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 "compiler/compileLog.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/oopMap.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "interpreter/interpreter.hpp"
31 #include "opto/callGenerator.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/escape.hpp"
36 #include "opto/locknode.hpp"
37 #include "opto/machnode.hpp"
38 #include "opto/matcher.hpp"
39 #include "opto/parse.hpp"
40 #include "opto/regalloc.hpp"
41 #include "opto/regmask.hpp"
42 #include "opto/rootnode.hpp"
43 #include "opto/runtime.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "utilities/powerOfTwo.hpp"
46 #include "code/vmreg.hpp"
47
48 // Portions of code courtesy of Clifford Click
49
50 // Optimization - Graph Style
51
52 //=============================================================================
53 uint StartNode::size_of() const { return sizeof(*this); }
54 bool StartNode::cmp( const Node &n ) const
55 { return _domain == ((StartNode&)n)._domain; }
56 const Type *StartNode::bottom_type() const { return _domain; }
57 const Type* StartNode::Value(PhaseGVN* phase) const { return _domain; }
58 #ifndef PRODUCT
59 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);}
60 void StartNode::dump_compact_spec(outputStream *st) const { /* empty */ }
61 #endif
62
63 //------------------------------Ideal------------------------------------------
64 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){
65 return remove_dead_region(phase, can_reshape) ? this : nullptr;
66 }
67
68 //------------------------------calling_convention-----------------------------
69 void StartNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
70 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt);
71 }
72
73 //------------------------------Registers--------------------------------------
74 const RegMask &StartNode::in_RegMask(uint) const {
75 return RegMask::Empty;
76 }
77
78 //------------------------------match------------------------------------------
79 // Construct projections for incoming parameters, and their RegMask info
80 Node *StartNode::match( const ProjNode *proj, const Matcher *match ) {
81 switch (proj->_con) {
82 case TypeFunc::Control:
83 case TypeFunc::I_O:
84 case TypeFunc::Memory:
85 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
86 case TypeFunc::FramePtr:
87 return new MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP);
88 case TypeFunc::ReturnAdr:
89 return new MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP);
90 case TypeFunc::Parms:
91 default: {
92 uint parm_num = proj->_con - TypeFunc::Parms;
93 const Type *t = _domain->field_at(proj->_con);
94 if (t->base() == Type::Half) // 2nd half of Longs and Doubles
95 return new ConNode(Type::TOP);
96 uint ideal_reg = t->ideal_reg();
97 RegMask &rm = match->_calling_convention_mask[parm_num];
98 return new MachProjNode(this,proj->_con,rm,ideal_reg);
99 }
100 }
101 return nullptr;
102 }
103
104 //------------------------------StartOSRNode----------------------------------
105 // The method start node for an on stack replacement adapter
106
107 //------------------------------osr_domain-----------------------------
108 const TypeTuple *StartOSRNode::osr_domain() {
109 const Type **fields = TypeTuple::fields(2);
110 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer
111
112 return TypeTuple::make(TypeFunc::Parms+1, fields);
113 }
114
115 //=============================================================================
116 const char * const ParmNode::names[TypeFunc::Parms+1] = {
117 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms"
118 };
119
120 #ifndef PRODUCT
121 void ParmNode::dump_spec(outputStream *st) const {
122 if( _con < TypeFunc::Parms ) {
123 st->print("%s", names[_con]);
124 } else {
125 st->print("Parm%d: ",_con-TypeFunc::Parms);
126 // Verbose and WizardMode dump bottom_type for all nodes
127 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st);
128 }
129 }
130
131 void ParmNode::dump_compact_spec(outputStream *st) const {
132 if (_con < TypeFunc::Parms) {
133 st->print("%s", names[_con]);
134 } else {
135 st->print("%d:", _con-TypeFunc::Parms);
136 // unconditionally dump bottom_type
137 bottom_type()->dump_on(st);
138 }
139 }
140 #endif
141
142 uint ParmNode::ideal_reg() const {
143 switch( _con ) {
144 case TypeFunc::Control : // fall through
145 case TypeFunc::I_O : // fall through
146 case TypeFunc::Memory : return 0;
147 case TypeFunc::FramePtr : // fall through
148 case TypeFunc::ReturnAdr: return Op_RegP;
149 default : assert( _con > TypeFunc::Parms, "" );
150 // fall through
151 case TypeFunc::Parms : {
152 // Type of argument being passed
153 const Type *t = in(0)->as_Start()->_domain->field_at(_con);
154 return t->ideal_reg();
155 }
156 }
157 ShouldNotReachHere();
158 return 0;
159 }
160
161 //=============================================================================
162 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) {
163 init_req(TypeFunc::Control,cntrl);
164 init_req(TypeFunc::I_O,i_o);
165 init_req(TypeFunc::Memory,memory);
166 init_req(TypeFunc::FramePtr,frameptr);
167 init_req(TypeFunc::ReturnAdr,retadr);
168 }
169
170 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){
171 return remove_dead_region(phase, can_reshape) ? this : nullptr;
172 }
173
174 const Type* ReturnNode::Value(PhaseGVN* phase) const {
175 return ( phase->type(in(TypeFunc::Control)) == Type::TOP)
176 ? Type::TOP
177 : Type::BOTTOM;
178 }
179
180 // Do we Match on this edge index or not? No edges on return nodes
181 uint ReturnNode::match_edge(uint idx) const {
182 return 0;
183 }
184
185
186 #ifndef PRODUCT
187 void ReturnNode::dump_req(outputStream *st, DumpConfig* dc) const {
188 // Dump the required inputs, after printing "returns"
189 uint i; // Exit value of loop
190 for (i = 0; i < req(); i++) { // For all required inputs
191 if (i == TypeFunc::Parms) st->print("returns ");
192 Node* p = in(i);
193 if (p != nullptr) {
194 p->dump_idx(false, st, dc);
195 st->print(" ");
196 } else {
197 st->print("_ ");
198 }
199 }
200 }
201 #endif
202
203 //=============================================================================
204 RethrowNode::RethrowNode(
205 Node* cntrl,
206 Node* i_o,
207 Node* memory,
208 Node* frameptr,
209 Node* ret_adr,
210 Node* exception
211 ) : Node(TypeFunc::Parms + 1) {
212 init_req(TypeFunc::Control , cntrl );
213 init_req(TypeFunc::I_O , i_o );
214 init_req(TypeFunc::Memory , memory );
215 init_req(TypeFunc::FramePtr , frameptr );
216 init_req(TypeFunc::ReturnAdr, ret_adr);
217 init_req(TypeFunc::Parms , exception);
218 }
219
220 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){
221 return remove_dead_region(phase, can_reshape) ? this : nullptr;
222 }
223
224 const Type* RethrowNode::Value(PhaseGVN* phase) const {
225 return (phase->type(in(TypeFunc::Control)) == Type::TOP)
226 ? Type::TOP
227 : Type::BOTTOM;
228 }
229
230 uint RethrowNode::match_edge(uint idx) const {
231 return 0;
232 }
233
234 #ifndef PRODUCT
235 void RethrowNode::dump_req(outputStream *st, DumpConfig* dc) const {
236 // Dump the required inputs, after printing "exception"
237 uint i; // Exit value of loop
238 for (i = 0; i < req(); i++) { // For all required inputs
239 if (i == TypeFunc::Parms) st->print("exception ");
240 Node* p = in(i);
241 if (p != nullptr) {
242 p->dump_idx(false, st, dc);
243 st->print(" ");
244 } else {
245 st->print("_ ");
246 }
247 }
248 }
249 #endif
250
251 //=============================================================================
252 // Do we Match on this edge index or not? Match only target address & method
253 uint TailCallNode::match_edge(uint idx) const {
254 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
255 }
256
257 //=============================================================================
258 // Do we Match on this edge index or not? Match only target address & oop
259 uint TailJumpNode::match_edge(uint idx) const {
260 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
261 }
262
263 //=============================================================================
264 JVMState::JVMState(ciMethod* method, JVMState* caller) :
265 _method(method) {
266 assert(method != nullptr, "must be valid call site");
267 _bci = InvocationEntryBci;
268 _reexecute = Reexecute_Undefined;
269 debug_only(_bci = -99); // random garbage value
270 debug_only(_map = (SafePointNode*)-1);
271 _caller = caller;
272 _depth = 1 + (caller == nullptr ? 0 : caller->depth());
273 _locoff = TypeFunc::Parms;
274 _stkoff = _locoff + _method->max_locals();
275 _monoff = _stkoff + _method->max_stack();
276 _scloff = _monoff;
277 _endoff = _monoff;
278 _sp = 0;
279 }
280 JVMState::JVMState(int stack_size) :
281 _method(nullptr) {
282 _bci = InvocationEntryBci;
283 _reexecute = Reexecute_Undefined;
284 debug_only(_map = (SafePointNode*)-1);
285 _caller = nullptr;
286 _depth = 1;
287 _locoff = TypeFunc::Parms;
288 _stkoff = _locoff;
289 _monoff = _stkoff + stack_size;
290 _scloff = _monoff;
291 _endoff = _monoff;
292 _sp = 0;
293 }
294
295 //--------------------------------of_depth-------------------------------------
296 JVMState* JVMState::of_depth(int d) const {
297 const JVMState* jvmp = this;
298 assert(0 < d && (uint)d <= depth(), "oob");
299 for (int skip = depth() - d; skip > 0; skip--) {
300 jvmp = jvmp->caller();
301 }
302 assert(jvmp->depth() == (uint)d, "found the right one");
303 return (JVMState*)jvmp;
304 }
305
306 //-----------------------------same_calls_as-----------------------------------
307 bool JVMState::same_calls_as(const JVMState* that) const {
308 if (this == that) return true;
309 if (this->depth() != that->depth()) return false;
310 const JVMState* p = this;
311 const JVMState* q = that;
312 for (;;) {
313 if (p->_method != q->_method) return false;
314 if (p->_method == nullptr) return true; // bci is irrelevant
315 if (p->_bci != q->_bci) return false;
316 if (p->_reexecute != q->_reexecute) return false;
317 p = p->caller();
318 q = q->caller();
319 if (p == q) return true;
320 assert(p != nullptr && q != nullptr, "depth check ensures we don't run off end");
321 }
322 }
323
324 //------------------------------debug_start------------------------------------
325 uint JVMState::debug_start() const {
326 debug_only(JVMState* jvmroot = of_depth(1));
327 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
328 return of_depth(1)->locoff();
329 }
330
331 //-------------------------------debug_end-------------------------------------
332 uint JVMState::debug_end() const {
333 debug_only(JVMState* jvmroot = of_depth(1));
334 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
335 return endoff();
336 }
337
338 //------------------------------debug_depth------------------------------------
339 uint JVMState::debug_depth() const {
340 uint total = 0;
341 for (const JVMState* jvmp = this; jvmp != nullptr; jvmp = jvmp->caller()) {
342 total += jvmp->debug_size();
343 }
344 return total;
345 }
346
347 #ifndef PRODUCT
348
349 //------------------------------format_helper----------------------------------
350 // Given an allocation (a Chaitin object) and a Node decide if the Node carries
351 // any defined value or not. If it does, print out the register or constant.
