1 /*
2 * Copyright (c) 1997, 2026, 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 "ci/bcEscapeAnalyzer.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciSymbols.hpp"
28 #include "code/vmreg.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "compiler/oopMap.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/c2/barrierSetC2.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "opto/callGenerator.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/escape.hpp"
39 #include "opto/inlinetypenode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/matcher.hpp"
43 #include "opto/movenode.hpp"
44 #include "opto/parse.hpp"
45 #include "opto/regalloc.hpp"
46 #include "opto/regmask.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "runtime/arguments.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "runtime/stubRoutines.hpp"
52 #include "utilities/powerOfTwo.hpp"
53
54 // Portions of code courtesy of Clifford Click
55
56 // Optimization - Graph Style
57
58 //=============================================================================
59 uint StartNode::size_of() const { return sizeof(*this); }
60 bool StartNode::cmp( const Node &n ) const
61 { return _domain == ((StartNode&)n)._domain; }
62 const Type *StartNode::bottom_type() const { return _domain; }
63 const Type* StartNode::Value(PhaseGVN* phase) const { return _domain; }
64 #ifndef PRODUCT
65 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);}
66 void StartNode::dump_compact_spec(outputStream *st) const { /* empty */ }
67 #endif
68
69 //------------------------------Ideal------------------------------------------
70 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){
71 return remove_dead_region(phase, can_reshape) ? this : nullptr;
72 }
73
74 //------------------------------calling_convention-----------------------------
75 void StartNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
76 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt);
77 }
78
79 //------------------------------Registers--------------------------------------
80 const RegMask &StartNode::in_RegMask(uint) const {
81 return RegMask::EMPTY;
82 }
83
84 //------------------------------match------------------------------------------
85 // Construct projections for incoming parameters, and their RegMask info
86 Node *StartNode::match(const ProjNode *proj, const Matcher *match, const RegMask* mask) {
87 switch (proj->_con) {
88 case TypeFunc::Control:
89 case TypeFunc::I_O:
90 case TypeFunc::Memory:
91 return new MachProjNode(this,proj->_con,RegMask::EMPTY,MachProjNode::unmatched_proj);
92 case TypeFunc::FramePtr:
93 return new MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP);
94 case TypeFunc::ReturnAdr:
95 return new MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP);
96 case TypeFunc::Parms:
97 default: {
98 uint parm_num = proj->_con - TypeFunc::Parms;
99 const Type *t = _domain->field_at(proj->_con);
100 if (t->base() == Type::Half) // 2nd half of Longs and Doubles
101 return new ConNode(Type::TOP);
102 uint ideal_reg = t->ideal_reg();
103 RegMask &rm = match->_calling_convention_mask[parm_num];
104 return new MachProjNode(this,proj->_con,rm,ideal_reg);
105 }
106 }
107 return nullptr;
108 }
109
110 //=============================================================================
111 const char * const ParmNode::names[TypeFunc::Parms+1] = {
112 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms"
113 };
114
115 #ifndef PRODUCT
116 void ParmNode::dump_spec(outputStream *st) const {
117 if( _con < TypeFunc::Parms ) {
118 st->print("%s", names[_con]);
119 } else {
120 st->print("Parm%d: ",_con-TypeFunc::Parms);
121 // Verbose and WizardMode dump bottom_type for all nodes
122 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st);
