1 /*
2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "ci/bcEscapeAnalyzer.hpp"
28 #include "compiler/oopMap.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/c2/barrierSetC2.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/escape.hpp"
37 #include "opto/locknode.hpp"
38 #include "opto/machnode.hpp"
39 #include "opto/matcher.hpp"
40 #include "opto/parse.hpp"
41 #include "opto/regalloc.hpp"
42 #include "opto/regmask.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "runtime/sharedRuntime.hpp"
46 #include "utilities/powerOfTwo.hpp"
47 #include "code/vmreg.hpp"
48
49 // Portions of code courtesy of Clifford Click
50
51 // Optimization - Graph Style
52
53 //=============================================================================
54 uint StartNode::size_of() const { return sizeof(*this); }
55 bool StartNode::cmp( const Node &n ) const
56 { return _domain == ((StartNode&)n)._domain; }
57 const Type *StartNode::bottom_type() const { return _domain; }
58 const Type* StartNode::Value(PhaseGVN* phase) const { return _domain; }
59 #ifndef PRODUCT
60 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);}
61 void StartNode::dump_compact_spec(outputStream *st) const { /* empty */ }
62 #endif
63
64 //------------------------------Ideal------------------------------------------
65 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){
66 return remove_dead_region(phase, can_reshape) ? this : nullptr;
67 }
68
69 //------------------------------calling_convention-----------------------------
70 void StartNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
71 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt);
72 }
73
74 //------------------------------Registers--------------------------------------
75 const RegMask &StartNode::in_RegMask(uint) const {
76 return RegMask::Empty;
77 }
78
79 //------------------------------match------------------------------------------
80 // Construct projections for incoming parameters, and their RegMask info
81 Node *StartNode::match( const ProjNode *proj, const Matcher *match ) {
82 switch (proj->_con) {
83 case TypeFunc::Control:
84 case TypeFunc::I_O:
85 case TypeFunc::Memory:
86 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
87 case TypeFunc::FramePtr:
88 return new MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP);
89 case TypeFunc::ReturnAdr:
90 return new MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP);
91 case TypeFunc::Parms:
92 default: {
93 uint parm_num = proj->_con - TypeFunc::Parms;
94 const Type *t = _domain->field_at(proj->_con);
95 if (t->base() == Type::Half) // 2nd half of Longs and Doubles
96 return new ConNode(Type::TOP);
97 uint ideal_reg = t->ideal_reg();
98 RegMask &rm = match->_calling_convention_mask[parm_num];
99 return new MachProjNode(this,proj->_con,rm,ideal_reg);
100 }
101 }
102 return nullptr;
103 }
104
105 //------------------------------StartOSRNode----------------------------------
106 // The method start node for an on stack replacement adapter
107
108 //------------------------------osr_domain-----------------------------
109 const TypeTuple *StartOSRNode::osr_domain() {
110 const Type **fields = TypeTuple::fields(2);
111 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer
112
113 return TypeTuple::make(TypeFunc::Parms+1, fields);
114 }
115
116 //=============================================================================
117 const char * const ParmNode::names[TypeFunc::Parms+1] = {
118 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms"
119 };
120
121 #ifndef PRODUCT
122 void ParmNode::dump_spec(outputStream *st) const {
123 if( _con < TypeFunc::Parms ) {
124 st->print("%s", names[_con]);
125 } else {
126 st->print("Parm%d: ",_con-TypeFunc::Parms);
127 // Verbose and WizardMode dump bottom_type for all nodes
128 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st);
129 }
130 }
131
132 void ParmNode::dump_compact_spec(outputStream *st) const {
133 if (_con < TypeFunc::Parms) {
134 st->print("%s", names[_con]);
135 } else {
136 st->print("%d:", _con-TypeFunc::Parms);
137 // unconditionally dump bottom_type
138 bottom_type()->dump_on(st);
139 }
140 }
141 #endif
142
143 uint ParmNode::ideal_reg() const {
144 switch( _con ) {
145 case TypeFunc::Control : // fall through
146 case TypeFunc::I_O : // fall through
147 case TypeFunc::Memory : return 0;
148 case TypeFunc::FramePtr : // fall through
149 case TypeFunc::ReturnAdr: return Op_RegP;
150 default : assert( _con > TypeFunc::Parms, "" );
151 // fall through
152 case TypeFunc::Parms : {
153 // Type of argument being passed
154 const Type *t = in(0)->as_Start()->_domain->field_at(_con);
155 return t->ideal_reg();
156 }
157 }
158 ShouldNotReachHere();
159 return 0;
160 }
161
162 //=============================================================================
163 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) {
164 init_req(TypeFunc::Control,cntrl);
165 init_req(TypeFunc::I_O,i_o);
166 init_req(TypeFunc::Memory,memory);
167 init_req(TypeFunc::FramePtr,frameptr);
168 init_req(TypeFunc::ReturnAdr,retadr);
169 }
170
171 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){
172 return remove_dead_region(phase, can_reshape) ? this : nullptr;
173 }
174
175 const Type* ReturnNode::Value(PhaseGVN* phase) const {
176 return ( phase->type(in(TypeFunc::Control)) == Type::TOP)
177 ? Type::TOP
178 : Type::BOTTOM;
179 }
180
181 // Do we Match on this edge index or not? No edges on return nodes
182 uint ReturnNode::match_edge(uint idx) const {
183 return 0;
184 }
185
186
187 #ifndef PRODUCT
188 void ReturnNode::dump_req(outputStream *st, DumpConfig* dc) const {
189 // Dump the required inputs, after printing "returns"
190 uint i; // Exit value of loop
191 for (i = 0; i < req(); i++) { // For all required inputs
192 if (i == TypeFunc::Parms) st->print("returns ");
193 Node* p = in(i);
194 if (p != nullptr) {
195 p->dump_idx(false, st, dc);
196 st->print(" ");
197 } else {
198 st->print("_ ");
199 }
200 }
201 }
202 #endif
203
204 //=============================================================================
205 RethrowNode::RethrowNode(
206 Node* cntrl,
207 Node* i_o,
208 Node* memory,
209 Node* frameptr,
210 Node* ret_adr,
211 Node* exception
212 ) : Node(TypeFunc::Parms + 1) {
213 init_req(TypeFunc::Control , cntrl );
214 init_req(TypeFunc::I_O , i_o );
215 init_req(TypeFunc::Memory , memory );
216 init_req(TypeFunc::FramePtr , frameptr );
217 init_req(TypeFunc::ReturnAdr, ret_adr);
218 init_req(TypeFunc::Parms , exception);
219 }
220
221 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){
222 return remove_dead_region(phase, can_reshape) ? this : nullptr;
223 }
224
225 const Type* RethrowNode::Value(PhaseGVN* phase) const {
226 return (phase->type(in(TypeFunc::Control)) == Type::TOP)
227 ? Type::TOP
228 : Type::BOTTOM;
229 }
230
231 uint RethrowNode::match_edge(uint idx) const {
232 return 0;
233 }
234
235 #ifndef PRODUCT
236 void RethrowNode::dump_req(outputStream *st, DumpConfig* dc) const {
237 // Dump the required inputs, after printing "exception"
238 uint i; // Exit value of loop
239 for (i = 0; i < req(); i++) { // For all required inputs
240 if (i == TypeFunc::Parms) st->print("exception ");
241 Node* p = in(i);
242 if (p != nullptr) {
243 p->dump_idx(false, st, dc);
244 st->print(" ");
245 } else {
246 st->print("_ ");
247 }
248 }
249 }
250 #endif
251
252 //=============================================================================
253 // Do we Match on this edge index or not? Match only target address & method
254 uint TailCallNode::match_edge(uint idx) const {
255 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
256 }
257
258 //=============================================================================
259 // Do we Match on this edge index or not? Match only target address & oop
260 uint TailJumpNode::match_edge(uint idx) const {
261 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
262 }
263
264 //=============================================================================
265 JVMState::JVMState(ciMethod* method, JVMState* caller) :
266 _method(method) {
267 assert(method != nullptr, "must be valid call site");
268 _bci = InvocationEntryBci;
269 _reexecute = Reexecute_Undefined;
270 debug_only(_bci = -99); // random garbage value
271 debug_only(_map = (SafePointNode*)-1);
272 _caller = caller;
273 _depth = 1 + (caller == nullptr ? 0 : caller->depth());
274 _locoff = TypeFunc::Parms;
275 _stkoff = _locoff + _method->max_locals();
276 _monoff = _stkoff + _method->max_stack();
277 _scloff = _monoff;
278 _endoff = _monoff;
279 _sp = 0;
280 }
281 JVMState::JVMState(int stack_size) :
282 _method(nullptr) {
283 _bci = InvocationEntryBci;
284 _reexecute = Reexecute_Undefined;
285 debug_only(_map = (SafePointNode*)-1);
286 _caller = nullptr;
287 _depth = 1;
288 _locoff = TypeFunc::Parms;
289 _stkoff = _locoff;
290 _monoff = _stkoff + stack_size;
291 _scloff = _monoff;
292 _endoff = _monoff;
293 _sp = 0;
294 }
295
296 //--------------------------------of_depth-------------------------------------
297 JVMState* JVMState::of_depth(int d) const {
298 const JVMState* jvmp = this;
299 assert(0 < d && (uint)d <= depth(), "oob");
300 for (int skip = depth() - d; skip > 0; skip--) {
301 jvmp = jvmp->caller();
302 }
303 assert(jvmp->depth() == (uint)d, "found the right one");
304 return (JVMState*)jvmp;
305 }
306
307 //-----------------------------same_calls_as-----------------------------------
308 bool JVMState::same_calls_as(const JVMState* that) const {
309 if (this == that) return true;
310 if (this->depth() != that->depth()) return false;
311 const JVMState* p = this;
312 const JVMState* q = that;
313 for (;;) {
314 if (p->_method != q->_method) return false;
315 if (p->_method == nullptr) return true; // bci is irrelevant
316 if (p->_bci != q->_bci) return false;
317 if (p->_reexecute != q->_reexecute) return false;
318 p = p->caller();
319 q = q->caller();
320 if (p == q) return true;
321 assert(p != nullptr && q != nullptr, "depth check ensures we don't run off end");
322 }
323 }
324
325 //------------------------------debug_start------------------------------------
326 uint JVMState::debug_start() const {
327 debug_only(JVMState* jvmroot = of_depth(1));
328 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
329 return of_depth(1)->locoff();
330 }
331
332 //-------------------------------debug_end-------------------------------------
333 uint JVMState::debug_end() const {
334 debug_only(JVMState* jvmroot = of_depth(1));
335 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
336 return endoff();
337 }
338
339 //------------------------------debug_depth------------------------------------
340 uint JVMState::debug_depth() const {
341 uint total = 0;
342 for (const JVMState* jvmp = this; jvmp != nullptr; jvmp = jvmp->caller()) {
343 total += jvmp->debug_size();
344 }
345 return total;
346 }
347
348 #ifndef PRODUCT
349
350 //------------------------------format_helper----------------------------------
351 // Given an allocation (a Chaitin object) and a Node decide if the Node carries
352 // any defined value or not. If it does, print out the register or constant.