352 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
353 if (n == nullptr) { st->print(" null"); return; }
354 if (n->is_SafePointScalarObject()) {
355 // Scalar replacement.
356 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
357 scobjs->append_if_missing(spobj);
358 int sco_n = scobjs->find(spobj);
359 assert(sco_n >= 0, "");
360 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
361 return;
362 }
363 if (regalloc->node_regs_max_index() > 0 &&
364 OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
365 char buf[50];
366 regalloc->dump_register(n,buf,sizeof(buf));
367 st->print(" %s%d]=%s",msg,i,buf);
368 } else { // No register, but might be constant
369 const Type *t = n->bottom_type();
370 switch (t->base()) {
371 case Type::Int:
372 st->print(" %s%d]=#" INT32_FORMAT,msg,i,t->is_int()->get_con());
373 break;
374 case Type::AnyPtr:
375 assert( t == TypePtr::NULL_PTR || n->in_dump(), "" );
376 st->print(" %s%d]=#null",msg,i);
377 break;
378 case Type::AryPtr:
379 case Type::InstPtr:
380 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->isa_oopptr()->const_oop()));
381 break;
382 case Type::KlassPtr:
383 case Type::AryKlassPtr:
384 case Type::InstKlassPtr:
385 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_klassptr()->exact_klass()));
386 break;
387 case Type::MetadataPtr:
388 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_metadataptr()->metadata()));
389 break;
390 case Type::NarrowOop:
391 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_oopptr()->const_oop()));
392 break;
393 case Type::RawPtr:
394 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,p2i(t->is_rawptr()));
395 break;
396 case Type::DoubleCon:
397 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
398 break;
399 case Type::FloatCon:
400 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
401 break;
402 case Type::Long:
403 st->print(" %s%d]=#" INT64_FORMAT,msg,i,(int64_t)(t->is_long()->get_con()));
404 break;
405 case Type::Half:
406 case Type::Top:
407 st->print(" %s%d]=_",msg,i);
408 break;
409 default: ShouldNotReachHere();
410 }
411 }
412 }
413
414 //---------------------print_method_with_lineno--------------------------------
415 void JVMState::print_method_with_lineno(outputStream* st, bool show_name) const {
416 if (show_name) _method->print_short_name(st);
417
418 int lineno = _method->line_number_from_bci(_bci);
419 if (lineno != -1) {
420 st->print(" @ bci:%d (line %d)", _bci, lineno);
421 } else {
422 st->print(" @ bci:%d", _bci);
423 }
424 }
425
426 //------------------------------format-----------------------------------------
427 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
428 st->print(" #");
429 if (_method) {
430 print_method_with_lineno(st, true);
431 } else {
432 st->print_cr(" runtime stub ");
433 return;
434 }
435 if (n->is_MachSafePoint()) {
436 GrowableArray<SafePointScalarObjectNode*> scobjs;
437 MachSafePointNode *mcall = n->as_MachSafePoint();
438 uint i;
439 // Print locals
440 for (i = 0; i < (uint)loc_size(); i++)
441 format_helper(regalloc, st, mcall->local(this, i), "L[", i, &scobjs);
442 // Print stack
443 for (i = 0; i < (uint)stk_size(); i++) {
444 if ((uint)(_stkoff + i) >= mcall->len())
445 st->print(" oob ");
446 else
447 format_helper(regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs);
448 }
449 for (i = 0; (int)i < nof_monitors(); i++) {
450 Node *box = mcall->monitor_box(this, i);
451 Node *obj = mcall->monitor_obj(this, i);
452 if (regalloc->node_regs_max_index() > 0 &&
453 OptoReg::is_valid(regalloc->get_reg_first(box))) {
454 box = BoxLockNode::box_node(box);
455 format_helper(regalloc, st, box, "MON-BOX[", i, &scobjs);
456 } else {
457 OptoReg::Name box_reg = BoxLockNode::reg(box);
458 st->print(" MON-BOX%d=%s+%d",
459 i,
460 OptoReg::regname(OptoReg::c_frame_pointer),
461 regalloc->reg2offset(box_reg));
462 }
463 const char* obj_msg = "MON-OBJ[";
464 if (EliminateLocks) {
465 if (BoxLockNode::box_node(box)->is_eliminated())
466 obj_msg = "MON-OBJ(LOCK ELIMINATED)[";
467 }
468 format_helper(regalloc, st, obj, obj_msg, i, &scobjs);
469 }
470
471 for (i = 0; i < (uint)scobjs.length(); i++) {
472 // Scalar replaced objects.
473 st->cr();
474 st->print(" # ScObj" INT32_FORMAT " ", i);
475 SafePointScalarObjectNode* spobj = scobjs.at(i);
476 ciKlass* cik = spobj->bottom_type()->is_oopptr()->exact_klass();
477 assert(cik->is_instance_klass() ||
478 cik->is_array_klass(), "Not supported allocation.");
479 ciInstanceKlass *iklass = nullptr;
480 if (cik->is_instance_klass()) {
481 cik->print_name_on(st);
482 iklass = cik->as_instance_klass();
483 } else if (cik->is_type_array_klass()) {
484 cik->as_array_klass()->base_element_type()->print_name_on(st);
485 st->print("[%d]", spobj->n_fields());
486 } else if (cik->is_obj_array_klass()) {
487 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass();
488 if (cie->is_instance_klass()) {
489 cie->print_name_on(st);
490 } else if (cie->is_type_array_klass()) {
491 cie->as_array_klass()->base_element_type()->print_name_on(st);
492 } else {
493 ShouldNotReachHere();
494 }
495 st->print("[%d]", spobj->n_fields());
496 int ndim = cik->as_array_klass()->dimension() - 1;
497 while (ndim-- > 0) {
498 st->print("[]");
499 }
500 }
501 st->print("={");
502 uint nf = spobj->n_fields();
503 if (nf > 0) {
504 uint first_ind = spobj->first_index(mcall->jvms());
505 Node* fld_node = mcall->in(first_ind);
506 ciField* cifield;
507 if (iklass != nullptr) {
508 st->print(" [");
509 cifield = iklass->nonstatic_field_at(0);
510 cifield->print_name_on(st);
511 format_helper(regalloc, st, fld_node, ":", 0, &scobjs);
512 } else {
513 format_helper(regalloc, st, fld_node, "[", 0, &scobjs);
514 }
515 for (uint j = 1; j < nf; j++) {
516 fld_node = mcall->in(first_ind+j);
517 if (iklass != nullptr) {
518 st->print(", [");
519 cifield = iklass->nonstatic_field_at(j);
520 cifield->print_name_on(st);
521 format_helper(regalloc, st, fld_node, ":", j, &scobjs);
522 } else {
523 format_helper(regalloc, st, fld_node, ", [", j, &scobjs);
524 }
525 }
526 }
527 st->print(" }");
528 }
529 }
530 st->cr();
531 if (caller() != nullptr) caller()->format(regalloc, n, st);
532 }
533
534
535 void JVMState::dump_spec(outputStream *st) const {
536 if (_method != nullptr) {
537 bool printed = false;
538 if (!Verbose) {
539 // The JVMS dumps make really, really long lines.
540 // Take out the most boring parts, which are the package prefixes.
541 char buf[500];
542 stringStream namest(buf, sizeof(buf));
543 _method->print_short_name(&namest);
544 if (namest.count() < sizeof(buf)) {
545 const char* name = namest.base();
546 if (name[0] == ' ') ++name;
547 const char* endcn = strchr(name, ':'); // end of class name
548 if (endcn == nullptr) endcn = strchr(name, '(');
549 if (endcn == nullptr) endcn = name + strlen(name);
550 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
551 --endcn;
552 st->print(" %s", endcn);
553 printed = true;
554 }
555 }
556 print_method_with_lineno(st, !printed);
557 if(_reexecute == Reexecute_True)
558 st->print(" reexecute");
559 } else {
560 st->print(" runtime stub");
561 }
562 if (caller() != nullptr) caller()->dump_spec(st);
563 }
564
565
566 void JVMState::dump_on(outputStream* st) const {
567 bool print_map = _map && !((uintptr_t)_map & 1) &&
568 ((caller() == nullptr) || (caller()->map() != _map));
569 if (print_map) {
570 if (_map->len() > _map->req()) { // _map->has_exceptions()
571 Node* ex = _map->in(_map->req()); // _map->next_exception()
572 // skip the first one; it's already being printed
573 while (ex != nullptr && ex->len() > ex->req()) {
574 ex = ex->in(ex->req()); // ex->next_exception()
575 ex->dump(1);
576 }
577 }
578 _map->dump(Verbose ? 2 : 1);
579 }
580 if (caller() != nullptr) {
581 caller()->dump_on(st);
582 }
583 st->print("JVMS depth=%d loc=%d stk=%d arg=%d mon=%d scalar=%d end=%d mondepth=%d sp=%d bci=%d reexecute=%s method=",
584 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false");
585 if (_method == nullptr) {
586 st->print_cr("(none)");
587 } else {
588 _method->print_name(st);
589 st->cr();
590 if (bci() >= 0 && bci() < _method->code_size()) {
591 st->print(" bc: ");
592 _method->print_codes_on(bci(), bci()+1, st);
593 }
594 }
595 }
596
597 // Extra way to dump a jvms from the debugger,
598 // to avoid a bug with C++ member function calls.
599 void dump_jvms(JVMState* jvms) {
600 jvms->dump();
601 }
602 #endif
603
604 //--------------------------clone_shallow--------------------------------------
605 JVMState* JVMState::clone_shallow(Compile* C) const {
606 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
607 n->set_bci(_bci);
608 n->_reexecute = _reexecute;
609 n->set_locoff(_locoff);
610 n->set_stkoff(_stkoff);
611 n->set_monoff(_monoff);
612 n->set_scloff(_scloff);
613 n->set_endoff(_endoff);
614 n->set_sp(_sp);
615 n->set_map(_map);
616 return n;
617 }
618
619 //---------------------------clone_deep----------------------------------------
620 JVMState* JVMState::clone_deep(Compile* C) const {
621 JVMState* n = clone_shallow(C);
622 for (JVMState* p = n; p->_caller != nullptr; p = p->_caller) {
623 p->_caller = p->_caller->clone_shallow(C);
624 }
625 assert(n->depth() == depth(), "sanity");
626 assert(n->debug_depth() == debug_depth(), "sanity");
627 return n;
628 }
629
630 /**
631 * Reset map for all callers
632 */
633 void JVMState::set_map_deep(SafePointNode* map) {
634 for (JVMState* p = this; p != nullptr; p = p->_caller) {
635 p->set_map(map);
636 }
637 }
638
639 // unlike set_map(), this is two-way setting.