123 }
124 }
125
126 void ParmNode::dump_compact_spec(outputStream *st) const {
127 if (_con < TypeFunc::Parms) {
128 st->print("%s", names[_con]);
129 } else {
130 st->print("%d:", _con-TypeFunc::Parms);
131 // unconditionally dump bottom_type
132 bottom_type()->dump_on(st);
133 }
134 }
135 #endif
136
137 uint ParmNode::ideal_reg() const {
138 switch( _con ) {
139 case TypeFunc::Control : // fall through
140 case TypeFunc::I_O : // fall through
141 case TypeFunc::Memory : return 0;
142 case TypeFunc::FramePtr : // fall through
143 case TypeFunc::ReturnAdr: return Op_RegP;
144 default : assert( _con > TypeFunc::Parms, "" );
145 // fall through
146 case TypeFunc::Parms : {
147 // Type of argument being passed
148 const Type *t = in(0)->as_Start()->_domain->field_at(_con);
149 return t->ideal_reg();
150 }
151 }
152 ShouldNotReachHere();
153 return 0;
154 }
155
156 //=============================================================================
157 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) {
158 init_req(TypeFunc::Control,cntrl);
159 init_req(TypeFunc::I_O,i_o);
160 init_req(TypeFunc::Memory,memory);
161 init_req(TypeFunc::FramePtr,frameptr);
162 init_req(TypeFunc::ReturnAdr,retadr);
163 }
164
165 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){
166 return remove_dead_region(phase, can_reshape) ? this : nullptr;
167 }
168
169 const Type* ReturnNode::Value(PhaseGVN* phase) const {
170 return ( phase->type(in(TypeFunc::Control)) == Type::TOP)
171 ? Type::TOP
172 : Type::BOTTOM;
173 }
174
175 // Do we Match on this edge index or not? No edges on return nodes
176 uint ReturnNode::match_edge(uint idx) const {
177 return 0;
178 }
179
180
181 #ifndef PRODUCT
182 void ReturnNode::dump_req(outputStream *st, DumpConfig* dc) const {
183 // Dump the required inputs, after printing "returns"
184 uint i; // Exit value of loop
185 for (i = 0; i < req(); i++) { // For all required inputs
186 if (i == TypeFunc::Parms) st->print("returns ");
187 Node* p = in(i);
188 if (p != nullptr) {
189 p->dump_idx(false, st, dc);
190 st->print(" ");
191 } else {
192 st->print("_ ");
193 }
194 }
195 }
196 #endif
197
198 //=============================================================================
199 RethrowNode::RethrowNode(
200 Node* cntrl,
201 Node* i_o,
202 Node* memory,
203 Node* frameptr,
204 Node* ret_adr,
205 Node* exception
206 ) : Node(TypeFunc::Parms + 1) {
207 init_req(TypeFunc::Control , cntrl );
208 init_req(TypeFunc::I_O , i_o );
209 init_req(TypeFunc::Memory , memory );
210 init_req(TypeFunc::FramePtr , frameptr );
211 init_req(TypeFunc::ReturnAdr, ret_adr);
212 init_req(TypeFunc::Parms , exception);
213 }
214
215 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){
216 return remove_dead_region(phase, can_reshape) ? this : nullptr;
217 }
218
219 const Type* RethrowNode::Value(PhaseGVN* phase) const {
220 return (phase->type(in(TypeFunc::Control)) == Type::TOP)
221 ? Type::TOP
222 : Type::BOTTOM;
223 }
224
225 uint RethrowNode::match_edge(uint idx) const {
226 return 0;
227 }
228
229 #ifndef PRODUCT
230 void RethrowNode::dump_req(outputStream *st, DumpConfig* dc) const {
231 // Dump the required inputs, after printing "exception"
232 uint i; // Exit value of loop
233 for (i = 0; i < req(); i++) { // For all required inputs
234 if (i == TypeFunc::Parms) st->print("exception ");
235 Node* p = in(i);
236 if (p != nullptr) {
237 p->dump_idx(false, st, dc);
238 st->print(" ");
239 } else {
240 st->print("_ ");
241 }
242 }
243 }
244 #endif
245
246 //=============================================================================
247 // Do we Match on this edge index or not? Match only target address & method
248 uint TailCallNode::match_edge(uint idx) const {
249 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