353 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
354 if (n == nullptr) { st->print(" null"); return; }
355 if (n->is_SafePointScalarObject()) {
356 // Scalar replacement.
357 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
358 scobjs->append_if_missing(spobj);
359 int sco_n = scobjs->find(spobj);
360 assert(sco_n >= 0, "");
361 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
362 return;
363 }
364 if (regalloc->node_regs_max_index() > 0 &&
365 OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
366 char buf[50];
367 regalloc->dump_register(n,buf,sizeof(buf));
368 st->print(" %s%d]=%s",msg,i,buf);
369 } else { // No register, but might be constant
370 const Type *t = n->bottom_type();
371 switch (t->base()) {
372 case Type::Int:
373 st->print(" %s%d]=#" INT32_FORMAT,msg,i,t->is_int()->get_con());
374 break;
375 case Type::AnyPtr:
376 assert( t == TypePtr::NULL_PTR || n->in_dump(), "" );
377 st->print(" %s%d]=#null",msg,i);
378 break;
379 case Type::AryPtr:
380 case Type::InstPtr:
381 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->isa_oopptr()->const_oop()));
382 break;
383 case Type::KlassPtr:
384 case Type::AryKlassPtr:
385 case Type::InstKlassPtr:
386 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_klassptr()->exact_klass()));
387 break;
388 case Type::MetadataPtr:
389 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_metadataptr()->metadata()));
390 break;
391 case Type::NarrowOop:
392 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_oopptr()->const_oop()));
393 break;
394 case Type::RawPtr:
395 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,p2i(t->is_rawptr()));
396 break;
397 case Type::DoubleCon:
398 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
399 break;
400 case Type::FloatCon:
401 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
402 break;
403 case Type::Long:
404 st->print(" %s%d]=#" INT64_FORMAT,msg,i,(int64_t)(t->is_long()->get_con()));
405 break;
406 case Type::Half:
407 case Type::Top:
408 st->print(" %s%d]=_",msg,i);
409 break;
410 default: ShouldNotReachHere();
411 }
412 }
413 }
414
415 //---------------------print_method_with_lineno--------------------------------
416 void JVMState::print_method_with_lineno(outputStream* st, bool show_name) const {
417 if (show_name) _method->print_short_name(st);
418
419 int lineno = _method->line_number_from_bci(_bci);
420 if (lineno != -1) {
421 st->print(" @ bci:%d (line %d)", _bci, lineno);
422 } else {
423 st->print(" @ bci:%d", _bci);
424 }
425 }
426
427 //------------------------------format-----------------------------------------
428 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
429 st->print(" #");
430 if (_method) {
431 print_method_with_lineno(st, true);
432 } else {
433 st->print_cr(" runtime stub ");
434 return;
435 }
436 if (n->is_MachSafePoint()) {
437 GrowableArray<SafePointScalarObjectNode*> scobjs;
438 MachSafePointNode *mcall = n->as_MachSafePoint();
439 uint i;
440 // Print locals
441 for (i = 0; i < (uint)loc_size(); i++)
442 format_helper(regalloc, st, mcall->local(this, i), "L[", i, &scobjs);
443 // Print stack
444 for (i = 0; i < (uint)stk_size(); i++) {
445 if ((uint)(_stkoff + i) >= mcall->len())
446 st->print(" oob ");
447 else
448 format_helper(regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs);
449 }
450 for (i = 0; (int)i < nof_monitors(); i++) {
451 Node *box = mcall->monitor_box(this, i);
452 Node *obj = mcall->monitor_obj(this, i);
453 if (regalloc->node_regs_max_index() > 0 &&
454 OptoReg::is_valid(regalloc->get_reg_first(box))) {
455 box = BoxLockNode::box_node(box);
456 format_helper(regalloc, st, box, "MON-BOX[", i, &scobjs);
457 } else {
458 OptoReg::Name box_reg = BoxLockNode::reg(box);
459 st->print(" MON-BOX%d=%s+%d",
460 i,
461 OptoReg::regname(OptoReg::c_frame_pointer),
462 regalloc->reg2offset(box_reg));
463 }
464 const char* obj_msg = "MON-OBJ[";
465 if (EliminateLocks) {
466 if (BoxLockNode::box_node(box)->is_eliminated())
467 obj_msg = "MON-OBJ(LOCK ELIMINATED)[";
468 }
469 format_helper(regalloc, st, obj, obj_msg, i, &scobjs);
470 }
471
472 for (i = 0; i < (uint)scobjs.length(); i++) {
473 // Scalar replaced objects.
474 st->cr();
475 st->print(" # ScObj" INT32_FORMAT " ", i);
476 SafePointScalarObjectNode* spobj = scobjs.at(i);
477 ciKlass* cik = spobj->bottom_type()->is_oopptr()->exact_klass();
478 assert(cik->is_instance_klass() ||
479 cik->is_array_klass(), "Not supported allocation.");
480 ciInstanceKlass *iklass = nullptr;
481 if (cik->is_instance_klass()) {
482 cik->print_name_on(st);
483 iklass = cik->as_instance_klass();
484 } else if (cik->is_type_array_klass()) {
485 cik->as_array_klass()->base_element_type()->print_name_on(st);
486 st->print("[%d]", spobj->n_fields());
487 } else if (cik->is_obj_array_klass()) {
488 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass();
489 if (cie->is_instance_klass()) {
490 cie->print_name_on(st);
491 } else if (cie->is_type_array_klass()) {
492 cie->as_array_klass()->base_element_type()->print_name_on(st);
493 } else {
494 ShouldNotReachHere();
495 }
496 st->print("[%d]", spobj->n_fields());
497 int ndim = cik->as_array_klass()->dimension() - 1;
498 while (ndim-- > 0) {
499 st->print("[]");
500 }
501 }
502 st->print("={");
503 uint nf = spobj->n_fields();
504 if (nf > 0) {
505 uint first_ind = spobj->first_index(mcall->jvms());
506 Node* fld_node = mcall->in(first_ind);
507 ciField* cifield;
508 if (iklass != nullptr) {
509 st->print(" [");
510 cifield = iklass->nonstatic_field_at(0);
511 cifield->print_name_on(st);
512 format_helper(regalloc, st, fld_node, ":", 0, &scobjs);
513 } else {
514 format_helper(regalloc, st, fld_node, "[", 0, &scobjs);
515 }
516 for (uint j = 1; j < nf; j++) {
517 fld_node = mcall->in(first_ind+j);
518 if (iklass != nullptr) {
519 st->print(", [");
520 cifield = iklass->nonstatic_field_at(j);
521 cifield->print_name_on(st);
522 format_helper(regalloc, st, fld_node, ":", j, &scobjs);
523 } else {
524 format_helper(regalloc, st, fld_node, ", [", j, &scobjs);
525 }
526 }
527 }
528 st->print(" }");
529 }
530 }
531 st->cr();
532 if (caller() != nullptr) caller()->format(regalloc, n, st);
533 }
534
535
536 void JVMState::dump_spec(outputStream *st) const {
537 if (_method != nullptr) {
538 bool printed = false;
539 if (!Verbose) {
540 // The JVMS dumps make really, really long lines.
541 // Take out the most boring parts, which are the package prefixes.
542 char buf[500];
543 stringStream namest(buf, sizeof(buf));
544 _method->print_short_name(&namest);
545 if (namest.count() < sizeof(buf)) {
546 const char* name = namest.base();
547 if (name[0] == ' ') ++name;
548 const char* endcn = strchr(name, ':'); // end of class name
549 if (endcn == nullptr) endcn = strchr(name, '(');
550 if (endcn == nullptr) endcn = name + strlen(name);
551 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
552 --endcn;
553 st->print(" %s", endcn);
554 printed = true;
555 }
556 }
557 print_method_with_lineno(st, !printed);
558 if(_reexecute == Reexecute_True)
559 st->print(" reexecute");
560 } else {
561 st->print(" runtime stub");
562 }
563 if (caller() != nullptr) caller()->dump_spec(st);
564 }
565
566
567 void JVMState::dump_on(outputStream* st) const {
568 bool print_map = _map && !((uintptr_t)_map & 1) &&
569 ((caller() == nullptr) || (caller()->map() != _map));
570 if (print_map) {
571 if (_map->len() > _map->req()) { // _map->has_exceptions()
572 Node* ex = _map->in(_map->req()); // _map->next_exception()
573 // skip the first one; it's already being printed
574 while (ex != nullptr && ex->len() > ex->req()) {
575 ex = ex->in(ex->req()); // ex->next_exception()
576 ex->dump(1);
577 }
578 }
579 _map->dump(Verbose ? 2 : 1);
580 }
581 if (caller() != nullptr) {
582 caller()->dump_on(st);
583 }
584 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=",
585 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false");
586 if (_method == nullptr) {
587 st->print_cr("(none)");
588 } else {
589 _method->print_name(st);
590 st->cr();
591 if (bci() >= 0 && bci() < _method->code_size()) {
592 st->print(" bc: ");
593 _method->print_codes_on(bci(), bci()+1, st);
594 }
595 }
596 }
597
598 // Extra way to dump a jvms from the debugger,
599 // to avoid a bug with C++ member function calls.