640 void JVMState::bind_map(SafePointNode* map) {
641 set_map(map);
642 _map->set_jvms(this);
643 }
644
645 // Adapt offsets in in-array after adding or removing an edge.
646 // Prerequisite is that the JVMState is used by only one node.
647 void JVMState::adapt_position(int delta) {
648 for (JVMState* jvms = this; jvms != nullptr; jvms = jvms->caller()) {
649 jvms->set_locoff(jvms->locoff() + delta);
650 jvms->set_stkoff(jvms->stkoff() + delta);
651 jvms->set_monoff(jvms->monoff() + delta);
652 jvms->set_scloff(jvms->scloff() + delta);
653 jvms->set_endoff(jvms->endoff() + delta);
654 }
655 }
656
657 // Mirror the stack size calculation in the deopt code
658 // How much stack space would we need at this point in the program in
659 // case of deoptimization?
660 int JVMState::interpreter_frame_size() const {
661 const JVMState* jvms = this;
662 int size = 0;
663 int callee_parameters = 0;
664 int callee_locals = 0;
665 int extra_args = method()->max_stack() - stk_size();
666
667 while (jvms != nullptr) {
668 int locks = jvms->nof_monitors();
669 int temps = jvms->stk_size();
670 bool is_top_frame = (jvms == this);
671 ciMethod* method = jvms->method();
672
673 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(),
674 temps + callee_parameters,
675 extra_args,
676 locks,
677 callee_parameters,
678 callee_locals,
679 is_top_frame);
680 size += frame_size;
681
682 callee_parameters = method->size_of_parameters();
683 callee_locals = method->max_locals();
684 extra_args = 0;
685 jvms = jvms->caller();
686 }
687 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord;
688 }
689
690 //=============================================================================
691 bool CallNode::cmp( const Node &n ) const
692 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
693 #ifndef PRODUCT
694 void CallNode::dump_req(outputStream *st, DumpConfig* dc) const {
695 // Dump the required inputs, enclosed in '(' and ')'
696 uint i; // Exit value of loop
697 for (i = 0; i < req(); i++) { // For all required inputs
698 if (i == TypeFunc::Parms) st->print("(");
699 Node* p = in(i);
700 if (p != nullptr) {
701 p->dump_idx(false, st, dc);
702 st->print(" ");
703 } else {
704 st->print("_ ");
705 }
706 }
707 st->print(")");
708 }
709
710 void CallNode::dump_spec(outputStream *st) const {
711 st->print(" ");
712 if (tf() != nullptr) tf()->dump_on(st);
713 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt);
714 if (jvms() != nullptr) jvms()->dump_spec(st);
715 }
716
717 void AllocateNode::dump_spec(outputStream* st) const {
718 st->print(" ");
719 if (tf() != nullptr) {
720 tf()->dump_on(st);
721 }
722 if (_cnt != COUNT_UNKNOWN) {
723 st->print(" C=%f", _cnt);
724 }
725 const Node* const klass_node = in(KlassNode);
726 if (klass_node != nullptr) {
727 const TypeKlassPtr* const klass_ptr = klass_node->bottom_type()->isa_klassptr();
728
729 if (klass_ptr != nullptr && klass_ptr->klass_is_exact()) {
730 st->print(" allocationKlass:");
731 klass_ptr->exact_klass()->print_name_on(st);
732 }
733 }
734 if (jvms() != nullptr) {
735 jvms()->dump_spec(st);
736 }
737 }
738 #endif
739
740 const Type *CallNode::bottom_type() const { return tf()->range(); }
741 const Type* CallNode::Value(PhaseGVN* phase) const {
742 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
743 return Type::TOP;
744 }
745 return tf()->range();
746 }
747
748 //------------------------------calling_convention-----------------------------
749 void CallNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
750 // Use the standard compiler calling convention
751 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt);
752 }
753
754
755 //------------------------------match------------------------------------------
756 // Construct projections for control, I/O, memory-fields, ..., and
757 // return result(s) along with their RegMask info
758 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
759 switch (proj->_con) {
760 case TypeFunc::Control:
761 case TypeFunc::I_O:
762 case TypeFunc::Memory:
763 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
764
765 case TypeFunc::Parms+1: // For LONG & DOUBLE returns
766 assert(tf()->range()->field_at(TypeFunc::Parms+1) == Type::HALF, "");
767 // 2nd half of doubles and longs
768 return new MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad);
769
770 case TypeFunc::Parms: { // Normal returns
771 uint ideal_reg = tf()->range()->field_at(TypeFunc::Parms)->ideal_reg();
772 OptoRegPair regs = Opcode() == Op_CallLeafVector
773 ? match->vector_return_value(ideal_reg) // Calls into assembly vector routine
774 : is_CallRuntime()
775 ? match->c_return_value(ideal_reg) // Calls into C runtime
776 : match-> return_value(ideal_reg); // Calls into compiled Java code
777 RegMask rm = RegMask(regs.first());
778
779 if (Opcode() == Op_CallLeafVector) {
780 // If the return is in vector, compute appropriate regmask taking into account the whole range
781 if(ideal_reg >= Op_VecA && ideal_reg <= Op_VecZ) {
782 if(OptoReg::is_valid(regs.second())) {
783 for (OptoReg::Name r = regs.first(); r <= regs.second(); r = OptoReg::add(r, 1)) {
784 rm.Insert(r);
785 }
786 }
787 }
788 }
789
790 if( OptoReg::is_valid(regs.second()) )
791 rm.Insert( regs.second() );
792 return new MachProjNode(this,proj->_con,rm,ideal_reg);
793 }
794
795 case TypeFunc::ReturnAdr:
796 case TypeFunc::FramePtr:
797 default:
798 ShouldNotReachHere();
799 }
800 return nullptr;
801 }
802
803 // Do we Match on this edge index or not? Match no edges
804 uint CallNode::match_edge(uint idx) const {
805 return 0;
806 }
807
808 //
809 // Determine whether the call could modify the field of the specified
810 // instance at the specified offset.
811 //
812 bool CallNode::may_modify(const TypeOopPtr* t_oop, PhaseValues* phase) {
813 assert((t_oop != nullptr), "sanity");
814 if (is_call_to_arraycopystub() && strcmp(_name, "unsafe_arraycopy") != 0) {
815 const TypeTuple* args = _tf->domain();
816 Node* dest = nullptr;
817 // Stubs that can be called once an ArrayCopyNode is expanded have
818 // different signatures. Look for the second pointer argument,
819 // that is the destination of the copy.
820 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
821 if (args->field_at(i)->isa_ptr()) {
822 j++;
823 if (j == 2) {
824 dest = in(i);
825 break;
826 }
827 }
828 }
829 guarantee(dest != nullptr, "Call had only one ptr in, broken IR!");
830 if (!dest->is_top() && may_modify_arraycopy_helper(phase->type(dest)->is_oopptr(), t_oop, phase)) {
831 return true;
832 }
833 return false;
834 }
835 if (t_oop->is_known_instance()) {
836 // The instance_id is set only for scalar-replaceable allocations which
837 // are not passed as arguments according to Escape Analysis.
838 return false;
839 }
840 if (t_oop->is_ptr_to_boxed_value()) {
841 ciKlass* boxing_klass = t_oop->is_instptr()->instance_klass();
842 if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) {
843 // Skip unrelated boxing methods.
844 Node* proj = proj_out_or_null(TypeFunc::Parms);
845 if ((proj == nullptr) || (phase->type(proj)->is_instptr()->instance_klass() != boxing_klass)) {
846 return false;
847 }
848 }
849 if (is_CallJava() && as_CallJava()->method() != nullptr) {
850 ciMethod* meth = as_CallJava()->method();
851 if (meth->is_getter()) {
852 return false;
853 }
854 // May modify (by reflection) if an boxing object is passed
855 // as argument or returned.
856 Node* proj = returns_pointer() ? proj_out_or_null(TypeFunc::Parms) : nullptr;
857 if (proj != nullptr) {
858 const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr();
859 if ((inst_t != nullptr) && (!inst_t->klass_is_exact() ||
860 (inst_t->instance_klass() == boxing_klass))) {
861 return true;
862 }
863 }
864 const TypeTuple* d = tf()->domain();
865 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
866 const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr();
867 if ((inst_t != nullptr) && (!inst_t->klass_is_exact() ||
868 (inst_t->instance_klass() == boxing_klass))) {
869 return true;
870 }
871 }
872 return false;
873 }
874 }
875 return true;
876 }
877
878 // Does this call have a direct reference to n other than debug information?
879 bool CallNode::has_non_debug_use(Node *n) {
880 const TypeTuple * d = tf()->domain();
881 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
882 Node *arg = in(i);
883 if (arg == n) {
884 return true;
885 }
886 }
887 return false;
888 }
889
890 // Returns the unique CheckCastPP of a call
891 // or 'this' if there are several CheckCastPP or unexpected uses
892 // or returns null if there is no one.
893 Node *CallNode::result_cast() {
894 Node *cast = nullptr;
895
896 Node *p = proj_out_or_null(TypeFunc::Parms);
897 if (p == nullptr)
898 return nullptr;
899
900 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
901 Node *use = p->fast_out(i);
902 if (use->is_CheckCastPP()) {
903 if (cast != nullptr) {
904 return this; // more than 1 CheckCastPP
905 }
906 cast = use;
907 } else if (!use->is_Initialize() &&
908 !use->is_AddP() &&
909 use->Opcode() != Op_MemBarStoreStore) {
910 // Expected uses are restricted to a CheckCastPP, an Initialize
911 // node, a MemBarStoreStore (clone) and AddP nodes. If we
912 // encounter any other use (a Phi node can be seen in rare
913 // cases) return this to prevent incorrect optimizations.