250 }
251
252 //=============================================================================
253 // Do we Match on this edge index or not? Match only target address & oop
254 uint TailJumpNode::match_edge(uint idx) const {
255 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
256 }
257
258 //=============================================================================
259 JVMState::JVMState(ciMethod* method, JVMState* caller) :
260 _method(method),
261 _receiver_info(nullptr) {
262 assert(method != nullptr, "must be valid call site");
263 _bci = InvocationEntryBci;
264 _reexecute = Reexecute_Undefined;
265 DEBUG_ONLY(_bci = -99); // random garbage value
266 DEBUG_ONLY(_map = (SafePointNode*)-1);
267 _caller = caller;
268 _depth = 1 + (caller == nullptr ? 0 : caller->depth());
269 _locoff = TypeFunc::Parms;
270 _stkoff = _locoff + _method->max_locals();
271 _monoff = _stkoff + _method->max_stack();
272 _scloff = _monoff;
273 _endoff = _monoff;
274 _sp = 0;
275 }
276 JVMState::JVMState(int stack_size) :
277 _method(nullptr),
278 _receiver_info(nullptr) {
279 _bci = InvocationEntryBci;
280 _reexecute = Reexecute_Undefined;
281 DEBUG_ONLY(_map = (SafePointNode*)-1);
282 _caller = nullptr;
283 _depth = 1;
284 _locoff = TypeFunc::Parms;
285 _stkoff = _locoff;
286 _monoff = _stkoff + stack_size;
287 _scloff = _monoff;
288 _endoff = _monoff;
289 _sp = 0;
290 }
291
292 //--------------------------------of_depth-------------------------------------
293 JVMState* JVMState::of_depth(int d) const {
294 const JVMState* jvmp = this;
295 assert(0 < d && (uint)d <= depth(), "oob");
296 for (int skip = depth() - d; skip > 0; skip--) {
297 jvmp = jvmp->caller();
298 }
299 assert(jvmp->depth() == (uint)d, "found the right one");
300 return (JVMState*)jvmp;
301 }
302
303 //-----------------------------same_calls_as-----------------------------------
304 bool JVMState::same_calls_as(const JVMState* that) const {
305 if (this == that) return true;
306 if (this->depth() != that->depth()) return false;
307 const JVMState* p = this;
308 const JVMState* q = that;
309 for (;;) {
310 if (p->_method != q->_method) return false;
311 if (p->_method == nullptr) return true; // bci is irrelevant
312 if (p->_bci != q->_bci) return false;
313 if (p->_reexecute != q->_reexecute) return false;
314 p = p->caller();
315 q = q->caller();
316 if (p == q) return true;
317 assert(p != nullptr && q != nullptr, "depth check ensures we don't run off end");
318 }
319 }
320
321 //------------------------------debug_start------------------------------------
322 uint JVMState::debug_start() const {
323 DEBUG_ONLY(JVMState* jvmroot = of_depth(1));
324 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
325 return of_depth(1)->locoff();
326 }
327
328 //-------------------------------debug_end-------------------------------------
329 uint JVMState::debug_end() const {
330 DEBUG_ONLY(JVMState* jvmroot = of_depth(1));
331 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
332 return endoff();
333 }
334
335 //------------------------------debug_depth------------------------------------
336 uint JVMState::debug_depth() const {
337 uint total = 0;
338 for (const JVMState* jvmp = this; jvmp != nullptr; jvmp = jvmp->caller()) {
339 total += jvmp->debug_size();
340 }
341 return total;
342 }
343
344 #ifndef PRODUCT
345
346 //------------------------------format_helper----------------------------------
347 // Given an allocation (a Chaitin object) and a Node decide if the Node carries
348 // any defined value or not. If it does, print out the register or constant.
349 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
350 if (n == nullptr) { st->print(" null"); return; }
351 if (n->is_SafePointScalarObject()) {
352 // Scalar replacement.