600 void dump_jvms(JVMState* jvms) {
601 jvms->dump();
602 }
603 #endif
604
605 //--------------------------clone_shallow--------------------------------------
606 JVMState* JVMState::clone_shallow(Compile* C) const {
607 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
608 n->set_bci(_bci);
609 n->_reexecute = _reexecute;
610 n->set_locoff(_locoff);
611 n->set_stkoff(_stkoff);
612 n->set_monoff(_monoff);
613 n->set_scloff(_scloff);
614 n->set_endoff(_endoff);
615 n->set_sp(_sp);
616 n->set_map(_map);
617 return n;
618 }
619
620 //---------------------------clone_deep----------------------------------------
621 JVMState* JVMState::clone_deep(Compile* C) const {
622 JVMState* n = clone_shallow(C);
623 for (JVMState* p = n; p->_caller != nullptr; p = p->_caller) {
624 p->_caller = p->_caller->clone_shallow(C);
625 }
626 assert(n->depth() == depth(), "sanity");
627 assert(n->debug_depth() == debug_depth(), "sanity");
628 return n;
629 }
630
631 /**
632 * Reset map for all callers
633 */
634 void JVMState::set_map_deep(SafePointNode* map) {
635 for (JVMState* p = this; p != nullptr; p = p->_caller) {
636 p->set_map(map);
637 }
638 }
639
640 // unlike set_map(), this is two-way setting.
641 void JVMState::bind_map(SafePointNode* map) {
642 set_map(map);
643 _map->set_jvms(this);
644 }
645
646 // Adapt offsets in in-array after adding or removing an edge.
647 // Prerequisite is that the JVMState is used by only one node.
648 void JVMState::adapt_position(int delta) {
649 for (JVMState* jvms = this; jvms != nullptr; jvms = jvms->caller()) {
650 jvms->set_locoff(jvms->locoff() + delta);
651 jvms->set_stkoff(jvms->stkoff() + delta);
652 jvms->set_monoff(jvms->monoff() + delta);
653 jvms->set_scloff(jvms->scloff() + delta);
654 jvms->set_endoff(jvms->endoff() + delta);
655 }
656 }
657
658 // Mirror the stack size calculation in the deopt code
659 // How much stack space would we need at this point in the program in
660 // case of deoptimization?
661 int JVMState::interpreter_frame_size() const {
662 const JVMState* jvms = this;
663 int size = 0;
664 int callee_parameters = 0;
665 int callee_locals = 0;
666 int extra_args = method()->max_stack() - stk_size();
667
668 while (jvms != nullptr) {
669 int locks = jvms->nof_monitors();
670 int temps = jvms->stk_size();
671 bool is_top_frame = (jvms == this);
672 ciMethod* method = jvms->method();
673
674 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(),
675 temps + callee_parameters,
676 extra_args,
677 locks,
678 callee_parameters,
679 callee_locals,
680 is_top_frame);
681 size += frame_size;
682
683 callee_parameters = method->size_of_parameters();
684 callee_locals = method->max_locals();
685 extra_args = 0;
686 jvms = jvms->caller();
687 }
688 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord;
689 }
690
691 //=============================================================================
692 bool CallNode::cmp( const Node &n ) const
693 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
694 #ifndef PRODUCT
695 void CallNode::dump_req(outputStream *st, DumpConfig* dc) const {
696 // Dump the required inputs, enclosed in '(' and ')'
697 uint i; // Exit value of loop
698 for (i = 0; i < req(); i++) { // For all required inputs
699 if (i == TypeFunc::Parms) st->print("(");
700 Node* p = in(i);
701 if (p != nullptr) {
702 p->dump_idx(false, st, dc);
703 st->print(" ");
704 } else {
705 st->print("_ ");
706 }
707 }
708 st->print(")");
709 }
710
711 void CallNode::dump_spec(outputStream *st) const {
712 st->print(" ");
713 if (tf() != nullptr) tf()->dump_on(st);
714 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt);
715 if (jvms() != nullptr) jvms()->dump_spec(st);
716 }
717 #endif
718
719 const Type *CallNode::bottom_type() const { return tf()->range(); }
720 const Type* CallNode::Value(PhaseGVN* phase) const {
721 if (phase->type(in(0)) == Type::TOP) return Type::TOP;
722 return tf()->range();
723 }
724
725 //------------------------------calling_convention-----------------------------
726 void CallNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
727 // Use the standard compiler calling convention
728 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt);
729 }
730
731
732 //------------------------------match------------------------------------------
733 // Construct projections for control, I/O, memory-fields, ..., and
734 // return result(s) along with their RegMask info
735 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
736 switch (proj->_con) {
737 case TypeFunc::Control:
738 case TypeFunc::I_O:
739 case TypeFunc::Memory:
740 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
741
742 case TypeFunc::Parms+1: // For LONG & DOUBLE returns
743 assert(tf()->range()->field_at(TypeFunc::Parms+1) == Type::HALF, "");
744 // 2nd half of doubles and longs
745 return new MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad);
746
747 case TypeFunc::Parms: { // Normal returns
748 uint ideal_reg = tf()->range()->field_at(TypeFunc::Parms)->ideal_reg();
749 OptoRegPair regs = Opcode() == Op_CallLeafVector
750 ? match->vector_return_value(ideal_reg) // Calls into assembly vector routine
751 : is_CallRuntime()
752 ? match->c_return_value(ideal_reg) // Calls into C runtime
753 : match-> return_value(ideal_reg); // Calls into compiled Java code
754 RegMask rm = RegMask(regs.first());
755
756 if (Opcode() == Op_CallLeafVector) {
757 // If the return is in vector, compute appropriate regmask taking into account the whole range
758 if(ideal_reg >= Op_VecS && ideal_reg <= Op_VecZ) {
759 if(OptoReg::is_valid(regs.second())) {
760 for (OptoReg::Name r = regs.first(); r <= regs.second(); r = OptoReg::add(r, 1)) {
761 rm.Insert(r);
762 }
763 }
764 }
765 }
766
767 if( OptoReg::is_valid(regs.second()) )
768 rm.Insert( regs.second() );
769 return new MachProjNode(this,proj->_con,rm,ideal_reg);
770 }
771
772 case TypeFunc::ReturnAdr:
773 case TypeFunc::FramePtr:
774 default:
775 ShouldNotReachHere();
776 }
777 return nullptr;
778 }
779
780 // Do we Match on this edge index or not? Match no edges
781 uint CallNode::match_edge(uint idx) const {
782 return 0;
783 }
784
785 //
786 // Determine whether the call could modify the field of the specified
787 // instance at the specified offset.
788 //
789 bool CallNode::may_modify(const TypeOopPtr* t_oop, PhaseValues* phase) {
790 assert((t_oop != nullptr), "sanity");
791 if (is_call_to_arraycopystub() && strcmp(_name, "unsafe_arraycopy") != 0) {
792 const TypeTuple* args = _tf->domain();
793 Node* dest = nullptr;
794 // Stubs that can be called once an ArrayCopyNode is expanded have
795 // different signatures. Look for the second pointer argument,
796 // that is the destination of the copy.
797 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
798 if (args->field_at(i)->isa_ptr()) {
799 j++;
800 if (j == 2) {
801 dest = in(i);
802 break;
803 }
804 }
805 }
806 guarantee(dest != nullptr, "Call had only one ptr in, broken IR!");
807 if (!dest->is_top() && may_modify_arraycopy_helper(phase->type(dest)->is_oopptr(), t_oop, phase)) {
808 return true;
809 }
810 return false;
811 }
812 if (t_oop->is_known_instance()) {
813 // The instance_id is set only for scalar-replaceable allocations which
814 // are not passed as arguments according to Escape Analysis.
815 return false;
816 }
817 if (t_oop->is_ptr_to_boxed_value()) {
818 ciKlass* boxing_klass = t_oop->is_instptr()->instance_klass();
819 if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) {
820 // Skip unrelated boxing methods.
821 Node* proj = proj_out_or_null(TypeFunc::Parms);
822 if ((proj == nullptr) || (phase->type(proj)->is_instptr()->instance_klass() != boxing_klass)) {
823 return false;
824 }
825 }
826 if (is_CallJava() && as_CallJava()->method() != nullptr) {
827 ciMethod* meth = as_CallJava()->method();
828 if (meth->is_getter()) {
829 return false;
830 }
831 // May modify (by reflection) if an boxing object is passed
832 // as argument or returned.
833 Node* proj = returns_pointer() ? proj_out_or_null(TypeFunc::Parms) : nullptr;
834 if (proj != nullptr) {
835 const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr();
836 if ((inst_t != nullptr) && (!inst_t->klass_is_exact() ||
837 (inst_t->instance_klass() == boxing_klass))) {
838 return true;
839 }
840 }
841 const TypeTuple* d = tf()->domain();
842 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
843 const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr();
844 if ((inst_t != nullptr) && (!inst_t->klass_is_exact() ||
845 (inst_t->instance_klass() == boxing_klass))) {
846 return true;
847 }
848 }
849 return false;
850 }
851 }
852 return true;
853 }
854
855 // Does this call have a direct reference to n other than debug information?
856 bool CallNode::has_non_debug_use(Node *n) {
857 const TypeTuple * d = tf()->domain();
858 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
859 Node *arg = in(i);
860 if (arg == n) {
861 return true;
862 }
863 }
864 return false;
865 }
866
867 // Returns the unique CheckCastPP of a call
868 // or 'this' if there are several CheckCastPP or unexpected uses
869 // or returns null if there is no one.
870 Node *CallNode::result_cast() {
871 Node *cast = nullptr;
872
873 Node *p = proj_out_or_null(TypeFunc::Parms);
874 if (p == nullptr)
875 return nullptr;
876
877 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
878 Node *use = p->fast_out(i);
879 if (use->is_CheckCastPP()) {
880 if (cast != nullptr) {
881 return this; // more than 1 CheckCastPP
882 }
883 cast = use;
884 } else if (!use->is_Initialize() &&
885 !use->is_AddP() &&
886 use->Opcode() != Op_MemBarStoreStore) {
887 // Expected uses are restricted to a CheckCastPP, an Initialize
888 // node, a MemBarStoreStore (clone) and AddP nodes. If we
889 // encounter any other use (a Phi node can be seen in rare
890 // cases) return this to prevent incorrect optimizations.