914 return this;
915 }
916 }
917 return cast;
918 }
919
920
921 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts) {
922 projs->fallthrough_proj = nullptr;
923 projs->fallthrough_catchproj = nullptr;
924 projs->fallthrough_ioproj = nullptr;
925 projs->catchall_ioproj = nullptr;
926 projs->catchall_catchproj = nullptr;
927 projs->fallthrough_memproj = nullptr;
928 projs->catchall_memproj = nullptr;
929 projs->resproj = nullptr;
930 projs->exobj = nullptr;
931
932 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
933 ProjNode *pn = fast_out(i)->as_Proj();
934 if (pn->outcnt() == 0) continue;
935 switch (pn->_con) {
936 case TypeFunc::Control:
937 {
938 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
939 projs->fallthrough_proj = pn;
940 const Node* cn = pn->unique_ctrl_out_or_null();
941 if (cn != nullptr && cn->is_Catch()) {
942 ProjNode *cpn = nullptr;
943 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
944 cpn = cn->fast_out(k)->as_Proj();
945 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
946 if (cpn->_con == CatchProjNode::fall_through_index)
947 projs->fallthrough_catchproj = cpn;
948 else {
949 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
950 projs->catchall_catchproj = cpn;
951 }
952 }
953 }
954 break;
955 }
956 case TypeFunc::I_O:
957 if (pn->_is_io_use)
958 projs->catchall_ioproj = pn;
959 else
960 projs->fallthrough_ioproj = pn;
961 for (DUIterator j = pn->outs(); pn->has_out(j); j++) {
962 Node* e = pn->out(j);
963 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) {
964 assert(projs->exobj == nullptr, "only one");
965 projs->exobj = e;
966 }
967 }
968 break;
969 case TypeFunc::Memory:
970 if (pn->_is_io_use)
971 projs->catchall_memproj = pn;
972 else
973 projs->fallthrough_memproj = pn;
974 break;
975 case TypeFunc::Parms:
976 projs->resproj = pn;
977 break;
978 default:
979 assert(false, "unexpected projection from allocation node.");
980 }
981 }
982
983 // The resproj may not exist because the result could be ignored
984 // and the exception object may not exist if an exception handler
985 // swallows the exception but all the other must exist and be found.
986 assert(projs->fallthrough_proj != nullptr, "must be found");
987 do_asserts = do_asserts && !Compile::current()->inlining_incrementally();
988 assert(!do_asserts || projs->fallthrough_catchproj != nullptr, "must be found");
989 assert(!do_asserts || projs->fallthrough_memproj != nullptr, "must be found");
990 assert(!do_asserts || projs->fallthrough_ioproj != nullptr, "must be found");
991 assert(!do_asserts || projs->catchall_catchproj != nullptr, "must be found");
992 if (separate_io_proj) {
993 assert(!do_asserts || projs->catchall_memproj != nullptr, "must be found");
994 assert(!do_asserts || projs->catchall_ioproj != nullptr, "must be found");
995 }
996 }
997
998 Node* CallNode::Ideal(PhaseGVN* phase, bool can_reshape) {
999 #ifdef ASSERT
1000 // Validate attached generator
1001 CallGenerator* cg = generator();
1002 if (cg != nullptr) {
1003 assert((is_CallStaticJava() && cg->is_mh_late_inline()) ||
1004 (is_CallDynamicJava() && cg->is_virtual_late_inline()), "mismatch");
1005 }
1006 #endif // ASSERT
1007 return SafePointNode::Ideal(phase, can_reshape);
1008 }
1009
1010 bool CallNode::is_call_to_arraycopystub() const {
1011 if (_name != nullptr && strstr(_name, "arraycopy") != nullptr) {
1012 return true;
1013 }
1014 return false;
1015 }
1016
1017 //=============================================================================
1018 uint CallJavaNode::size_of() const { return sizeof(*this); }
1019 bool CallJavaNode::cmp( const Node &n ) const {
1020 CallJavaNode &call = (CallJavaNode&)n;
1021 return CallNode::cmp(call) && _method == call._method &&
1022 _override_symbolic_info == call._override_symbolic_info;
1023 }
1024
1025 void CallJavaNode::copy_call_debug_info(PhaseIterGVN* phase, SafePointNode* sfpt) {
1026 // Copy debug information and adjust JVMState information
1027 uint old_dbg_start = sfpt->is_Call() ? sfpt->as_Call()->tf()->domain()->cnt() : (uint)TypeFunc::Parms+1;
1028 uint new_dbg_start = tf()->domain()->cnt();
1029 int jvms_adj = new_dbg_start - old_dbg_start;
1030 assert (new_dbg_start == req(), "argument count mismatch");
1031 Compile* C = phase->C;
1032
1033 // SafePointScalarObject node could be referenced several times in debug info.
1034 // Use Dict to record cloned nodes.
1035 Dict* sosn_map = new Dict(cmpkey,hashkey);
1036 for (uint i = old_dbg_start; i < sfpt->req(); i++) {
1037 Node* old_in = sfpt->in(i);
1038 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
1039 if (old_in != nullptr && old_in->is_SafePointScalarObject()) {
1040 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
1041 bool new_node;
1042 Node* new_in = old_sosn->clone(sosn_map, new_node);
1043 if (new_node) { // New node?
1044 new_in->set_req(0, C->root()); // reset control edge
1045 new_in = phase->transform(new_in); // Register new node.
1046 }
1047 old_in = new_in;
1048 }
1049 add_req(old_in);
1050 }
1051
1052 // JVMS may be shared so clone it before we modify it
1053 set_jvms(sfpt->jvms() != nullptr ? sfpt->jvms()->clone_deep(C) : nullptr);
1054 for (JVMState *jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1055 jvms->set_map(this);
1056 jvms->set_locoff(jvms->locoff()+jvms_adj);
1057 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
1058 jvms->set_monoff(jvms->monoff()+jvms_adj);
1059 jvms->set_scloff(jvms->scloff()+jvms_adj);
1060 jvms->set_endoff(jvms->endoff()+jvms_adj);
1061 }
1062 }
1063
1064 #ifdef ASSERT
1065 bool CallJavaNode::validate_symbolic_info() const {
1066 if (method() == nullptr) {
1067 return true; // call into runtime or uncommon trap
1068 }
1069 ciMethod* symbolic_info = jvms()->method()->get_method_at_bci(jvms()->bci());
1070 ciMethod* callee = method();
1071 if (symbolic_info->is_method_handle_intrinsic() && !callee->is_method_handle_intrinsic()) {
1072 assert(override_symbolic_info(), "should be set");
1073 }
1074 assert(ciMethod::is_consistent_info(symbolic_info, callee), "inconsistent info");
1075 return true;
1076 }
1077 #endif
1078
1079 #ifndef PRODUCT
1080 void CallJavaNode::dump_spec(outputStream* st) const {
1081 if( _method ) _method->print_short_name(st);
1082 CallNode::dump_spec(st);
1083 }
1084
1085 void CallJavaNode::dump_compact_spec(outputStream* st) const {
1086 if (_method) {
1087 _method->print_short_name(st);
1088 } else {
1089 st->print("<?>");
1090 }
1091 }
1092 #endif
1093
1094 void CallJavaNode::register_for_late_inline() {
1095 if (generator() != nullptr) {
1096 Compile::current()->prepend_late_inline(generator());
1097 set_generator(nullptr);
1098 } else {
1099 assert(false, "repeated inline attempt");
1100 }
1101 }
1102
1103 //=============================================================================
1104 uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
1105 bool CallStaticJavaNode::cmp( const Node &n ) const {
1106 CallStaticJavaNode &call = (CallStaticJavaNode&)n;
1107 return CallJavaNode::cmp(call);
1108 }
1109
1110 Node* CallStaticJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1111 CallGenerator* cg = generator();
1112 if (can_reshape && cg != nullptr) {
1113 if (cg->is_mh_late_inline()) {
1114 assert(IncrementalInlineMH, "required");
1115 assert(cg->call_node() == this, "mismatch");
1116 assert(cg->method()->is_method_handle_intrinsic(), "required");
1117
1118 // Check whether this MH handle call becomes a candidate for inlining.
1119 ciMethod* callee = cg->method();
1120 vmIntrinsics::ID iid = callee->intrinsic_id();
1121 if (iid == vmIntrinsics::_invokeBasic) {
1122 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) {
1123 register_for_late_inline();
1124 }
1125 } else if (iid == vmIntrinsics::_linkToNative) {
1126 // never retry
1127 } else {
1128 assert(callee->has_member_arg(), "wrong type of call?");
1129 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) {
1130 register_for_late_inline();
1131 phase->C->inc_number_of_mh_late_inlines();
1132 }
1133 }
1134 } else {
1135 assert(IncrementalInline, "required");
1136 assert(!cg->method()->is_method_handle_intrinsic(), "required");
1137 if (phase->C->print_inlining()) {
1138 phase->C->inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE,
1139 "static call node changed: trying again");
1140 }
1141 register_for_late_inline();
1142 }
1143 }
1144 return CallNode::Ideal(phase, can_reshape);
1145 }
1146
1147 //----------------------------is_uncommon_trap----------------------------
1148 // Returns true if this is an uncommon trap.
1149 bool CallStaticJavaNode::is_uncommon_trap() const {
1150 return (_name != nullptr && !strcmp(_name, "uncommon_trap"));
1151 }
1152
1153 //----------------------------uncommon_trap_request----------------------------
1154 // If this is an uncommon trap, return the request code, else zero.
1155 int CallStaticJavaNode::uncommon_trap_request() const {
1156 return is_uncommon_trap() ? extract_uncommon_trap_request(this) : 0;
1157 }
1158 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
1159 #ifndef PRODUCT
1160 if (!(call->req() > TypeFunc::Parms &&
1161 call->in(TypeFunc::Parms) != nullptr &&
1162 call->in(TypeFunc::Parms)->is_Con() &&
1163 call->in(TypeFunc::Parms)->bottom_type()->isa_int())) {
1164 assert(in_dump() != 0, "OK if dumping");
1165 tty->print("[bad uncommon trap]");
1166 return 0;
1167 }
1168 #endif
1169 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
1170 }
1171
1172 #ifndef PRODUCT
1173 void CallStaticJavaNode::dump_spec(outputStream *st) const {
1174 st->print("# Static ");
1175 if (_name != nullptr) {
1176 st->print("%s", _name);
1177 int trap_req = uncommon_trap_request();
1178 if (trap_req != 0) {
1179 char buf[100];
1180 st->print("(%s)",
1181 Deoptimization::format_trap_request(buf, sizeof(buf),
1182 trap_req));
1183 }
1184 st->print(" ");
1185 }
1186 CallJavaNode::dump_spec(st);
1187 }
1188
1189 void CallStaticJavaNode::dump_compact_spec(outputStream* st) const {
1190 if (_method) {
1191 _method->print_short_name(st);
1192 } else if (_name) {
1193 st->print("%s", _name);
1194 } else {
1195 st->print("<?>");
1196 }
1197 }
1198 #endif
1199
1200 //=============================================================================
1201 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
1202 bool CallDynamicJavaNode::cmp( const Node &n ) const {
1203 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
1204 return CallJavaNode::cmp(call);
1205 }
1206
1207 Node* CallDynamicJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1208 CallGenerator* cg = generator();
1209 if (can_reshape && cg != nullptr) {
1210 if (cg->is_virtual_late_inline()) {
1211 assert(IncrementalInlineVirtual, "required");
1212 assert(cg->call_node() == this, "mismatch");
1213
1214 // Recover symbolic info for method resolution.