353 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
354 scobjs->append_if_missing(spobj);
355 int sco_n = scobjs->find(spobj);
356 assert(sco_n >= 0, "");
357 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
358 return;
359 }
360 if (regalloc->node_regs_max_index() > 0 &&
361 OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
362 char buf[50];
363 regalloc->dump_register(n,buf,sizeof(buf));
364 st->print(" %s%d]=%s",msg,i,buf);
365 } else { // No register, but might be constant
366 const Type *t = n->bottom_type();
367 switch (t->base()) {
368 case Type::Int:
369 st->print(" %s%d]=#" INT32_FORMAT,msg,i,t->is_int()->get_con());
370 break;
371 case Type::AnyPtr:
372 assert( t == TypePtr::NULL_PTR || n->in_dump(), "" );
373 st->print(" %s%d]=#null",msg,i);
374 break;
375 case Type::AryPtr:
376 case Type::InstPtr:
377 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->isa_oopptr()->const_oop()));
378 break;
379 case Type::KlassPtr:
380 case Type::AryKlassPtr:
381 case Type::InstKlassPtr:
382 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_klassptr()->exact_klass()));
383 break;
384 case Type::MetadataPtr:
385 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_metadataptr()->metadata()));
386 break;
387 case Type::NarrowOop:
388 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_oopptr()->const_oop()));
389 break;
390 case Type::RawPtr:
391 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,p2i(t->is_rawptr()));
392 break;
393 case Type::DoubleCon:
394 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
395 break;
396 case Type::FloatCon:
397 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
398 break;
399 case Type::Long:
400 st->print(" %s%d]=#" INT64_FORMAT,msg,i,(int64_t)(t->is_long()->get_con()));
401 break;
402 case Type::Half:
403 case Type::Top:
404 st->print(" %s%d]=_",msg,i);
405 break;
406 default: ShouldNotReachHere();
407 }
408 }
409 }
410
411 //---------------------print_method_with_lineno--------------------------------
412 void JVMState::print_method_with_lineno(outputStream* st, bool show_name) const {
413 if (show_name) _method->print_short_name(st);
414
415 int lineno = _method->line_number_from_bci(_bci);
416 if (lineno != -1) {
417 st->print(" @ bci:%d (line %d)", _bci, lineno);
418 } else {
419 st->print(" @ bci:%d", _bci);
420 }
421 }
422
423 //------------------------------format-----------------------------------------
424 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
425 st->print(" #");
426 if (_method) {
427 print_method_with_lineno(st, true);
428 } else {
429 st->print_cr(" runtime stub ");
430 return;
431 }
432 if (n->is_MachSafePoint()) {
433 GrowableArray<SafePointScalarObjectNode*> scobjs;
434 MachSafePointNode *mcall = n->as_MachSafePoint();
435 uint i;
436 // Print locals
437 for (i = 0; i < (uint)loc_size(); i++)
438 format_helper(regalloc, st, mcall->local(this, i), "L[", i, &scobjs);
439 // Print stack
440 for (i = 0; i < (uint)stk_size(); i++) {
441 if ((uint)(_stkoff + i) >= mcall->len())
442 st->print(" oob ");
443 else
444 format_helper(regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs);
445 }
446 for (i = 0; (int)i < nof_monitors(); i++) {
447 Node *box = mcall->monitor_box(this, i);
448 Node *obj = mcall->monitor_obj(this, i);
449 if (regalloc->node_regs_max_index() > 0 &&
450 OptoReg::is_valid(regalloc->get_reg_first(box))) {
451 box = BoxLockNode::box_node(box);
452 format_helper(regalloc, st, box, "MON-BOX[", i, &scobjs);
453 } else {
454 OptoReg::Name box_reg = BoxLockNode::reg(box);
455 st->print(" MON-BOX%d=%s+%d",
456 i,
457 OptoReg::regname(OptoReg::c_frame_pointer),
458 regalloc->reg2offset(box_reg));
459 }
460 const char* obj_msg = "MON-OBJ[";
461 if (EliminateLocks) {
462 if (BoxLockNode::box_node(box)->is_eliminated())
463 obj_msg = "MON-OBJ(LOCK ELIMINATED)[";
464 }
465 format_helper(regalloc, st, obj, obj_msg, i, &scobjs);
466 }
467
468 for (i = 0; i < (uint)scobjs.length(); i++) {
469 // Scalar replaced objects.
