(proj->is_Proj() && // AND NOT a rethrow
3255 proj->in(0)->is_Call() &&
3256 (call = proj->in(0)->as_Call()) &&
3257 call->entry_point() == OptoRuntime::rethrow_stub()))
3258 return this;
3259
3260 // Search for any other path being control
3261 for (uint i = 0; i < t->cnt(); i++) {
3262 if (i != _con && t->field_at(i) == Type::CONTROL)
3263 return this;
3264 }
3265 // Only my path is possible; I am identity on control to the jump
3266 return in(0)->in(0);
3267 }
3268
3269
3270 #ifndef PRODUCT
3271 void CatchProjNode::dump_spec(outputStream *st) const {
3272 ProjNode::dump_spec(st);
3273 st->print("@bci %d ",_handler_bci);
3274 }
3275 #endif
3276
3277 //=============================================================================
3278 //------------------------------Identity---------------------------------------
3279 // Check for CreateEx being Identity.
3280 Node* CreateExNode::Identity(PhaseGVN* phase) {
3281 if( phase->type(in(1)) == Type::TOP ) return in(1);
3282 if( phase->type(in(0)) == Type::TOP ) return in(0);
3283 if (phase->type(in(0)->in(0)) == Type::TOP) {
3284 assert(in(0)->is_CatchProj(), "control is CatchProj");
3285 return phase->C->top(); // dead code
3286 }
3287 // We only come from CatchProj, unless the CatchProj goes away.
3288 // If the CatchProj is optimized away, then we just carry the
3289 // exception oop through.
3290
3291 // CheckCastPPNode::Ideal() for inline types reuses the exception
3292 // paths of a call to perform an allocation: we can see a Phi here.
3293 if (in(1)->is_Phi()) {
3294 return this;
3295 }
3296 CallNode *call = in(1)->in(0)->as_Call();
3297
3298 return (in(0)->is_CatchProj() && in(0)->in(0)->is_Catch() &&
3299 in(0)->in(0)->in(1) == in(1)) ? this : call->in(TypeFunc::Parms);
3300 }
3301
3302 //=============================================================================
3303 //------------------------------Value------------------------------------------
3304 // Check for being unreachable.
3305 const Type* NeverBranchNode::Value(PhaseGVN* phase) const {
3306 if (!in(0) || in(0)->is_top()) return Type::TOP;
3307 return bottom_type();
3308 }
3309
3310 //------------------------------Ideal------------------------------------------
3311 // Check for no longer being part of a loop
3312 Node *NeverBranchNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3313 if (can_reshape && !in(0)->is_Region()) {
3314 // Dead code elimination can sometimes delete this projection so
3315 // if it's not there, there's nothing to do.
3316 Node* fallthru = proj_out_or_null(0);
3317 if (fallthru != nullptr) {
3318 phase->is_IterGVN()->replace_node(fallthru, in(0));
3319 }
3320 return phase->C->top();
3321 }
3322 return nullptr;
3323 }
3324
3325 #ifndef PRODUCT
3326 void NeverBranchNode::format( PhaseRegAlloc *ra_, outputStream *st) const {
3327 st->print("%s", Name());
3328 }
3329 #endif
3330
3331 Node* BlackholeNode::Ideal(PhaseGVN* phase, bool can_reshape) {
3332 return remove_dead_region(phase, can_reshape) ? this : nullptr;
3333 }
3334
3335 #ifndef PRODUCT
3336 void BlackholeNode::format(PhaseRegAlloc* ra, outputStream* st) const {
3337 st->print("blackhole ");
3338 bool first = true;
3339 for (uint i = 0; i < req(); i++) {
3340 Node* n = in(i);
3341 if (n != nullptr && OptoReg::is_valid(ra->get_reg_first(n))) {
3342 if (first) {
3343 first = false;
3344 } else {
3345 st->print(", ");
3346 }
3347 char buf[128];
3348 ra->dump_register(n, buf, sizeof(buf));
3349 st->print("%s", buf);
3350 }
3351 }
3352 st->cr();
3353 }
3354 #endif
3355