1215 ciMethod* caller = jvms()->method();
1216 ciBytecodeStream iter(caller);
1217 iter.force_bci(jvms()->bci());
1218
1219 bool not_used1;
1220 ciSignature* not_used2;
1221 ciMethod* orig_callee = iter.get_method(not_used1, ¬_used2); // callee in the bytecode
1222 ciKlass* holder = iter.get_declared_method_holder();
1223 if (orig_callee->is_method_handle_intrinsic()) {
1224 assert(_override_symbolic_info, "required");
1225 orig_callee = method();
1226 holder = method()->holder();
1227 }
1228
1229 ciInstanceKlass* klass = ciEnv::get_instance_klass_for_declared_method_holder(holder);
1230
1231 Node* receiver_node = in(TypeFunc::Parms);
1232 const TypeOopPtr* receiver_type = phase->type(receiver_node)->isa_oopptr();
1233
1234 int not_used3;
1235 bool call_does_dispatch;
1236 ciMethod* callee = phase->C->optimize_virtual_call(caller, klass, holder, orig_callee, receiver_type, true /*is_virtual*/,
1237 call_does_dispatch, not_used3); // out-parameters
1238 if (!call_does_dispatch) {
1239 // Register for late inlining.
1240 cg->set_callee_method(callee);
1241 register_for_late_inline(); // MH late inlining prepends to the list, so do the same
1242 }
1243 } else {
1244 assert(IncrementalInline, "required");
1245 if (phase->C->print_inlining()) {
1246 phase->C->inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE,
1247 "dynamic call node changed: trying again");
1248 }
1249 register_for_late_inline();
1250 }
1251 }
1252 return CallNode::Ideal(phase, can_reshape);
1253 }
1254
1255 #ifndef PRODUCT
1256 void CallDynamicJavaNode::dump_spec(outputStream *st) const {
1257 st->print("# Dynamic ");
1258 CallJavaNode::dump_spec(st);
1259 }
1260 #endif
1261
1262 //=============================================================================
1263 uint CallRuntimeNode::size_of() const { return sizeof(*this); }
1264 bool CallRuntimeNode::cmp( const Node &n ) const {
1265 CallRuntimeNode &call = (CallRuntimeNode&)n;
1266 return CallNode::cmp(call) && !strcmp(_name,call._name);
1267 }
1268 #ifndef PRODUCT
1269 void CallRuntimeNode::dump_spec(outputStream *st) const {
1270 st->print("# ");
1271 st->print("%s", _name);
1272 CallNode::dump_spec(st);
1273 }
1274 #endif
1275 uint CallLeafVectorNode::size_of() const { return sizeof(*this); }
1276 bool CallLeafVectorNode::cmp( const Node &n ) const {
1277 CallLeafVectorNode &call = (CallLeafVectorNode&)n;
1278 return CallLeafNode::cmp(call) && _num_bits == call._num_bits;
1279 }
1280
1281 //------------------------------calling_convention-----------------------------
1282 void CallRuntimeNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
1283 SharedRuntime::c_calling_convention(sig_bt, parm_regs, argcnt);
1284 }
1285
1286 void CallLeafVectorNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
1287 #ifdef ASSERT
1288 assert(tf()->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1289 "return vector size must match");
1290 const TypeTuple* d = tf()->domain();
1291 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1292 Node* arg = in(i);
1293 assert(arg->bottom_type()->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1294 "vector argument size must match");
1295 }
1296 #endif
1297
1298 SharedRuntime::vector_calling_convention(parm_regs, _num_bits, argcnt);
1299 }
1300
1301 //=============================================================================
1302 //------------------------------calling_convention-----------------------------
1303
1304
1305 //=============================================================================
1306 #ifndef PRODUCT
1307 void CallLeafNode::dump_spec(outputStream *st) const {
1308 st->print("# ");
1309 st->print("%s", _name);
1310 CallNode::dump_spec(st);
1311 }
1312 #endif
1313
1314 //=============================================================================
1315
1316 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) {
1317 assert(verify_jvms(jvms), "jvms must match");
1318 int loc = jvms->locoff() + idx;
1319 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
1320 // If current local idx is top then local idx - 1 could
1321 // be a long/double that needs to be killed since top could
1322 // represent the 2nd half of the long/double.
1323 uint ideal = in(loc -1)->ideal_reg();
1324 if (ideal == Op_RegD || ideal == Op_RegL) {
1325 // set other (low index) half to top
1326 set_req(loc - 1, in(loc));
1327 }
1328 }
1329 set_req(loc, c);
1330 }
1331
1332 uint SafePointNode::size_of() const { return sizeof(*this); }
1333 bool SafePointNode::cmp( const Node &n ) const {
1334 return (&n == this); // Always fail except on self
1335 }
1336
1337 //-------------------------set_next_exception----------------------------------
1338 void SafePointNode::set_next_exception(SafePointNode* n) {
1339 assert(n == nullptr || n->Opcode() == Op_SafePoint, "correct value for next_exception");
1340 if (len() == req()) {
1341 if (n != nullptr) add_prec(n);
1342 } else {
1343 set_prec(req(), n);
1344 }
1345 }
1346
1347
1348 //----------------------------next_exception-----------------------------------
1349 SafePointNode* SafePointNode::next_exception() const {
1350 if (len() == req()) {
1351 return nullptr;
1352 } else {
1353 Node* n = in(req());
1354 assert(n == nullptr || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
1355 return (SafePointNode*) n;
1356 }
1357 }
1358
1359
1360 //------------------------------Ideal------------------------------------------
1361 // Skip over any collapsed Regions
1362 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1363 assert(_jvms == nullptr || ((uintptr_t)_jvms->map() & 1) || _jvms->map() == this, "inconsistent JVMState");
1364 return remove_dead_region(phase, can_reshape) ? this : nullptr;
1365 }
1366
1367 //------------------------------Identity---------------------------------------
1368 // Remove obviously duplicate safepoints
1369 Node* SafePointNode::Identity(PhaseGVN* phase) {
1370
1371 // If you have back to back safepoints, remove one
1372 if (in(TypeFunc::Control)->is_SafePoint()) {
1373 Node* out_c = unique_ctrl_out_or_null();
1374 // This can be the safepoint of an outer strip mined loop if the inner loop's backedge was removed. Replacing the
1375 // outer loop's safepoint could confuse removal of the outer loop.
1376 if (out_c != nullptr && !out_c->is_OuterStripMinedLoopEnd()) {
1377 return in(TypeFunc::Control);
1378 }
1379 }
1380
1381 // Transforming long counted loops requires a safepoint node. Do not
1382 // eliminate a safepoint until loop opts are over.
1383 if (in(0)->is_Proj() && !phase->C->major_progress()) {
1384 Node *n0 = in(0)->in(0);
1385 // Check if he is a call projection (except Leaf Call)
1386 if( n0->is_Catch() ) {
1387 n0 = n0->in(0)->in(0);
1388 assert( n0->is_Call(), "expect a call here" );
1389 }
1390 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
1391 // Don't remove a safepoint belonging to an OuterStripMinedLoopEndNode.
1392 // If the loop dies, they will be removed together.
1393 if (has_out_with(Op_OuterStripMinedLoopEnd)) {
1394 return this;
1395 }
1396 // Useless Safepoint, so remove it
1397 return in(TypeFunc::Control);
1398 }
1399 }
1400
1401 return this;
1402 }
1403
1404 //------------------------------Value------------------------------------------
1405 const Type* SafePointNode::Value(PhaseGVN* phase) const {
1406 if (phase->type(in(0)) == Type::TOP) {
1407 return Type::TOP;
1408 }
1409 if (in(0) == this) {
1410 return Type::TOP; // Dead infinite loop
1411 }
1412 return Type::CONTROL;
1413 }
1414
1415 #ifndef PRODUCT
1416 void SafePointNode::dump_spec(outputStream *st) const {
1417 st->print(" SafePoint ");
1418 _replaced_nodes.dump(st);
1419 }
1420 #endif
1421
1422 const RegMask &SafePointNode::in_RegMask(uint idx) const {
1423 if( idx < TypeFunc::Parms ) return RegMask::Empty;
1424 // Values outside the domain represent debug info
1425 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1426 }
1427 const RegMask &SafePointNode::out_RegMask() const {
1428 return RegMask::Empty;
1429 }
1430
1431
1432 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
1433 assert((int)grow_by > 0, "sanity");
1434 int monoff = jvms->monoff();
1435 int scloff = jvms->scloff();
1436 int endoff = jvms->endoff();
1437 assert(endoff == (int)req(), "no other states or debug info after me");
1438 Node* top = Compile::current()->top();
1439 for (uint i = 0; i < grow_by; i++) {
1440 ins_req(monoff, top);
1441 }
1442 jvms->set_monoff(monoff + grow_by);
1443 jvms->set_scloff(scloff + grow_by);
1444 jvms->set_endoff(endoff + grow_by);
1445 }
1446
1447 void SafePointNode::push_monitor(const FastLockNode *lock) {
1448 // Add a LockNode, which points to both the original BoxLockNode (the
1449 // stack space for the monitor) and the Object being locked.
1450 const int MonitorEdges = 2;
1451 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1452 assert(req() == jvms()->endoff(), "correct sizing");
1453 int nextmon = jvms()->scloff();
1454 if (GenerateSynchronizationCode) {
1455 ins_req(nextmon, lock->box_node());
1456 ins_req(nextmon+1, lock->obj_node());
1457 } else {
1458 Node* top = Compile::current()->top();
1459 ins_req(nextmon, top);
1460 ins_req(nextmon, top);
1461 }
1462 jvms()->set_scloff(nextmon + MonitorEdges);
1463 jvms()->set_endoff(req());
1464 }
1465
1466 void SafePointNode::pop_monitor() {
1467 // Delete last monitor from debug info
1468 debug_only(int num_before_pop = jvms()->nof_monitors());
1469 const int MonitorEdges = 2;
1470 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1471 int scloff = jvms()->scloff();
1472 int endoff = jvms()->endoff();
1473 int new_scloff = scloff - MonitorEdges;
1474 int new_endoff = endoff - MonitorEdges;
1475 jvms()->set_scloff(new_scloff);
1476 jvms()->set_endoff(new_endoff);
1477 while (scloff > new_scloff) del_req_ordered(--scloff);
1478 assert(jvms()->nof_monitors() == num_before_pop-1, "");
1479 }
1480
1481 Node *SafePointNode::peek_monitor_box() const {
1482 int mon = jvms()->nof_monitors() - 1;
1483 assert(mon >= 0, "must have a monitor");
1484 return monitor_box(jvms(), mon);
1485 }
1486
1487 Node *SafePointNode::peek_monitor_obj() const {
1488 int mon = jvms()->nof_monitors() - 1;
1489 assert(mon >= 0, "must have a monitor");
1490 return monitor_obj(jvms(), mon);
1491 }
1492
1493 Node* SafePointNode::peek_operand(uint off) const {
1494 assert(jvms()->sp() > 0, "must have an operand");
1495 assert(off < jvms()->sp(), "off is out-of-range");
1496 return stack(jvms(), jvms()->sp() - off - 1);
1497 }
1498
1499 // Do we Match on this edge index or not? Match no edges
1500 uint SafePointNode::match_edge(uint idx) const {
1501 return (TypeFunc::Parms == idx);
1502 }
1503
1504 void SafePointNode::disconnect_from_root(PhaseIterGVN *igvn) {
1505 assert(Opcode() == Op_SafePoint, "only value for safepoint in loops");
1506 int nb = igvn->C->root()->find_prec_edge(this);
1507 if (nb != -1) {
1508 igvn->delete_precedence_of(igvn->C->root(), nb);
1509 }
1510 }
1511
1512 //============== SafePointScalarObjectNode ==============
1513
1514 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp, Node* alloc, uint first_index, uint depth, uint n_fields) :
1515 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1516 _first_index(first_index),
1517 _depth(depth),
1518 _n_fields(n_fields),
1519 _alloc(alloc)
1520 {
1521 #ifdef ASSERT
1522 if (!alloc->is_Allocate() && !(alloc->Opcode() == Op_VectorBox)) {
1523 alloc->dump();
1524 assert(false, "unexpected call node");
1525 }
1526 #endif
1527 init_class_id(Class_SafePointScalarObject);
1528 }
1529
1530 // Do not allow value-numbering for SafePointScalarObject node.