891 return this;
892 }
893 }
894 return cast;
895 }
896
897
898 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts) {
899 projs->fallthrough_proj = nullptr;
900 projs->fallthrough_catchproj = nullptr;
901 projs->fallthrough_ioproj = nullptr;
902 projs->catchall_ioproj = nullptr;
903 projs->catchall_catchproj = nullptr;
904 projs->fallthrough_memproj = nullptr;
905 projs->catchall_memproj = nullptr;
906 projs->resproj = nullptr;
907 projs->exobj = nullptr;
908
909 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
910 ProjNode *pn = fast_out(i)->as_Proj();
911 if (pn->outcnt() == 0) continue;
912 switch (pn->_con) {
913 case TypeFunc::Control:
914 {
915 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
916 projs->fallthrough_proj = pn;
917 const Node* cn = pn->unique_ctrl_out_or_null();
918 if (cn != nullptr && cn->is_Catch()) {
919 ProjNode *cpn = nullptr;
920 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
921 cpn = cn->fast_out(k)->as_Proj();
922 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
923 if (cpn->_con == CatchProjNode::fall_through_index)
924 projs->fallthrough_catchproj = cpn;
925 else {
926 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
927 projs->catchall_catchproj = cpn;
928 }
929 }
930 }
931 break;
932 }
933 case TypeFunc::I_O:
934 if (pn->_is_io_use)
935 projs->catchall_ioproj = pn;
936 else
937 projs->fallthrough_ioproj = pn;
938 for (DUIterator j = pn->outs(); pn->has_out(j); j++) {
939 Node* e = pn->out(j);
940 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) {
941 assert(projs->exobj == nullptr, "only one");
942 projs->exobj = e;
943 }
944 }
945 break;
946 case TypeFunc::Memory:
947 if (pn->_is_io_use)
948 projs->catchall_memproj = pn;
949 else
950 projs->fallthrough_memproj = pn;
951 break;
952 case TypeFunc::Parms:
953 projs->resproj = pn;
954 break;
955 default:
956 assert(false, "unexpected projection from allocation node.");
957 }
958 }
959
960 // The resproj may not exist because the result could be ignored
961 // and the exception object may not exist if an exception handler
962 // swallows the exception but all the other must exist and be found.
963 assert(projs->fallthrough_proj != nullptr, "must be found");
964 do_asserts = do_asserts && !Compile::current()->inlining_incrementally();
965 assert(!do_asserts || projs->fallthrough_catchproj != nullptr, "must be found");
966 assert(!do_asserts || projs->fallthrough_memproj != nullptr, "must be found");
967 assert(!do_asserts || projs->fallthrough_ioproj != nullptr, "must be found");
968 assert(!do_asserts || projs->catchall_catchproj != nullptr, "must be found");
969 if (separate_io_proj) {
970 assert(!do_asserts || projs->catchall_memproj != nullptr, "must be found");
971 assert(!do_asserts || projs->catchall_ioproj != nullptr, "must be found");
972 }
973 }
974
975 Node* CallNode::Ideal(PhaseGVN* phase, bool can_reshape) {
976 #ifdef ASSERT
977 // Validate attached generator
978 CallGenerator* cg = generator();
979 if (cg != nullptr) {
980 assert(is_CallStaticJava() && cg->is_mh_late_inline() ||
981 is_CallDynamicJava() && cg->is_virtual_late_inline(), "mismatch");
982 }
983 #endif // ASSERT
984 return SafePointNode::Ideal(phase, can_reshape);
985 }
986
987 bool CallNode::is_call_to_arraycopystub() const {
988 if (_name != nullptr && strstr(_name, "arraycopy") != 0) {
989 return true;
990 }
991 return false;
992 }
993
994 //=============================================================================
995 uint CallJavaNode::size_of() const { return sizeof(*this); }
996 bool CallJavaNode::cmp( const Node &n ) const {
997 CallJavaNode &call = (CallJavaNode&)n;
998 return CallNode::cmp(call) && _method == call._method &&
999 _override_symbolic_info == call._override_symbolic_info;
1000 }
1001
1002 void CallJavaNode::copy_call_debug_info(PhaseIterGVN* phase, SafePointNode* sfpt) {
1003 // Copy debug information and adjust JVMState information
1004 uint old_dbg_start = sfpt->is_Call() ? sfpt->as_Call()->tf()->domain()->cnt() : (uint)TypeFunc::Parms+1;
1005 uint new_dbg_start = tf()->domain()->cnt();
1006 int jvms_adj = new_dbg_start - old_dbg_start;
1007 assert (new_dbg_start == req(), "argument count mismatch");
1008 Compile* C = phase->C;
1009
1010 // SafePointScalarObject node could be referenced several times in debug info.
1011 // Use Dict to record cloned nodes.
1012 Dict* sosn_map = new Dict(cmpkey,hashkey);
1013 for (uint i = old_dbg_start; i < sfpt->req(); i++) {
1014 Node* old_in = sfpt->in(i);
1015 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
1016 if (old_in != nullptr && old_in->is_SafePointScalarObject()) {
1017 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
1018 bool new_node;
1019 Node* new_in = old_sosn->clone(sosn_map, new_node);
1020 if (new_node) { // New node?
1021 new_in->set_req(0, C->root()); // reset control edge
1022 new_in = phase->transform(new_in); // Register new node.
1023 }
1024 old_in = new_in;
1025 }
1026 add_req(old_in);
1027 }
1028
1029 // JVMS may be shared so clone it before we modify it
1030 set_jvms(sfpt->jvms() != nullptr ? sfpt->jvms()->clone_deep(C) : nullptr);
1031 for (JVMState *jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1032 jvms->set_map(this);
1033 jvms->set_locoff(jvms->locoff()+jvms_adj);
1034 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
1035 jvms->set_monoff(jvms->monoff()+jvms_adj);
1036 jvms->set_scloff(jvms->scloff()+jvms_adj);
1037 jvms->set_endoff(jvms->endoff()+jvms_adj);
1038 }
1039 }
1040
1041 #ifdef ASSERT
1042 bool CallJavaNode::validate_symbolic_info() const {
1043 if (method() == nullptr) {
1044 return true; // call into runtime or uncommon trap
1045 }
1046 ciMethod* symbolic_info = jvms()->method()->get_method_at_bci(jvms()->bci());
1047 ciMethod* callee = method();
1048 if (symbolic_info->is_method_handle_intrinsic() && !callee->is_method_handle_intrinsic()) {
1049 assert(override_symbolic_info(), "should be set");
1050 }
1051 assert(ciMethod::is_consistent_info(symbolic_info, callee), "inconsistent info");
1052 return true;
1053 }
1054 #endif
1055
1056 #ifndef PRODUCT
1057 void CallJavaNode::dump_spec(outputStream* st) const {
1058 if( _method ) _method->print_short_name(st);
1059 CallNode::dump_spec(st);
1060 }
1061
1062 void CallJavaNode::dump_compact_spec(outputStream* st) const {
1063 if (_method) {
1064 _method->print_short_name(st);
1065 } else {
1066 st->print("<?>");
1067 }
1068 }
1069 #endif
1070
1071 //=============================================================================
1072 uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
1073 bool CallStaticJavaNode::cmp( const Node &n ) const {
1074 CallStaticJavaNode &call = (CallStaticJavaNode&)n;
1075 return CallJavaNode::cmp(call);
1076 }
1077
1078 Node* CallStaticJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1079 CallGenerator* cg = generator();
1080 if (can_reshape && cg != nullptr) {
1081 assert(IncrementalInlineMH, "required");
1082 assert(cg->call_node() == this, "mismatch");
1083 assert(cg->is_mh_late_inline(), "not virtual");
1084
1085 // Check whether this MH handle call becomes a candidate for inlining.
1086 ciMethod* callee = cg->method();
1087 vmIntrinsics::ID iid = callee->intrinsic_id();
1088 if (iid == vmIntrinsics::_invokeBasic) {
1089 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) {
1090 phase->C->prepend_late_inline(cg);
1091 set_generator(nullptr);
1092 }
1093 } else if (iid == vmIntrinsics::_linkToNative) {
1094 // never retry
1095 } else {
1096 assert(callee->has_member_arg(), "wrong type of call?");
1097 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) {
1098 phase->C->prepend_late_inline(cg);
1099 set_generator(nullptr);
1100 }
1101 }
1102 }
1103 return CallNode::Ideal(phase, can_reshape);
1104 }
1105
1106 //----------------------------uncommon_trap_request----------------------------
1107 // If this is an uncommon trap, return the request code, else zero.
1108 int CallStaticJavaNode::uncommon_trap_request() const {
1109 if (_name != nullptr && !strcmp(_name, "uncommon_trap")) {
1110 return extract_uncommon_trap_request(this);
1111 }
1112 return 0;
1113 }
1114 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
1115 #ifndef PRODUCT
1116 if (!(call->req() > TypeFunc::Parms &&
1117 call->in(TypeFunc::Parms) != nullptr &&
1118 call->in(TypeFunc::Parms)->is_Con() &&
1119 call->in(TypeFunc::Parms)->bottom_type()->isa_int())) {
1120 assert(in_dump() != 0, "OK if dumping");
1121 tty->print("[bad uncommon trap]");
1122 return 0;
1123 }
1124 #endif
1125 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
1126 }
1127
1128 #ifndef PRODUCT
1129 void CallStaticJavaNode::dump_spec(outputStream *st) const {
1130 st->print("# Static ");
1131 if (_name != nullptr) {
1132 st->print("%s", _name);
1133 int trap_req = uncommon_trap_request();
1134 if (trap_req != 0) {
1135 char buf[100];
1136 st->print("(%s)",
1137 Deoptimization::format_trap_request(buf, sizeof(buf),
1138 trap_req));
1139 }
1140 st->print(" ");
1141 }
1142 CallJavaNode::dump_spec(st);
1143 }
1144
1145 void CallStaticJavaNode::dump_compact_spec(outputStream* st) const {
1146 if (_method) {
1147 _method->print_short_name(st);
1148 } else if (_name) {
1149 st->print("%s", _name);
1150 } else {
1151 st->print("<?>");
1152 }
1153 }
1154 #endif
1155
1156 //=============================================================================
1157 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
1158 bool CallDynamicJavaNode::cmp( const Node &n ) const {
1159 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
1160 return CallJavaNode::cmp(call);
1161 }
1162
1163 Node* CallDynamicJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1164 CallGenerator* cg = generator();
1165 if (can_reshape && cg != nullptr) {
1166 assert(IncrementalInlineVirtual, "required");
1167 assert(cg->call_node() == this, "mismatch");
1168 assert(cg->is_virtual_late_inline(), "not virtual");
1169
1170 // Recover symbolic info for method resolution.