470 st->cr();
471 st->print(" # ScObj" INT32_FORMAT " ", i);
472 SafePointScalarObjectNode* spobj = scobjs.at(i);
473 ciKlass* cik = spobj->bottom_type()->is_oopptr()->exact_klass();
474 assert(cik->is_instance_klass() ||
475 cik->is_array_klass(), "Not supported allocation.");
476 ciInstanceKlass *iklass = nullptr;
477 if (cik->is_instance_klass()) {
478 cik->print_name_on(st);
479 iklass = cik->as_instance_klass();
480 } else if (cik->is_type_array_klass()) {
481 cik->as_array_klass()->base_element_type()->print_name_on(st);
482 st->print("[%d]", spobj->n_fields());
483 } else if (cik->is_obj_array_klass()) {
484 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass();
485 if (cie->is_instance_klass()) {
486 cie->print_name_on(st);
487 } else if (cie->is_type_array_klass()) {
488 cie->as_array_klass()->base_element_type()->print_name_on(st);
489 } else {
490 ShouldNotReachHere();
491 }
492 st->print("[%d]", spobj->n_fields());
493 int ndim = cik->as_array_klass()->dimension() - 1;
494 while (ndim-- > 0) {
495 st->print("[]");
496 }
497 } else {
498 assert(false, "unexpected type %s", cik->name()->as_utf8());
499 }
500 st->print("={");
501 uint nf = spobj->n_fields();
502 if (nf > 0) {
503 uint first_ind = spobj->first_index(mcall->jvms());
504 if (iklass != nullptr && iklass->is_inlinetype()) {
505 Node* null_marker = mcall->in(first_ind++);
506 if (!null_marker->is_top()) {
507 st->print(" [null marker");
508 format_helper(regalloc, st, null_marker, ":", -1, nullptr);
509 }
510 }
511 Node* fld_node = mcall->in(first_ind);
512 if (iklass != nullptr) {
513 st->print(" [");
514 iklass->nonstatic_field_at(0)->print_name_on(st);
515 format_helper(regalloc, st, fld_node, ":", 0, &scobjs);
516 } else {
517 format_helper(regalloc, st, fld_node, "[", 0, &scobjs);
518 }
519 for (uint j = 1; j < nf; j++) {
520 fld_node = mcall->in(first_ind+j);
521 if (iklass != nullptr) {
522 st->print(", [");
523 iklass->nonstatic_field_at(j)->print_name_on(st);
524 format_helper(regalloc, st, fld_node, ":", j, &scobjs);
525 } else {
526 format_helper(regalloc, st, fld_node, ", [", j, &scobjs);
527 }
528 }
529 }
530 st->print(" }");
531 }
532 }
533 st->cr();
534 if (caller() != nullptr) caller()->format(regalloc, n, st);
535 }
536
537
538 void JVMState::dump_spec(outputStream *st) const {
539 if (_method != nullptr) {
540 bool printed = false;
541 if (!Verbose) {
542 // The JVMS dumps make really, really long lines.
543 // Take out the most boring parts, which are the package prefixes.
544 char buf[500];
545 stringStream namest(buf, sizeof(buf));
546 _method->print_short_name(&namest);
547 if (namest.count() < sizeof(buf)) {
548 const char* name = namest.base();
549 if (name[0] == ' ') ++name;
550 const char* endcn = strchr(name, ':'); // end of class name
551 if (endcn == nullptr) endcn = strchr(name, '(');
552 if (endcn == nullptr) endcn = name + strlen(name);
553 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
554 --endcn;
555 st->print(" %s", endcn);
556 printed = true;
557 }
558 }
559 print_method_with_lineno(st, !printed);
560 if(_reexecute == Reexecute_True)
561 st->print(" reexecute");
562 } else {
563 st->print(" runtime stub");
564 }
565 if (caller() != nullptr) caller()->dump_spec(st);
566 }
567
568
569 void JVMState::dump_on(outputStream* st) const {
570 bool print_map = _map && !((uintptr_t)_map & 1) &&
571 ((caller() == nullptr) || (caller()->map() != _map));
572 if (print_map) {
573 if (_map->len() > _map->req()) { // _map->has_exceptions()
574 Node* ex = _map->in(_map->req()); // _map->next_exception()
575 // skip the first one; it's already being printed
576 while (ex != nullptr && ex->len() > ex->req()) {
577 ex = ex->in(ex->req()); // ex->next_exception()
578 ex->dump(1);
579 }
580 }
581 _map->dump(Verbose ? 2 : 1);
582 }
583 if (caller() != nullptr) {
584 caller()->dump_on(st);
585 }
586 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=",
587 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false");
588 if (_method == nullptr) {
589 st->print_cr("(none)");
590 } else {
591 _method->print_name(st);
592 st->cr();
593 if (bci() >= 0 && bci() < _method->code_size()) {
594 st->print(" bc: ");
595 _method->print_codes_on(bci(), bci()+1, st);
596 }
597 }
598 }
599
600 // Extra way to dump a jvms from the debugger,
601 // to avoid a bug with C++ member function calls.