1531 uint SafePointScalarObjectNode::hash() const { return NO_HASH; }
1532 bool SafePointScalarObjectNode::cmp( const Node &n ) const {
1533 return (&n == this); // Always fail except on self
1534 }
1535
1536 uint SafePointScalarObjectNode::ideal_reg() const {
1537 return 0; // No matching to machine instruction
1538 }
1539
1540 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1541 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1542 }
1543
1544 const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1545 return RegMask::Empty;
1546 }
1547
1548 uint SafePointScalarObjectNode::match_edge(uint idx) const {
1549 return 0;
1550 }
1551
1552 SafePointScalarObjectNode*
1553 SafePointScalarObjectNode::clone(Dict* sosn_map, bool& new_node) const {
1554 void* cached = (*sosn_map)[(void*)this];
1555 if (cached != nullptr) {
1556 new_node = false;
1557 return (SafePointScalarObjectNode*)cached;
1558 }
1559 new_node = true;
1560 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1561 sosn_map->Insert((void*)this, (void*)res);
1562 return res;
1563 }
1564
1565
1566 #ifndef PRODUCT
1567 void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1568 st->print(" # fields@[%d..%d]", first_index(), first_index() + n_fields() - 1);
1569 }
1570 #endif
1571
1572 //============== SafePointScalarMergeNode ==============
1573
1574 SafePointScalarMergeNode::SafePointScalarMergeNode(const TypeOopPtr* tp, int merge_pointer_idx) :
1575 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1576 _merge_pointer_idx(merge_pointer_idx)
1577 {
1578 init_class_id(Class_SafePointScalarMerge);
1579 }
1580
1581 // Do not allow value-numbering for SafePointScalarMerge node.
1582 uint SafePointScalarMergeNode::hash() const { return NO_HASH; }
1583 bool SafePointScalarMergeNode::cmp( const Node &n ) const {
1584 return (&n == this); // Always fail except on self
1585 }
1586
1587 uint SafePointScalarMergeNode::ideal_reg() const {
1588 return 0; // No matching to machine instruction
1589 }
1590
1591 const RegMask &SafePointScalarMergeNode::in_RegMask(uint idx) const {
1592 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1593 }
1594
1595 const RegMask &SafePointScalarMergeNode::out_RegMask() const {
1596 return RegMask::Empty;
1597 }
1598
1599 uint SafePointScalarMergeNode::match_edge(uint idx) const {
1600 return 0;
1601 }
1602
1603 SafePointScalarMergeNode*
1604 SafePointScalarMergeNode::clone(Dict* sosn_map, bool& new_node) const {
1605 void* cached = (*sosn_map)[(void*)this];
1606 if (cached != nullptr) {
1607 new_node = false;
1608 return (SafePointScalarMergeNode*)cached;
1609 }
1610 new_node = true;
1611 SafePointScalarMergeNode* res = (SafePointScalarMergeNode*)Node::clone();
1612 sosn_map->Insert((void*)this, (void*)res);
1613 return res;
1614 }
1615
1616 #ifndef PRODUCT
1617 void SafePointScalarMergeNode::dump_spec(outputStream *st) const {
1618 st->print(" # merge_pointer_idx=%d, scalarized_objects=%d", _merge_pointer_idx, req()-1);
1619 }
1620 #endif
1621
1622 //=============================================================================
1623 uint AllocateNode::size_of() const { return sizeof(*this); }
1624
1625 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1626 Node *ctrl, Node *mem, Node *abio,
1627 Node *size, Node *klass_node, Node *initial_test)
1628 : CallNode(atype, nullptr, TypeRawPtr::BOTTOM)
1629 {
1630 init_class_id(Class_Allocate);
1631 init_flags(Flag_is_macro);
1632 _is_scalar_replaceable = false;
1633 _is_non_escaping = false;
1634 _is_allocation_MemBar_redundant = false;
1635 Node *topnode = C->top();
1636
1637 init_req( TypeFunc::Control , ctrl );
1638 init_req( TypeFunc::I_O , abio );
1639 init_req( TypeFunc::Memory , mem );
1640 init_req( TypeFunc::ReturnAdr, topnode );
1641 init_req( TypeFunc::FramePtr , topnode );
1642 init_req( AllocSize , size);
1643 init_req( KlassNode , klass_node);
1644 init_req( InitialTest , initial_test);
1645 init_req( ALength , topnode);
1646 init_req( ValidLengthTest , topnode);
1647 C->add_macro_node(this);
1648 }
1649
1650 void AllocateNode::compute_MemBar_redundancy(ciMethod* initializer)
1651 {
1652 assert(initializer != nullptr && initializer->is_object_initializer(),
1653 "unexpected initializer method");
1654 BCEscapeAnalyzer* analyzer = initializer->get_bcea();
1655 if (analyzer == nullptr) {
1656 return;
1657 }
1658
1659 // Allocation node is first parameter in its initializer
1660 if (analyzer->is_arg_stack(0) || analyzer->is_arg_local(0)) {
1661 _is_allocation_MemBar_redundant = true;
1662 }
1663 }
1664 Node *AllocateNode::make_ideal_mark(PhaseGVN *phase, Node* obj, Node* control, Node* mem) {
1665 Node* mark_node = nullptr;
1666 if (UseCompactObjectHeaders) {
1667 Node* klass_node = in(AllocateNode::KlassNode);
1668 Node* proto_adr = phase->transform(new AddPNode(klass_node, klass_node, phase->MakeConX(in_bytes(Klass::prototype_header_offset()))));
1669 mark_node = LoadNode::make(*phase, control, mem, proto_adr, TypeRawPtr::BOTTOM, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
1670 } else {
1671 // For now only enable fast locking for non-array types
1672 mark_node = phase->MakeConX(markWord::prototype().value());
1673 }
1674 return mark_node;
1675 }
1676
1677 // Retrieve the length from the AllocateArrayNode. Narrow the type with a
1678 // CastII, if appropriate. If we are not allowed to create new nodes, and
1679 // a CastII is appropriate, return null.
1680 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseValues* phase, bool allow_new_nodes) {
1681 Node *length = in(AllocateNode::ALength);
1682 assert(length != nullptr, "length is not null");
1683
1684 const TypeInt* length_type = phase->find_int_type(length);
1685 const TypeAryPtr* ary_type = oop_type->isa_aryptr();
1686
1687 if (ary_type != nullptr && length_type != nullptr) {
1688 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type);
1689 if (narrow_length_type != length_type) {
1690 // Assert one of:
1691 // - the narrow_length is 0
1692 // - the narrow_length is not wider than length
1693 assert(narrow_length_type == TypeInt::ZERO ||
1694 (length_type->is_con() && narrow_length_type->is_con() &&
1695 (narrow_length_type->_hi <= length_type->_lo)) ||
1696 (narrow_length_type->_hi <= length_type->_hi &&
1697 narrow_length_type->_lo >= length_type->_lo),
1698 "narrow type must be narrower than length type");
1699
1700 // Return null if new nodes are not allowed
1701 if (!allow_new_nodes) {
1702 return nullptr;
1703 }
1704 // Create a cast which is control dependent on the initialization to
1705 // propagate the fact that the array length must be positive.
1706 InitializeNode* init = initialization();
1707 if (init != nullptr) {
1708 length = new CastIINode(init->proj_out_or_null(TypeFunc::Control), length, narrow_length_type);
1709 }
1710 }
1711 }
1712
1713 return length;
1714 }
1715
1716 //=============================================================================
1717 const TypeFunc* LockNode::_lock_type_Type = nullptr;
1718
1719 uint LockNode::size_of() const { return sizeof(*this); }
1720
1721 // Redundant lock elimination
1722 //
1723 // There are various patterns of locking where we release and
1724 // immediately reacquire a lock in a piece of code where no operations
1725 // occur in between that would be observable. In those cases we can
1726 // skip releasing and reacquiring the lock without violating any
1727 // fairness requirements. Doing this around a loop could cause a lock
1728 // to be held for a very long time so we concentrate on non-looping
1729 // control flow. We also require that the operations are fully
1730 // redundant meaning that we don't introduce new lock operations on
1731 // some paths so to be able to eliminate it on others ala PRE. This
1732 // would probably require some more extensive graph manipulation to
1733 // guarantee that the memory edges were all handled correctly.
1734 //
1735 // Assuming p is a simple predicate which can't trap in any way and s
1736 // is a synchronized method consider this code:
1737 //
1738 // s();
1739 // if (p)
1740 // s();
1741 // else
1742 // s();
1743 // s();
1744 //
1745 // 1. The unlocks of the first call to s can be eliminated if the
1746 // locks inside the then and else branches are eliminated.
1747 //
1748 // 2. The unlocks of the then and else branches can be eliminated if
1749 // the lock of the final call to s is eliminated.
1750 //
1751 // Either of these cases subsumes the simple case of sequential control flow
1752 //
1753 // Additionally we can eliminate versions without the else case:
1754 //
1755 // s();
1756 // if (p)
1757 // s();
1758 // s();
1759 //
1760 // 3. In this case we eliminate the unlock of the first s, the lock
1761 // and unlock in the then case and the lock in the final s.
1762 //
1763 // Note also that in all these cases the then/else pieces don't have
1764 // to be trivial as long as they begin and end with synchronization
1765 // operations.