1171 ciMethod* caller = jvms()->method();
1172 ciBytecodeStream iter(caller);
1173 iter.force_bci(jvms()->bci());
1174
1175 bool not_used1;
1176 ciSignature* not_used2;
1177 ciMethod* orig_callee = iter.get_method(not_used1, ¬_used2); // callee in the bytecode
1178 ciKlass* holder = iter.get_declared_method_holder();
1179 if (orig_callee->is_method_handle_intrinsic()) {
1180 assert(_override_symbolic_info, "required");
1181 orig_callee = method();
1182 holder = method()->holder();
1183 }
1184
1185 ciInstanceKlass* klass = ciEnv::get_instance_klass_for_declared_method_holder(holder);
1186
1187 Node* receiver_node = in(TypeFunc::Parms);
1188 const TypeOopPtr* receiver_type = phase->type(receiver_node)->isa_oopptr();
1189
1190 int not_used3;
1191 bool call_does_dispatch;
1192 ciMethod* callee = phase->C->optimize_virtual_call(caller, klass, holder, orig_callee, receiver_type, true /*is_virtual*/,
1193 call_does_dispatch, not_used3); // out-parameters
1194 if (!call_does_dispatch) {
1195 // Register for late inlining.
1196 cg->set_callee_method(callee);
1197 phase->C->prepend_late_inline(cg); // MH late inlining prepends to the list, so do the same
1198 set_generator(nullptr);
1199 }
1200 }
1201 return CallNode::Ideal(phase, can_reshape);
1202 }
1203
1204 #ifndef PRODUCT
1205 void CallDynamicJavaNode::dump_spec(outputStream *st) const {
1206 st->print("# Dynamic ");
1207 CallJavaNode::dump_spec(st);
1208 }
1209 #endif
1210
1211 //=============================================================================
1212 uint CallRuntimeNode::size_of() const { return sizeof(*this); }
1213 bool CallRuntimeNode::cmp( const Node &n ) const {
1214 CallRuntimeNode &call = (CallRuntimeNode&)n;
1215 return CallNode::cmp(call) && !strcmp(_name,call._name);
1216 }
1217 #ifndef PRODUCT
1218 void CallRuntimeNode::dump_spec(outputStream *st) const {
1219 st->print("# ");
1220 st->print("%s", _name);
1221 CallNode::dump_spec(st);
1222 }
1223 #endif
1224 uint CallLeafVectorNode::size_of() const { return sizeof(*this); }
1225 bool CallLeafVectorNode::cmp( const Node &n ) const {
1226 CallLeafVectorNode &call = (CallLeafVectorNode&)n;
1227 return CallLeafNode::cmp(call) && _num_bits == call._num_bits;
1228 }
1229
1230 //------------------------------calling_convention-----------------------------
1231 void CallRuntimeNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
1232 SharedRuntime::c_calling_convention(sig_bt, parm_regs, /*regs2=*/nullptr, argcnt);
1233 }
1234
1235 void CallLeafVectorNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
1236 #ifdef ASSERT
1237 assert(tf()->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1238 "return vector size must match");
1239 const TypeTuple* d = tf()->domain();
1240 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1241 Node* arg = in(i);
1242 assert(arg->bottom_type()->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1243 "vector argument size must match");
1244 }
1245 #endif
1246
1247 SharedRuntime::vector_calling_convention(parm_regs, _num_bits, argcnt);
1248 }
1249
1250 //=============================================================================
1251 //------------------------------calling_convention-----------------------------
1252
1253
1254 //=============================================================================
1255 #ifndef PRODUCT
1256 void CallLeafNode::dump_spec(outputStream *st) const {
1257 st->print("# ");
1258 st->print("%s", _name);
1259 CallNode::dump_spec(st);
1260 }
1261 #endif
1262
1263 //=============================================================================
1264
1265 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) {
1266 assert(verify_jvms(jvms), "jvms must match");
1267 int loc = jvms->locoff() + idx;
1268 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
1269 // If current local idx is top then local idx - 1 could
1270 // be a long/double that needs to be killed since top could
1271 // represent the 2nd half of the long/double.
1272 uint ideal = in(loc -1)->ideal_reg();
1273 if (ideal == Op_RegD || ideal == Op_RegL) {
1274 // set other (low index) half to top
1275 set_req(loc - 1, in(loc));
1276 }
1277 }
1278 set_req(loc, c);
1279 }
1280
1281 uint SafePointNode::size_of() const { return sizeof(*this); }
1282 bool SafePointNode::cmp( const Node &n ) const {
1283 return (&n == this); // Always fail except on self
1284 }
1285
1286 //-------------------------set_next_exception----------------------------------
1287 void SafePointNode::set_next_exception(SafePointNode* n) {
1288 assert(n == nullptr || n->Opcode() == Op_SafePoint, "correct value for next_exception");
1289 if (len() == req()) {
1290 if (n != nullptr) add_prec(n);
1291 } else {
1292 set_prec(req(), n);
1293 }
1294 }
1295
1296
1297 //----------------------------next_exception-----------------------------------
1298 SafePointNode* SafePointNode::next_exception() const {
1299 if (len() == req()) {
1300 return nullptr;
1301 } else {
1302 Node* n = in(req());
1303 assert(n == nullptr || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
1304 return (SafePointNode*) n;
1305 }
1306 }
1307
1308
1309 //------------------------------Ideal------------------------------------------
1310 // Skip over any collapsed Regions
1311 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1312 assert(_jvms == nullptr || ((uintptr_t)_jvms->map() & 1) || _jvms->map() == this, "inconsistent JVMState");
1313 return remove_dead_region(phase, can_reshape) ? this : nullptr;
1314 }
1315
1316 //------------------------------Identity---------------------------------------
1317 // Remove obviously duplicate safepoints
1318 Node* SafePointNode::Identity(PhaseGVN* phase) {
1319
1320 // If you have back to back safepoints, remove one
1321 if (in(TypeFunc::Control)->is_SafePoint()) {
1322 Node* out_c = unique_ctrl_out_or_null();
1323 // This can be the safepoint of an outer strip mined loop if the inner loop's backedge was removed. Replacing the
1324 // outer loop's safepoint could confuse removal of the outer loop.
1325 if (out_c != nullptr && !out_c->is_OuterStripMinedLoopEnd()) {
1326 return in(TypeFunc::Control);
1327 }
1328 }
1329
1330 // Transforming long counted loops requires a safepoint node. Do not
1331 // eliminate a safepoint until loop opts are over.
1332 if (in(0)->is_Proj() && !phase->C->major_progress()) {
1333 Node *n0 = in(0)->in(0);
1334 // Check if he is a call projection (except Leaf Call)
1335 if( n0->is_Catch() ) {
1336 n0 = n0->in(0)->in(0);
1337 assert( n0->is_Call(), "expect a call here" );
1338 }
1339 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
1340 // Don't remove a safepoint belonging to an OuterStripMinedLoopEndNode.
1341 // If the loop dies, they will be removed together.
1342 if (has_out_with(Op_OuterStripMinedLoopEnd)) {
1343 return this;
1344 }
1345 // Useless Safepoint, so remove it
1346 return in(TypeFunc::Control);
1347 }
1348 }
1349
1350 return this;
1351 }
1352
1353 //------------------------------Value------------------------------------------
1354 const Type* SafePointNode::Value(PhaseGVN* phase) const {
1355 if (phase->type(in(0)) == Type::TOP) {
1356 return Type::TOP;
1357 }
1358 if (in(0) == this) {
1359 return Type::TOP; // Dead infinite loop
1360 }
1361 return Type::CONTROL;
1362 }
1363
1364 #ifndef PRODUCT
1365 void SafePointNode::dump_spec(outputStream *st) const {
1366 st->print(" SafePoint ");
1367 _replaced_nodes.dump(st);
1368 }
1369 #endif
1370
1371 const RegMask &SafePointNode::in_RegMask(uint idx) const {
1372 if( idx < TypeFunc::Parms ) return RegMask::Empty;
1373 // Values outside the domain represent debug info
1374 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1375 }
1376 const RegMask &SafePointNode::out_RegMask() const {
1377 return RegMask::Empty;
1378 }
1379
1380
1381 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
1382 assert((int)grow_by > 0, "sanity");
1383 int monoff = jvms->monoff();
1384 int scloff = jvms->scloff();
1385 int endoff = jvms->endoff();
1386 assert(endoff == (int)req(), "no other states or debug info after me");
1387 Node* top = Compile::current()->top();
1388 for (uint i = 0; i < grow_by; i++) {
1389 ins_req(monoff, top);
1390 }
1391 jvms->set_monoff(monoff + grow_by);
1392 jvms->set_scloff(scloff + grow_by);
1393 jvms->set_endoff(endoff + grow_by);
1394 }
1395
1396 void SafePointNode::push_monitor(const FastLockNode *lock) {
1397 // Add a LockNode, which points to both the original BoxLockNode (the
1398 // stack space for the monitor) and the Object being locked.