602 void dump_jvms(JVMState* jvms) {
603 jvms->dump();
604 }
605 #endif
606
607 //--------------------------clone_shallow--------------------------------------
608 JVMState* JVMState::clone_shallow(Compile* C) const {
609 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
610 n->set_bci(_bci);
611 n->_reexecute = _reexecute;
612 n->set_locoff(_locoff);
613 n->set_stkoff(_stkoff);
614 n->set_monoff(_monoff);
615 n->set_scloff(_scloff);
616 n->set_endoff(_endoff);
617 n->set_sp(_sp);
618 n->set_map(_map);
619 n->set_receiver_info(_receiver_info);
620 return n;
621 }
622
623 //---------------------------clone_deep----------------------------------------
624 JVMState* JVMState::clone_deep(Compile* C) const {
625 JVMState* n = clone_shallow(C);
626 for (JVMState* p = n; p->_caller != nullptr; p = p->_caller) {
627 p->_caller = p->_caller->clone_shallow(C);
628 }
629 assert(n->depth() == depth(), "sanity");
630 assert(n->debug_depth() == debug_depth(), "sanity");
631 return n;
632 }
633
634 /**
635 * Reset map for all callers
636 */
637 void JVMState::set_map_deep(SafePointNode* map) {
638 for (JVMState* p = this; p != nullptr; p = p->_caller) {
639 p->set_map(map);
640 }
641 }
642
643 // unlike set_map(), this is two-way setting.
644 void JVMState::bind_map(SafePointNode* map) {
645 set_map(map);
646 _map->set_jvms(this);
647 }
648
649 // Adapt offsets in in-array after adding or removing an edge.
650 // Prerequisite is that the JVMState is used by only one node.
651 void JVMState::adapt_position(int delta) {
652 for (JVMState* jvms = this; jvms != nullptr; jvms = jvms->caller()) {
653 jvms->set_locoff(jvms->locoff() + delta);
654 jvms->set_stkoff(jvms->stkoff() + delta);
655 jvms->set_monoff(jvms->monoff() + delta);
656 jvms->set_scloff(jvms->scloff() + delta);
657 jvms->set_endoff(jvms->endoff() + delta);
658 }
659 }
660
661 // Mirror the stack size calculation in the deopt code
662 // How much stack space would we need at this point in the program in
663 // case of deoptimization?
664 int JVMState::interpreter_frame_size() const {
665 const JVMState* jvms = this;
666 int size = 0;
667 int callee_parameters = 0;
668 int callee_locals = 0;
669 int extra_args = method()->max_stack() - stk_size();
670
671 while (jvms != nullptr) {
672 int locks = jvms->nof_monitors();
673 int temps = jvms->stk_size();
674 bool is_top_frame = (jvms == this);
675 ciMethod* method = jvms->method();
676
677 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(),
678 temps + callee_parameters,
679 extra_args,
680 locks,
681 callee_parameters,
682 callee_locals,
683 is_top_frame);
684 size += frame_size;
685
686 callee_parameters = method->size_of_parameters();
687 callee_locals = method->max_locals();
688 extra_args = 0;
689 jvms = jvms->caller();
690 }
691 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord;
692 }
693
694 // Compute receiver info for a compiled lambda form at call site.