1766 //
1767 // s();
1768 // if (p)
1769 // s();
1770 // f();
1771 // s();
1772 // s();
1773 //
1774 // The code will work properly for this case, leaving in the unlock
1775 // before the call to f and the relock after it.
1776 //
1777 // A potentially interesting case which isn't handled here is when the
1778 // locking is partially redundant.
1779 //
1780 // s();
1781 // if (p)
1782 // s();
1783 //
1784 // This could be eliminated putting unlocking on the else case and
1785 // eliminating the first unlock and the lock in the then side.
1786 // Alternatively the unlock could be moved out of the then side so it
1787 // was after the merge and the first unlock and second lock
1788 // eliminated. This might require less manipulation of the memory
1789 // state to get correct.
1790 //
1791 // Additionally we might allow work between a unlock and lock before
1792 // giving up eliminating the locks. The current code disallows any
1793 // conditional control flow between these operations. A formulation
1794 // similar to partial redundancy elimination computing the
1795 // availability of unlocking and the anticipatability of locking at a
1796 // program point would allow detection of fully redundant locking with
1797 // some amount of work in between. I'm not sure how often I really
1798 // think that would occur though. Most of the cases I've seen
1799 // indicate it's likely non-trivial work would occur in between.
1800 // There may be other more complicated constructs where we could
1801 // eliminate locking but I haven't seen any others appear as hot or
1802 // interesting.
1803 //
1804 // Locking and unlocking have a canonical form in ideal that looks
1805 // roughly like this:
1806 //
1807 // <obj>
1808 // | \\------+
1809 // | \ \
1810 // | BoxLock \
1811 // | | | \
1812 // | | \ \
1813 // | | FastLock
1814 // | | /
1815 // | | /
1816 // | | |
1817 //
1818 // Lock
1819 // |
1820 // Proj #0
1821 // |
1822 // MembarAcquire
1823 // |
1824 // Proj #0
1825 //
1826 // MembarRelease
1827 // |
1828 // Proj #0
1829 // |
1830 // Unlock
1831 // |
1832 // Proj #0
1833 //
1834 //
1835 // This code proceeds by processing Lock nodes during PhaseIterGVN
1836 // and searching back through its control for the proper code
1837 // patterns. Once it finds a set of lock and unlock operations to
1838 // eliminate they are marked as eliminatable which causes the
1839 // expansion of the Lock and Unlock macro nodes to make the operation a NOP
1840 //
1841 //=============================================================================
1842
1843 //
1844 // Utility function to skip over uninteresting control nodes. Nodes skipped are:
1845 // - copy regions. (These may not have been optimized away yet.)
1846 // - eliminated locking nodes
1847 //
1848 static Node *next_control(Node *ctrl) {
1849 if (ctrl == nullptr)
1850 return nullptr;
1851 while (1) {
1852 if (ctrl->is_Region()) {
1853 RegionNode *r = ctrl->as_Region();
1854 Node *n = r->is_copy();
1855 if (n == nullptr)
1856 break; // hit a region, return it
1857 else
1858 ctrl = n;
1859 } else if (ctrl->is_Proj()) {
1860 Node *in0 = ctrl->in(0);
1861 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
1862 ctrl = in0->in(0);
1863 } else {
1864 break;
1865 }
1866 } else {
1867 break; // found an interesting control
1868 }
1869 }
1870 return ctrl;
1871 }
1872 //
1873 // Given a control, see if it's the control projection of an Unlock which
1874 // operating on the same object as lock.
1875 //
1876 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
1877 GrowableArray<AbstractLockNode*> &lock_ops) {
1878 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : nullptr;
1879 if (ctrl_proj != nullptr && ctrl_proj->_con == TypeFunc::Control) {
1880 Node *n = ctrl_proj->in(0);
1881 if (n != nullptr && n->is_Unlock()) {
1882 UnlockNode *unlock = n->as_Unlock();
1883 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1884 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1885 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
1886 if (lock_obj->eqv_uncast(unlock_obj) &&
1887 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) &&
1888 !unlock->is_eliminated()) {
1889 lock_ops.append(unlock);
1890 return true;
1891 }
1892 }
1893 }
1894 return false;
1895 }
1896
1897 //
1898 // Find the lock matching an unlock. Returns null if a safepoint
1899 // or complicated control is encountered first.
1900 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
1901 LockNode *lock_result = nullptr;
1902 // find the matching lock, or an intervening safepoint
1903 Node *ctrl = next_control(unlock->in(0));
1904 while (1) {
1905 assert(ctrl != nullptr, "invalid control graph");
1906 assert(!ctrl->is_Start(), "missing lock for unlock");
1907 if (ctrl->is_top()) break; // dead control path
1908 if (ctrl->is_Proj()) ctrl = ctrl->in(0);
1909 if (ctrl->is_SafePoint()) {
1910 break; // found a safepoint (may be the lock we are searching for)
1911 } else if (ctrl->is_Region()) {
1912 // Check for a simple diamond pattern. Punt on anything more complicated
1913 if (ctrl->req() == 3 && ctrl->in(1) != nullptr && ctrl->in(2) != nullptr) {
1914 Node *in1 = next_control(ctrl->in(1));
1915 Node *in2 = next_control(ctrl->in(2));
1916 if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
1917 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
1918 ctrl = next_control(in1->in(0)->in(0));
1919 } else {
1920 break;
1921 }
1922 } else {
1923 break;
1924 }
1925 } else {
1926 ctrl = next_control(ctrl->in(0)); // keep searching
1927 }
1928 }
1929 if (ctrl->is_Lock()) {
1930 LockNode *lock = ctrl->as_Lock();
1931 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1932 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1933 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
1934 if (lock_obj->eqv_uncast(unlock_obj) &&
1935 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) {
1936 lock_result = lock;
1937 }
1938 }
1939 return lock_result;
1940 }
1941
1942 // This code corresponds to case 3 above.
1943
1944 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
1945 GrowableArray<AbstractLockNode*> &lock_ops) {
1946 Node* if_node = node->in(0);
1947 bool if_true = node->is_IfTrue();
1948
1949 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
1950 Node *lock_ctrl = next_control(if_node->in(0));
1951 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
1952 Node* lock1_node = nullptr;
1953 ProjNode* proj = if_node->as_If()->proj_out(!if_true);
1954 if (if_true) {
1955 if (proj->is_IfFalse() && proj->outcnt() == 1) {
1956 lock1_node = proj->unique_out();
1957 }
1958 } else {
1959 if (proj->is_IfTrue() && proj->outcnt() == 1) {
1960 lock1_node = proj->unique_out();
1961 }
1962 }
1963 if (lock1_node != nullptr && lock1_node->is_Lock()) {
1964 LockNode *lock1 = lock1_node->as_Lock();
1965 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1966 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1967 Node* lock1_obj = bs->step_over_gc_barrier(lock1->obj_node());
1968 if (lock_obj->eqv_uncast(lock1_obj) &&
1969 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) &&
1970 !lock1->is_eliminated()) {
1971 lock_ops.append(lock1);
1972 return true;
1973 }
1974 }
1975 }
1976 }
1977
1978 lock_ops.trunc_to(0);
1979 return false;
1980 }
1981
1982 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
1983 GrowableArray<AbstractLockNode*> &lock_ops) {
1984 // check each control merging at this point for a matching unlock.
1985 // in(0) should be self edge so skip it.
1986 for (int i = 1; i < (int)region->req(); i++) {
1987 Node *in_node = next_control(region->in(i));
1988 if (in_node != nullptr) {
1989 if (find_matching_unlock(in_node, lock, lock_ops)) {
1990 // found a match so keep on checking.
1991 continue;
1992 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
1993 continue;
1994 }
1995
1996 // If we fall through to here then it was some kind of node we
1997 // don't understand or there wasn't a matching unlock, so give
1998 // up trying to merge locks.
1999 lock_ops.trunc_to(0);
2000 return false;
2001 }
2002 }
2003 return true;
2004
2005 }
2006
2007 // Check that all locks/unlocks associated with object come from balanced regions.
2008 bool AbstractLockNode::is_balanced() {
2009 Node* obj = obj_node();
2010 for (uint j = 0; j < obj->outcnt(); j++) {
2011 Node* n = obj->raw_out(j);
2012 if (n->is_AbstractLock() &&
2013 n->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2014 BoxLockNode* n_box = n->as_AbstractLock()->box_node()->as_BoxLock();
2015 if (n_box->is_unbalanced()) {
2016 return false;
2017 }
2018 }
2019 }
2020 return true;
2021 }
2022
2023 const char* AbstractLockNode::_kind_names[] = {"Regular", "NonEscObj", "Coarsened", "Nested"};
2024
2025 const char * AbstractLockNode::kind_as_string() const {
2026 return _kind_names[_kind];
2027 }
2028
2029 #ifndef PRODUCT
2030 //
2031 // Create a counter which counts the number of times this lock is acquired
2032 //
2033 void AbstractLockNode::create_lock_counter(JVMState* state) {
2034 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
2035 }
2036
2037 void AbstractLockNode::set_eliminated_lock_counter() {
2038 if (_counter) {
2039 // Update the counter to indicate that this lock was eliminated.
2040 // The counter update code will stay around even though the
2041 // optimizer will eliminate the lock operation itself.
2042 _counter->set_tag(NamedCounter::EliminatedLockCounter);
2043 }
2044 }
2045
2046 void AbstractLockNode::dump_spec(outputStream* st) const {
2047 st->print("%s ", _kind_names[_kind]);
2048 CallNode::dump_spec(st);
2049 }
2050
2051 void AbstractLockNode::dump_compact_spec(outputStream* st) const {
2052 st->print("%s", _kind_names[_kind]);
2053 }
2054 #endif
2055
2056 //=============================================================================
2057 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2058
2059 // perform any generic optimizations first (returns 'this' or null)
2060 Node *result = SafePointNode::Ideal(phase, can_reshape);
2061 if (result != nullptr) return result;
2062 // Don't bother trying to transform a dead node
2063 if (in(0) && in(0)->is_top()) return nullptr;
2064
2065 // Now see if we can optimize away this lock. We don't actually
2066 // remove the locking here, we simply set the _eliminate flag which
2067 // prevents macro expansion from expanding the lock. Since we don't
2068 // modify the graph, the value returned from this function is the
2069 // one computed above.
2070 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
2071 //
2072 // If we are locking an non-escaped object, the lock/unlock is unnecessary
2073 //
2074 ConnectionGraph *cgr = phase->C->congraph();
2075 if (cgr != nullptr && cgr->can_eliminate_lock(this)) {
2076 assert(!is_eliminated() || is_coarsened(), "sanity");
2077 // The lock could be marked eliminated by lock coarsening
2078 // code during first IGVN before EA. Replace coarsened flag
2079 // to eliminate all associated locks/unlocks.