1399 const int MonitorEdges = 2;
1400 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1401 assert(req() == jvms()->endoff(), "correct sizing");
1402 int nextmon = jvms()->scloff();
1403 if (GenerateSynchronizationCode) {
1404 ins_req(nextmon, lock->box_node());
1405 ins_req(nextmon+1, lock->obj_node());
1406 } else {
1407 Node* top = Compile::current()->top();
1408 ins_req(nextmon, top);
1409 ins_req(nextmon, top);
1410 }
1411 jvms()->set_scloff(nextmon + MonitorEdges);
1412 jvms()->set_endoff(req());
1413 }
1414
1415 void SafePointNode::pop_monitor() {
1416 // Delete last monitor from debug info
1417 debug_only(int num_before_pop = jvms()->nof_monitors());
1418 const int MonitorEdges = 2;
1419 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1420 int scloff = jvms()->scloff();
1421 int endoff = jvms()->endoff();
1422 int new_scloff = scloff - MonitorEdges;
1423 int new_endoff = endoff - MonitorEdges;
1424 jvms()->set_scloff(new_scloff);
1425 jvms()->set_endoff(new_endoff);
1426 while (scloff > new_scloff) del_req_ordered(--scloff);
1427 assert(jvms()->nof_monitors() == num_before_pop-1, "");
1428 }
1429
1430 Node *SafePointNode::peek_monitor_box() const {
1431 int mon = jvms()->nof_monitors() - 1;
1432 assert(mon >= 0, "must have a monitor");
1433 return monitor_box(jvms(), mon);
1434 }
1435
1436 Node *SafePointNode::peek_monitor_obj() const {
1437 int mon = jvms()->nof_monitors() - 1;
1438 assert(mon >= 0, "must have a monitor");
1439 return monitor_obj(jvms(), mon);
1440 }
1441
1442 Node* SafePointNode::peek_operand(uint off) const {
1443 assert(jvms()->sp() > 0, "must have an operand");
1444 assert(off < jvms()->sp(), "off is out-of-range");
1445 return stack(jvms(), jvms()->sp() - off - 1);
1446 }
1447
1448 // Do we Match on this edge index or not? Match no edges
1449 uint SafePointNode::match_edge(uint idx) const {
1450 return (TypeFunc::Parms == idx);
1451 }
1452
1453 void SafePointNode::disconnect_from_root(PhaseIterGVN *igvn) {
1454 assert(Opcode() == Op_SafePoint, "only value for safepoint in loops");
1455 int nb = igvn->C->root()->find_prec_edge(this);
1456 if (nb != -1) {
1457 igvn->delete_precedence_of(igvn->C->root(), nb);
1458 }
1459 }
1460
1461 //============== SafePointScalarObjectNode ==============
1462
1463 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp,
1464 #ifdef ASSERT
1465 Node* alloc,
1466 #endif
1467 uint first_index,
1468 uint depth,
1469 uint n_fields) :
1470 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1471 _first_index(first_index),
1472 _depth(depth),
1473 _n_fields(n_fields)
1474 #ifdef ASSERT
1475 , _alloc(alloc)
1476 #endif
1477 {
1478 #ifdef ASSERT
1479 if (!alloc->is_Allocate()
1480 && !(alloc->Opcode() == Op_VectorBox)) {
1481 alloc->dump();
1482 assert(false, "unexpected call node");
1483 }
1484 #endif
1485 init_class_id(Class_SafePointScalarObject);
1486 }
1487
1488 // Do not allow value-numbering for SafePointScalarObject node.
1489 uint SafePointScalarObjectNode::hash() const { return NO_HASH; }
1490 bool SafePointScalarObjectNode::cmp( const Node &n ) const {
1491 return (&n == this); // Always fail except on self
1492 }
1493
1494 uint SafePointScalarObjectNode::ideal_reg() const {
1495 return 0; // No matching to machine instruction
1496 }
1497
1498 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1499 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1500 }
1501
1502 const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1503 return RegMask::Empty;
1504 }
1505
1506 uint SafePointScalarObjectNode::match_edge(uint idx) const {
1507 return 0;
1508 }
1509
1510 SafePointScalarObjectNode*
1511 SafePointScalarObjectNode::clone(Dict* sosn_map, bool& new_node) const {
1512 void* cached = (*sosn_map)[(void*)this];
1513 if (cached != nullptr) {
1514 new_node = false;
1515 return (SafePointScalarObjectNode*)cached;
1516 }
1517 new_node = true;
1518 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1519 sosn_map->Insert((void*)this, (void*)res);
1520 return res;
1521 }
1522
1523
1524 #ifndef PRODUCT
1525 void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1526 st->print(" # fields@[%d..%d]", first_index(),
1527 first_index() + n_fields() - 1);
1528 }
1529
1530 #endif
1531
1532 //=============================================================================
1533 uint AllocateNode::size_of() const { return sizeof(*this); }
1534
1535 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1536 Node *ctrl, Node *mem, Node *abio,
1537 Node *size, Node *klass_node, Node *initial_test)
1538 : CallNode(atype, nullptr, TypeRawPtr::BOTTOM)
1539 {
1540 init_class_id(Class_Allocate);
1541 init_flags(Flag_is_macro);
1542 _is_scalar_replaceable = false;
1543 _is_non_escaping = false;
1544 _is_allocation_MemBar_redundant = false;
1545 Node *topnode = C->top();
1546
1547 init_req( TypeFunc::Control , ctrl );
1548 init_req( TypeFunc::I_O , abio );
1549 init_req( TypeFunc::Memory , mem );
1550 init_req( TypeFunc::ReturnAdr, topnode );
1551 init_req( TypeFunc::FramePtr , topnode );
1552 init_req( AllocSize , size);
1553 init_req( KlassNode , klass_node);
1554 init_req( InitialTest , initial_test);
1555 init_req( ALength , topnode);
1556 init_req( ValidLengthTest , topnode);
1557 C->add_macro_node(this);
1558 }
1559
1560 void AllocateNode::compute_MemBar_redundancy(ciMethod* initializer)
1561 {
1562 assert(initializer != nullptr &&
1563 initializer->is_initializer() &&
1564 !initializer->is_static(),
1565 "unexpected initializer method");
1566 BCEscapeAnalyzer* analyzer = initializer->get_bcea();
1567 if (analyzer == nullptr) {
1568 return;
1569 }
1570
1571 // Allocation node is first parameter in its initializer
1572 if (analyzer->is_arg_stack(0) || analyzer->is_arg_local(0)) {
1573 _is_allocation_MemBar_redundant = true;
1574 }
1575 }
1576 Node *AllocateNode::make_ideal_mark(PhaseGVN *phase, Node* obj, Node* control, Node* mem) {
1577 Node* mark_node = nullptr;
1578 // For now only enable fast locking for non-array types
1579 mark_node = phase->MakeConX(markWord::prototype().value());
1580 return mark_node;
1581 }
1582
1583 // Retrieve the length from the AllocateArrayNode. Narrow the type with a
1584 // CastII, if appropriate. If we are not allowed to create new nodes, and
1585 // a CastII is appropriate, return null.
1586 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseValues* phase, bool allow_new_nodes) {
1587 Node *length = in(AllocateNode::ALength);
1588 assert(length != nullptr, "length is not null");
1589
1590 const TypeInt* length_type = phase->find_int_type(length);
1591 const TypeAryPtr* ary_type = oop_type->isa_aryptr();
1592
1593 if (ary_type != nullptr && length_type != nullptr) {
1594 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type);
1595 if (narrow_length_type != length_type) {
1596 // Assert one of:
1597 // - the narrow_length is 0
1598 // - the narrow_length is not wider than length
1599 assert(narrow_length_type == TypeInt::ZERO ||
1600 length_type->is_con() && narrow_length_type->is_con() &&
1601 (narrow_length_type->_hi <= length_type->_lo) ||
1602 (narrow_length_type->_hi <= length_type->_hi &&
1603 narrow_length_type->_lo >= length_type->_lo),
1604 "narrow type must be narrower than length type");
1605
1606 // Return null if new nodes are not allowed
1607 if (!allow_new_nodes) {
1608 return nullptr;
1609 }
1610 // Create a cast which is control dependent on the initialization to
1611 // propagate the fact that the array length must be positive.
1612 InitializeNode* init = initialization();
1613 if (init != nullptr) {
1614 length = new CastIINode(length, narrow_length_type);
1615 length->set_req(TypeFunc::Control, init->proj_out_or_null(TypeFunc::Control));
1616 }
1617 }
1618 }
1619
1620 return length;
1621 }
1622
1623 //=============================================================================
1624 uint LockNode::size_of() const { return sizeof(*this); }
1625
1626 // Redundant lock elimination
1627 //
1628 // There are various patterns of locking where we release and
1629 // immediately reacquire a lock in a piece of code where no operations
1630 // occur in between that would be observable. In those cases we can
1631 // skip releasing and reacquiring the lock without violating any
1632 // fairness requirements. Doing this around a loop could cause a lock
1633 // to be held for a very long time so we concentrate on non-looping
1634 // control flow. We also require that the operations are fully
1635 // redundant meaning that we don't introduce new lock operations on
1636 // some paths so to be able to eliminate it on others ala PRE. This
1637 // would probably require some more extensive graph manipulation to
1638 // guarantee that the memory edges were all handled correctly.
1639 //
1640 // Assuming p is a simple predicate which can't trap in any way and s
1641 // is a synchronized method consider this code:
1642 //
1643 // s();
1644 // if (p)
1645 // s();
1646 // else
1647 // s();
1648 // s();
1649 //
1650 // 1. The unlocks of the first call to s can be eliminated if the
1651 // locks inside the then and else branches are eliminated.
1652 //
1653 // 2. The unlocks of the then and else branches can be eliminated if
1654 // the lock of the final call to s is eliminated.
1655 //
1656 // Either of these cases subsumes the simple case of sequential control flow
1657 //
1658 // Additionally we can eliminate versions without the else case:
1659 //
1660 // s();
1661 // if (p)
1662 // s();
1663 // s();
1664 //
1665 // 3. In this case we eliminate the unlock of the first s, the lock
1666 // and unlock in the then case and the lock in the final s.
1667 //
1668 // Note also that in all these cases the then/else pieces don't have
1669 // to be trivial as long as they begin and end with synchronization
1670 // operations.
1671 //
1672 // s();
1673 // if (p)
1674 // s();
1675 // f();
1676 // s();
1677 // s();
1678 //
1679 // The code will work properly for this case, leaving in the unlock
1680 // before the call to f and the relock after it.
1681 //
1682 // A potentially interesting case which isn't handled here is when the
1683 // locking is partially redundant.
1684 //
1685 // s();
1686 // if (p)
1687 // s();
1688 //
1689 // This could be eliminated putting unlocking on the else case and
1690 // eliminating the first unlock and the lock in the then side.
1691 // Alternatively the unlock could be moved out of the then side so it
1692 // was after the merge and the first unlock and second lock
1693 // eliminated. This might require less manipulation of the memory
1694 // state to get correct.
1695 //
1696 // Additionally we might allow work between a unlock and lock before
1697 // giving up eliminating the locks. The current code disallows any
1698 // conditional control flow between these operations. A formulation
1699 // similar to partial redundancy elimination computing the
1700 // availability of unlocking and the anticipatability of locking at a
1701 // program point would allow detection of fully redundant locking with
1702 // some amount of work in between. I'm not sure how often I really
1703 // think that would occur though. Most of the cases I've seen
1704 // indicate it's likely non-trivial work would occur in between.
1705 // There may be other more complicated constructs where we could
1706 // eliminate locking but I haven't seen any others appear as hot or
1707 // interesting.
1708 //
1709 // Locking and unlocking have a canonical form in ideal that looks
1710 // roughly like this:
1711 //
1712 // <obj>
1713 // | \\------+
1714 // | \ \
1715 // | BoxLock \
1716 // | | | \
1717 // | | \ \
1718 // | | FastLock
1719 // | | /
1720 // | | /
1721 // | | |
1722 //
1723 // Lock
1724 // |
1725 // Proj #0
1726 // |
1727 // MembarAcquire
1728 // |
1729 // Proj #0
1730 //
1731 // MembarRelease
1732 // |
1733 // Proj #0
1734 // |
1735 // Unlock
1736 // |
1737 // Proj #0
1738 //
1739 //
1740 // This code proceeds by processing Lock nodes during PhaseIterGVN
1741 // and searching back through its control for the proper code
1742 // patterns. Once it finds a set of lock and unlock operations to
1743 // eliminate they are marked as eliminatable which causes the
1744 // expansion of the Lock and Unlock macro nodes to make the operation a NOP
1745 //
1746 //=============================================================================
1747
1748 //
1749 // Utility function to skip over uninteresting control nodes. Nodes skipped are:
1750 // - copy regions. (These may not have been optimized away yet.)
1751 // - eliminated locking nodes
1752 //
1753 static Node *next_control(Node *ctrl) {
1754 if (ctrl == nullptr)
1755 return nullptr;
1756 while (1) {
1757 if (ctrl->is_Region()) {
1758 RegionNode *r = ctrl->as_Region();
1759 Node *n = r->is_copy();
1760 if (n == nullptr)
1761 break; // hit a region, return it
1762 else
1763 ctrl = n;
1764 } else if (ctrl->is_Proj()) {
1765 Node *in0 = ctrl->in(0);
1766 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
1767 ctrl = in0->in(0);
1768 } else {
1769 break;
1770 }
1771 } else {
1772 break; // found an interesting control
1773 }
1774 }
1775 return ctrl;
1776 }
1777 //
1778 // Given a control, see if it's the control projection of an Unlock which
1779 // operating on the same object as lock.
1780 //
1781 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
1782 GrowableArray<AbstractLockNode*> &lock_ops) {
1783 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : nullptr;
1784 if (ctrl_proj != nullptr && ctrl_proj->_con == TypeFunc::Control) {
1785 Node *n = ctrl_proj->in(0);
1786 if (n != nullptr && n->is_Unlock()) {
1787 UnlockNode *unlock = n->as_Unlock();
1788 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1789 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1790 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
1791 if (lock_obj->eqv_uncast(unlock_obj) &&
1792 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) &&
1793 !unlock->is_eliminated()) {
1794 lock_ops.append(unlock);
1795 return true;
1796 }
1797 }
1798 }
1799 return false;
1800 }
1801
1802 //
1803 // Find the lock matching an unlock. Returns null if a safepoint
1804 // or complicated control is encountered first.
1805 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
1806 LockNode *lock_result = nullptr;
1807 // find the matching lock, or an intervening safepoint
1808 Node *ctrl = next_control(unlock->in(0));
1809 while (1) {
1810 assert(ctrl != nullptr, "invalid control graph");
1811 assert(!ctrl->is_Start(), "missing lock for unlock");
1812 if (ctrl->is_top()) break; // dead control path
1813 if (ctrl->is_Proj()) ctrl = ctrl->in(0);
1814 if (ctrl->is_SafePoint()) {
1815 break; // found a safepoint (may be the lock we are searching for)
1816 } else if (ctrl->is_Region()) {
1817 // Check for a simple diamond pattern. Punt on anything more complicated
1818 if (ctrl->req() == 3 && ctrl->in(1) != nullptr && ctrl->in(2) != nullptr) {
1819 Node *in1 = next_control(ctrl->in(1));
1820 Node *in2 = next_control(ctrl->in(2));
1821 if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
1822 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
1823 ctrl = next_control(in1->in(0)->in(0));
1824 } else {
1825 break;
1826 }
1827 } else {
1828 break;
1829 }
1830 } else {
1831 ctrl = next_control(ctrl->in(0)); // keep searching
1832 }
1833 }
1834 if (ctrl->is_Lock()) {
1835 LockNode *lock = ctrl->as_Lock();
1836 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1837 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1838 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
1839 if (lock_obj->eqv_uncast(unlock_obj) &&
1840 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) {
1841 lock_result = lock;
1842 }
1843 }
1844 return lock_result;
1845 }
1846
1847 // This code corresponds to case 3 above.
1848
1849 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
1850 GrowableArray<AbstractLockNode*> &lock_ops) {
1851 Node* if_node = node->in(0);
1852 bool if_true = node->is_IfTrue();
1853
1854 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
1855 Node *lock_ctrl = next_control(if_node->in(0));
1856 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
1857 Node* lock1_node = nullptr;
1858 ProjNode* proj = if_node->as_If()->proj_out(!if_true);
1859 if (if_true) {
1860 if (proj->is_IfFalse() && proj->outcnt() == 1) {
1861 lock1_node = proj->unique_out();
1862 }
1863 } else {
1864 if (proj->is_IfTrue() && proj->outcnt() == 1) {
1865 lock1_node = proj->unique_out();
1866 }
1867 }
1868 if (lock1_node != nullptr && lock1_node->is_Lock()) {
1869 LockNode *lock1 = lock1_node->as_Lock();
1870 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1871 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1872 Node* lock1_obj = bs->step_over_gc_barrier(lock1->obj_node());
1873 if (lock_obj->eqv_uncast(lock1_obj) &&
1874 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) &&
1875 !lock1->is_eliminated()) {
1876 lock_ops.append(lock1);
1877 return true;
1878 }
1879 }
1880 }
1881 }
1882
1883 lock_ops.trunc_to(0);
1884 return false;
1885 }
1886
1887 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
1888 GrowableArray<AbstractLockNode*> &lock_ops) {
1889 // check each control merging at this point for a matching unlock.
1890 // in(0) should be self edge so skip it.
1891 for (int i = 1; i < (int)region->req(); i++) {
1892 Node *in_node = next_control(region->in(i));
1893 if (in_node != nullptr) {
1894 if (find_matching_unlock(in_node, lock, lock_ops)) {
1895 // found a match so keep on checking.
1896 continue;
1897 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
1898 continue;
1899 }
1900
1901 // If we fall through to here then it was some kind of node we
1902 // don't understand or there wasn't a matching unlock, so give
1903 // up trying to merge locks.
1904 lock_ops.trunc_to(0);
1905 return false;
1906 }
1907 }
1908 return true;
1909
1910 }
1911
1912 // Check that all locks/unlocks associated with object come from balanced regions.
1913 bool AbstractLockNode::is_balanced() {
1914 Node* obj = obj_node();
1915 for (uint j = 0; j < obj->outcnt(); j++) {
1916 Node* n = obj->raw_out(j);
1917 if (n->is_AbstractLock() &&
1918 n->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
1919 BoxLockNode* n_box = n->as_AbstractLock()->box_node()->as_BoxLock();
1920 if (n_box->is_unbalanced()) {
1921 return false;
1922 }
1923 }
1924 }
1925 return true;
1926 }
1927
1928 const char* AbstractLockNode::_kind_names[] = {"Regular", "NonEscObj", "Coarsened", "Nested"};
1929
1930 const char * AbstractLockNode::kind_as_string() const {
1931 return _kind_names[_kind];
1932 }
1933
1934 #ifndef PRODUCT
1935 //
1936 // Create a counter which counts the number of times this lock is acquired
1937 //
1938 void AbstractLockNode::create_lock_counter(JVMState* state) {
1939 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
1940 }
1941
1942 void AbstractLockNode::set_eliminated_lock_counter() {
1943 if (_counter) {
1944 // Update the counter to indicate that this lock was eliminated.
1945 // The counter update code will stay around even though the
1946 // optimizer will eliminate the lock operation itself.
1947 _counter->set_tag(NamedCounter::EliminatedLockCounter);
1948 }
1949 }
1950
1951 void AbstractLockNode::dump_spec(outputStream* st) const {
1952 st->print("%s ", _kind_names[_kind]);
1953 CallNode::dump_spec(st);
1954 }
1955
1956 void AbstractLockNode::dump_compact_spec(outputStream* st) const {
1957 st->print("%s", _kind_names[_kind]);
1958 }
1959 #endif
1960
1961 //=============================================================================
1962 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1963
1964 // perform any generic optimizations first (returns 'this' or null)
1965 Node *result = SafePointNode::Ideal(phase, can_reshape);
1966 if (result != nullptr) return result;
1967 // Don't bother trying to transform a dead node
1968 if (in(0) && in(0)->is_top()) return nullptr;
1969
1970 // Now see if we can optimize away this lock. We don't actually
1971 // remove the locking here, we simply set the _eliminate flag which
1972 // prevents macro expansion from expanding the lock. Since we don't
1973 // modify the graph, the value returned from this function is the
1974 // one computed above.
1975 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
1976 //
1977 // If we are locking an non-escaped object, the lock/unlock is unnecessary
1978 //
1979 ConnectionGraph *cgr = phase->C->congraph();
1980 if (cgr != nullptr && cgr->can_eliminate_lock(this)) {
1981 assert(!is_eliminated() || is_coarsened(), "sanity");
1982 // The lock could be marked eliminated by lock coarsening
1983 // code during first IGVN before EA. Replace coarsened flag
1984 // to eliminate all associated locks/unlocks.