2080 #ifdef ASSERT
2081 this->log_lock_optimization(phase->C,"eliminate_lock_set_non_esc1");
2082 #endif
2083 this->set_non_esc_obj();
2084 return result;
2085 }
2086
2087 if (!phase->C->do_locks_coarsening()) {
2088 return result; // Compiling without locks coarsening
2089 }
2090 //
2091 // Try lock coarsening
2092 //
2093 PhaseIterGVN* iter = phase->is_IterGVN();
2094 if (iter != nullptr && !is_eliminated()) {
2095
2096 GrowableArray<AbstractLockNode*> lock_ops;
2097
2098 Node *ctrl = next_control(in(0));
2099
2100 // now search back for a matching Unlock
2101 if (find_matching_unlock(ctrl, this, lock_ops)) {
2102 // found an unlock directly preceding this lock. This is the
2103 // case of single unlock directly control dependent on a
2104 // single lock which is the trivial version of case 1 or 2.
2105 } else if (ctrl->is_Region() ) {
2106 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
2107 // found lock preceded by multiple unlocks along all paths
2108 // joining at this point which is case 3 in description above.
2109 }
2110 } else {
2111 // see if this lock comes from either half of an if and the
2112 // predecessors merges unlocks and the other half of the if
2113 // performs a lock.
2114 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
2115 // found unlock splitting to an if with locks on both branches.
2116 }
2117 }
2118
2119 if (lock_ops.length() > 0) {
2120 // add ourselves to the list of locks to be eliminated.
2121 lock_ops.append(this);
2122
2123 #ifndef PRODUCT
2124 if (PrintEliminateLocks) {
2125 int locks = 0;
2126 int unlocks = 0;
2127 if (Verbose) {
2128 tty->print_cr("=== Locks coarsening ===");
2129 tty->print("Obj: ");
2130 obj_node()->dump();
2131 }
2132 for (int i = 0; i < lock_ops.length(); i++) {
2133 AbstractLockNode* lock = lock_ops.at(i);
2134 if (lock->Opcode() == Op_Lock)
2135 locks++;
2136 else
2137 unlocks++;
2138 if (Verbose) {
2139 tty->print("Box %d: ", i);
2140 box_node()->dump();
2141 tty->print(" %d: ", i);
2142 lock->dump();
2143 }
2144 }
2145 tty->print_cr("=== Coarsened %d unlocks and %d locks", unlocks, locks);
2146 }
2147 #endif
2148
2149 // for each of the identified locks, mark them
2150 // as eliminatable
2151 for (int i = 0; i < lock_ops.length(); i++) {
2152 AbstractLockNode* lock = lock_ops.at(i);
2153
2154 // Mark it eliminated by coarsening and update any counters
2155 #ifdef ASSERT
2156 lock->log_lock_optimization(phase->C, "eliminate_lock_set_coarsened");
2157 #endif
2158 lock->set_coarsened();
2159 }
2160 // Record this coarsened group.
2161 phase->C->add_coarsened_locks(lock_ops);
2162 } else if (ctrl->is_Region() &&
2163 iter->_worklist.member(ctrl)) {
2164 // We weren't able to find any opportunities but the region this
2165 // lock is control dependent on hasn't been processed yet so put
2166 // this lock back on the worklist so we can check again once any
2167 // region simplification has occurred.
2168 iter->_worklist.push(this);
2169 }
2170 }
2171 }
2172
2173 return result;
2174 }
2175
2176 //=============================================================================
2177 bool LockNode::is_nested_lock_region() {
2178 return is_nested_lock_region(nullptr);
2179 }
2180
2181 // p is used for access to compilation log; no logging if null
2182 bool LockNode::is_nested_lock_region(Compile * c) {
2183 BoxLockNode* box = box_node()->as_BoxLock();
2184 int stk_slot = box->stack_slot();
2185 if (stk_slot <= 0) {
2186 #ifdef ASSERT
2187 this->log_lock_optimization(c, "eliminate_lock_INLR_1");
2188 #endif
2189 return false; // External lock or it is not Box (Phi node).
2190 }
2191
2192 // Ignore complex cases: merged locks or multiple locks.
2193 Node* obj = obj_node();
2194 LockNode* unique_lock = nullptr;
2195 Node* bad_lock = nullptr;
2196 if (!box->is_simple_lock_region(&unique_lock, obj, &bad_lock)) {
2197 #ifdef ASSERT
2198 this->log_lock_optimization(c, "eliminate_lock_INLR_2a", bad_lock);
2199 #endif
2200 return false;
2201 }
2202 if (unique_lock != this) {
2203 #ifdef ASSERT
2204 this->log_lock_optimization(c, "eliminate_lock_INLR_2b", (unique_lock != nullptr ? unique_lock : bad_lock));
2205 if (PrintEliminateLocks && Verbose) {
2206 tty->print_cr("=============== unique_lock != this ============");
2207 tty->print(" this: ");
2208 this->dump();
2209 tty->print(" box: ");
2210 box->dump();
2211 tty->print(" obj: ");
2212 obj->dump();
2213 if (unique_lock != nullptr) {
2214 tty->print(" unique_lock: ");
2215 unique_lock->dump();
2216 }
2217 if (bad_lock != nullptr) {
2218 tty->print(" bad_lock: ");
2219 bad_lock->dump();
2220 }
2221 tty->print_cr("===============");
2222 }
2223 #endif
2224 return false;
2225 }
2226
2227 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2228 obj = bs->step_over_gc_barrier(obj);
2229 // Look for external lock for the same object.
2230 SafePointNode* sfn = this->as_SafePoint();
2231 JVMState* youngest_jvms = sfn->jvms();
2232 int max_depth = youngest_jvms->depth();
2233 for (int depth = 1; depth <= max_depth; depth++) {
2234 JVMState* jvms = youngest_jvms->of_depth(depth);
2235 int num_mon = jvms->nof_monitors();
2236 // Loop over monitors
2237 for (int idx = 0; idx < num_mon; idx++) {
2238 Node* obj_node = sfn->monitor_obj(jvms, idx);
2239 obj_node = bs->step_over_gc_barrier(obj_node);
2240 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock();
2241 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) {
2242 box->set_nested();
2243 return true;
2244 }
2245 }
2246 }
2247 #ifdef ASSERT
2248 this->log_lock_optimization(c, "eliminate_lock_INLR_3");
2249 #endif
2250 return false;
2251 }
2252
2253 //=============================================================================
2254 uint UnlockNode::size_of() const { return sizeof(*this); }
2255
2256 //=============================================================================
2257 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2258
2259 // perform any generic optimizations first (returns 'this' or null)
2260 Node *result = SafePointNode::Ideal(phase, can_reshape);
2261 if (result != nullptr) return result;
2262 // Don't bother trying to transform a dead node
2263 if (in(0) && in(0)->is_top()) return nullptr;
2264
2265 // Now see if we can optimize away this unlock. We don't actually
2266 // remove the unlocking here, we simply set the _eliminate flag which
2267 // prevents macro expansion from expanding the unlock. Since we don't
2268 // modify the graph, the value returned from this function is the
2269 // one computed above.
2270 // Escape state is defined after Parse phase.
2271 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
2272 //
2273 // If we are unlocking an non-escaped object, the lock/unlock is unnecessary.
2274 //
2275 ConnectionGraph *cgr = phase->C->congraph();
2276 if (cgr != nullptr && cgr->can_eliminate_lock(this)) {
2277 assert(!is_eliminated() || is_coarsened(), "sanity");
2278 // The lock could be marked eliminated by lock coarsening
2279 // code during first IGVN before EA. Replace coarsened flag
2280 // to eliminate all associated locks/unlocks.
2281 #ifdef ASSERT
2282 this->log_lock_optimization(phase->C, "eliminate_lock_set_non_esc2");
2283 #endif
2284 this->set_non_esc_obj();
2285 }
2286 }
2287 return result;
2288 }
2289
2290 void AbstractLockNode::log_lock_optimization(Compile *C, const char * tag, Node* bad_lock) const {
2291 if (C == nullptr) {
2292 return;
2293 }
2294 CompileLog* log = C->log();
2295 if (log != nullptr) {
2296 Node* box = box_node();
2297 Node* obj = obj_node();
2298 int box_id = box != nullptr ? box->_idx : -1;
2299 int obj_id = obj != nullptr ? obj->_idx : -1;
2300
2301 log->begin_head("%s compile_id='%d' lock_id='%d' class='%s' kind='%s' box_id='%d' obj_id='%d' bad_id='%d'",
2302 tag, C->compile_id(), this->_idx,
2303 is_Unlock() ? "unlock" : is_Lock() ? "lock" : "?",
2304 kind_as_string(), box_id, obj_id, (bad_lock != nullptr ? bad_lock->_idx : -1));
2305 log->stamp();
2306 log->end_head();
2307 JVMState* p = is_Unlock() ? (as_Unlock()->dbg_jvms()) : jvms();
2308 while (p != nullptr) {
2309 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
2310 p = p->caller();
2311 }
2312 log->tail(tag);
2313 }
2314 }
2315
2316 bool CallNode::may_modify_arraycopy_helper(const TypeOopPtr* dest_t, const TypeOopPtr* t_oop, PhaseValues* phase) {
2317 if (dest_t->is_known_instance() && t_oop->is_known_instance()) {
2318 return dest_t->instance_id() == t_oop->instance_id();
2319 }
2320
2321 if (dest_t->isa_instptr() && !dest_t->is_instptr()->instance_klass()->equals(phase->C->env()->Object_klass())) {
2322 // clone
2323 if (t_oop->isa_aryptr()) {
2324 return false;
2325 }
2326 if (!t_oop->isa_instptr()) {
2327 return true;
2328 }
2329 if (dest_t->maybe_java_subtype_of(t_oop) || t_oop->maybe_java_subtype_of(dest_t)) {
2330 return true;
2331 }
2332 // unrelated
2333 return false;
2334 }
2335
2336 if (dest_t->isa_aryptr()) {
2337 // arraycopy or array clone
2338 if (t_oop->isa_instptr()) {
2339 return false;
2340 }
2341 if (!t_oop->isa_aryptr()) {
2342 return true;
2343 }
2344
2345 const Type* elem = dest_t->is_aryptr()->elem();
2346 if (elem == Type::BOTTOM) {
2347 // An array but we don't know what elements are
2348 return true;
2349 }
2350
2351 dest_t = dest_t->add_offset(Type::OffsetBot)->is_oopptr();
2352 uint dest_alias = phase->C->get_alias_index(dest_t);
2353 uint t_oop_alias = phase->C->get_alias_index(t_oop);
2354
2355 return dest_alias == t_oop_alias;
2356 }
2357
2358 return true;
2359 }