1985 #ifdef ASSERT
1986 this->log_lock_optimization(phase->C,"eliminate_lock_set_non_esc1");
1987 #endif
1988 this->set_non_esc_obj();
1989 return result;
1990 }
1991
1992 if (!phase->C->do_locks_coarsening()) {
1993 return result; // Compiling without locks coarsening
1994 }
1995 //
1996 // Try lock coarsening
1997 //
1998 PhaseIterGVN* iter = phase->is_IterGVN();
1999 if (iter != nullptr && !is_eliminated()) {
2000
2001 GrowableArray<AbstractLockNode*> lock_ops;
2002
2003 Node *ctrl = next_control(in(0));
2004
2005 // now search back for a matching Unlock
2006 if (find_matching_unlock(ctrl, this, lock_ops)) {
2007 // found an unlock directly preceding this lock. This is the
2008 // case of single unlock directly control dependent on a
2009 // single lock which is the trivial version of case 1 or 2.
2010 } else if (ctrl->is_Region() ) {
2011 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
2012 // found lock preceded by multiple unlocks along all paths
2013 // joining at this point which is case 3 in description above.
2014 }
2015 } else {
2016 // see if this lock comes from either half of an if and the
2017 // predecessors merges unlocks and the other half of the if
2018 // performs a lock.
2019 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
2020 // found unlock splitting to an if with locks on both branches.
2021 }
2022 }
2023
2024 if (lock_ops.length() > 0) {
2025 // add ourselves to the list of locks to be eliminated.
2026 lock_ops.append(this);
2027
2028 #ifndef PRODUCT
2029 if (PrintEliminateLocks) {
2030 int locks = 0;
2031 int unlocks = 0;
2032 if (Verbose) {
2033 tty->print_cr("=== Locks coarsening ===");
2034 tty->print("Obj: ");
2035 obj_node()->dump();
2036 }
2037 for (int i = 0; i < lock_ops.length(); i++) {
2038 AbstractLockNode* lock = lock_ops.at(i);
2039 if (lock->Opcode() == Op_Lock)
2040 locks++;
2041 else
2042 unlocks++;
2043 if (Verbose) {
2044 tty->print("Box %d: ", i);
2045 box_node()->dump();
2046 tty->print(" %d: ", i);
2047 lock->dump();
2048 }
2049 }
2050 tty->print_cr("=== Coarsened %d unlocks and %d locks", unlocks, locks);
2051 }
2052 #endif
2053
2054 // for each of the identified locks, mark them
2055 // as eliminatable
2056 for (int i = 0; i < lock_ops.length(); i++) {
2057 AbstractLockNode* lock = lock_ops.at(i);
2058
2059 // Mark it eliminated by coarsening and update any counters
2060 #ifdef ASSERT
2061 lock->log_lock_optimization(phase->C, "eliminate_lock_set_coarsened");
2062 #endif
2063 lock->set_coarsened();
2064 }
2065 // Record this coarsened group.
2066 phase->C->add_coarsened_locks(lock_ops);
2067 } else if (ctrl->is_Region() &&
2068 iter->_worklist.member(ctrl)) {
2069 // We weren't able to find any opportunities but the region this
2070 // lock is control dependent on hasn't been processed yet so put
2071 // this lock back on the worklist so we can check again once any
2072 // region simplification has occurred.
2073 iter->_worklist.push(this);
2074 }
2075 }
2076 }
2077
2078 return result;
2079 }
2080
2081 //=============================================================================
2082 bool LockNode::is_nested_lock_region() {
2083 return is_nested_lock_region(nullptr);
2084 }
2085
2086 // p is used for access to compilation log; no logging if null
2087 bool LockNode::is_nested_lock_region(Compile * c) {
2088 BoxLockNode* box = box_node()->as_BoxLock();
2089 int stk_slot = box->stack_slot();
2090 if (stk_slot <= 0) {
2091 #ifdef ASSERT
2092 this->log_lock_optimization(c, "eliminate_lock_INLR_1");
2093 #endif
2094 return false; // External lock or it is not Box (Phi node).
2095 }
2096
2097 // Ignore complex cases: merged locks or multiple locks.
2098 Node* obj = obj_node();
2099 LockNode* unique_lock = nullptr;
2100 Node* bad_lock = nullptr;
2101 if (!box->is_simple_lock_region(&unique_lock, obj, &bad_lock)) {
2102 #ifdef ASSERT
2103 this->log_lock_optimization(c, "eliminate_lock_INLR_2a", bad_lock);
2104 #endif
2105 return false;
2106 }
2107 if (unique_lock != this) {
2108 #ifdef ASSERT
2109 this->log_lock_optimization(c, "eliminate_lock_INLR_2b", (unique_lock != nullptr ? unique_lock : bad_lock));
2110 if (PrintEliminateLocks && Verbose) {
2111 tty->print_cr("=============== unique_lock != this ============");
2112 tty->print(" this: ");
2113 this->dump();
2114 tty->print(" box: ");
2115 box->dump();
2116 tty->print(" obj: ");
2117 obj->dump();
2118 if (unique_lock != nullptr) {
2119 tty->print(" unique_lock: ");
2120 unique_lock->dump();
2121 }
2122 if (bad_lock != nullptr) {
2123 tty->print(" bad_lock: ");
2124 bad_lock->dump();
2125 }
2126 tty->print_cr("===============");
2127 }
2128 #endif
2129 return false;
2130 }
2131
2132 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2133 obj = bs->step_over_gc_barrier(obj);
2134 // Look for external lock for the same object.
2135 SafePointNode* sfn = this->as_SafePoint();
2136 JVMState* youngest_jvms = sfn->jvms();
2137 int max_depth = youngest_jvms->depth();
2138 for (int depth = 1; depth <= max_depth; depth++) {
2139 JVMState* jvms = youngest_jvms->of_depth(depth);
2140 int num_mon = jvms->nof_monitors();
2141 // Loop over monitors
2142 for (int idx = 0; idx < num_mon; idx++) {
2143 Node* obj_node = sfn->monitor_obj(jvms, idx);
2144 obj_node = bs->step_over_gc_barrier(obj_node);
2145 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock();
2146 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) {
2147 box->set_nested();
2148 return true;
2149 }
2150 }
2151 }
2152 #ifdef ASSERT
2153 this->log_lock_optimization(c, "eliminate_lock_INLR_3");
2154 #endif
2155 return false;
2156 }
2157
2158 //=============================================================================
2159 uint UnlockNode::size_of() const { return sizeof(*this); }
2160
2161 //=============================================================================
2162 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2163
2164 // perform any generic optimizations first (returns 'this' or null)
2165 Node *result = SafePointNode::Ideal(phase, can_reshape);
2166 if (result != nullptr) return result;
2167 // Don't bother trying to transform a dead node
2168 if (in(0) && in(0)->is_top()) return nullptr;
2169
2170 // Now see if we can optimize away this unlock. We don't actually
2171 // remove the unlocking here, we simply set the _eliminate flag which
2172 // prevents macro expansion from expanding the unlock. Since we don't
2173 // modify the graph, the value returned from this function is the
2174 // one computed above.
2175 // Escape state is defined after Parse phase.
2176 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
2177 //
2178 // If we are unlocking an non-escaped object, the lock/unlock is unnecessary.
2179 //
2180 ConnectionGraph *cgr = phase->C->congraph();
2181 if (cgr != nullptr && cgr->can_eliminate_lock(this)) {
2182 assert(!is_eliminated() || is_coarsened(), "sanity");
2183 // The lock could be marked eliminated by lock coarsening
2184 // code during first IGVN before EA. Replace coarsened flag
2185 // to eliminate all associated locks/unlocks.
2186 #ifdef ASSERT
2187 this->log_lock_optimization(phase->C, "eliminate_lock_set_non_esc2");
2188 #endif
2189 this->set_non_esc_obj();
2190 }
2191 }
2192 return result;
2193 }
2194
2195 void AbstractLockNode::log_lock_optimization(Compile *C, const char * tag, Node* bad_lock) const {
2196 if (C == nullptr) {
2197 return;
2198 }
2199 CompileLog* log = C->log();
2200 if (log != nullptr) {
2201 Node* box = box_node();
2202 Node* obj = obj_node();
2203 int box_id = box != nullptr ? box->_idx : -1;
2204 int obj_id = obj != nullptr ? obj->_idx : -1;
2205
2206 log->begin_head("%s compile_id='%d' lock_id='%d' class='%s' kind='%s' box_id='%d' obj_id='%d' bad_id='%d'",
2207 tag, C->compile_id(), this->_idx,
2208 is_Unlock() ? "unlock" : is_Lock() ? "lock" : "?",
2209 kind_as_string(), box_id, obj_id, (bad_lock != nullptr ? bad_lock->_idx : -1));
2210 log->stamp();
2211 log->end_head();
2212 JVMState* p = is_Unlock() ? (as_Unlock()->dbg_jvms()) : jvms();
2213 while (p != nullptr) {
2214 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
2215 p = p->caller();
2216 }
2217 log->tail(tag);
2218 }
2219 }
2220
2221 bool CallNode::may_modify_arraycopy_helper(const TypeOopPtr* dest_t, const TypeOopPtr* t_oop, PhaseValues* phase) {
2222 if (dest_t->is_known_instance() && t_oop->is_known_instance()) {
2223 return dest_t->instance_id() == t_oop->instance_id();
2224 }
2225
2226 if (dest_t->isa_instptr() && !dest_t->is_instptr()->instance_klass()->equals(phase->C->env()->Object_klass())) {
2227 // clone
2228 if (t_oop->isa_aryptr()) {
2229 return false;
2230 }
2231 if (!t_oop->isa_instptr()) {
2232 return true;
2233 }
2234 if (dest_t->maybe_java_subtype_of(t_oop) || t_oop->maybe_java_subtype_of(dest_t)) {
2235 return true;
2236 }
2237 // unrelated
2238 return false;
2239 }
2240
2241 if (dest_t->isa_aryptr()) {
2242 // arraycopy or array clone
2243 if (t_oop->isa_instptr()) {
2244 return false;
2245 }
2246 if (!t_oop->isa_aryptr()) {
2247 return true;
2248 }
2249
2250 const Type* elem = dest_t->is_aryptr()->elem();
2251 if (elem == Type::BOTTOM) {
2252 // An array but we don't know what elements are
2253 return true;
2254 }
2255
2256 dest_t = dest_t->add_offset(Type::OffsetBot)->is_oopptr();
2257 uint dest_alias = phase->C->get_alias_index(dest_t);
2258 uint t_oop_alias = phase->C->get_alias_index(t_oop);
2259
2260 return dest_alias == t_oop_alias;
2261 }
2262
2263 return true;
2264 }