695 ciInstance* JVMState::compute_receiver_info(ciMethod* callee) const {
696 assert(callee != nullptr && callee->is_compiled_lambda_form(), "");
697 if (has_method() && method()->is_compiled_lambda_form()) { // callee is not a MH invoker
698 Node* recv = map()->argument(this, 0);
699 assert(recv != nullptr, "");
700 const TypeOopPtr* recv_toop = recv->bottom_type()->isa_oopptr();
701 if (recv_toop != nullptr && recv_toop->const_oop() != nullptr) {
702 return recv_toop->const_oop()->as_instance();
703 }
704 }
705 return nullptr;
706 }
707
708 //=============================================================================
709 bool CallNode::cmp( const Node &n ) const
710 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
711 #ifndef PRODUCT
712 void CallNode::dump_req(outputStream *st, DumpConfig* dc) const {
713 // Dump the required inputs, enclosed in '(' and ')'
714 uint i; // Exit value of loop
715 for (i = 0; i < req(); i++) { // For all required inputs
716 if (i == TypeFunc::Parms) st->print("(");
717 Node* p = in(i);
718 if (p != nullptr) {
719 p->dump_idx(false, st, dc);
720 st->print(" ");
721 } else {
722 st->print("_ ");
723 }
724 }
725 st->print(")");
726 }
727
728 void CallNode::dump_spec(outputStream *st) const {
729 st->print(" ");
730 if (tf() != nullptr) tf()->dump_on(st);
731 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt);
732 if (jvms() != nullptr) jvms()->dump_spec(st);
733 }
734
735 void AllocateNode::dump_spec(outputStream* st) const {
736 st->print(" ");
737 if (tf() != nullptr) {
738 tf()->dump_on(st);
739 }
740 if (_cnt != COUNT_UNKNOWN) {
741 st->print(" C=%f", _cnt);
742 }
743 const Node* const klass_node = in(KlassNode);
744 if (klass_node != nullptr) {
745 const TypeKlassPtr* const klass_ptr = klass_node->bottom_type()->isa_klassptr();
746
747 if (klass_ptr != nullptr && klass_ptr->klass_is_exact()) {
748 st->print(" allocationKlass:");
749 klass_ptr->exact_klass()->print_name_on(st);
750 }
751 }
752 if (jvms() != nullptr) {
753 jvms()->dump_spec(st);
754 }
755 }
756 #endif
757
758 const Type *CallNode::bottom_type() const { return tf()->range_cc(); }
759 const Type* CallNode::Value(PhaseGVN* phase) const {
760 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
761 return Type::TOP;
762 }
763 return tf()->range_cc();
764 }
765
766 //------------------------------calling_convention-----------------------------
767 void CallNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
768 if (_entry_point == StubRoutines::store_inline_type_fields_to_buf()) {
769 // The call to that stub is a special case: its inputs are
770 // multiple values returned from a call and so it should follow
771 // the return convention.
772 SharedRuntime::java_return_convention(sig_bt, parm_regs, argcnt);
773 return;
774 }
775 // Use the standard compiler calling convention
776 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt);
777 }
778
779
780 //------------------------------match------------------------------------------
781 // Construct projections for control, I/O, memory-fields, ..., and
782 // return result(s) along with their RegMask info
783 Node *CallNode::match(const ProjNode *proj, const Matcher *match, const RegMask* mask) {
784 uint con = proj->_con;
785 const TypeTuple* range_cc = tf()->range_cc();
786 if (con >= TypeFunc::Parms) {
787 if (tf()->returns_inline_type_as_fields()) {
788 // The call returns multiple values (inline type fields): we
789 // create one projection per returned value.
790 assert(con <= TypeFunc::Parms+1 || InlineTypeReturnedAsFields, "only for multi value return");
791 uint ideal_reg = range_cc->field_at(con)->ideal_reg();
792 return new MachProjNode(this, con, mask[con-TypeFunc::Parms], ideal_reg);
793 } else {
794 if (con == TypeFunc::Parms) {
795 uint ideal_reg = range_cc->field_at(TypeFunc::Parms)->ideal_reg();
796 OptoRegPair regs = Opcode() == Op_CallLeafVector
797 ? match->vector_return_value(ideal_reg) // Calls into assembly vector routine
798 : match->c_return_value(ideal_reg);
799 RegMask rm = RegMask(regs.first());
800
801 if (Opcode() == Op_CallLeafVector) {
802 // If the return is in vector, compute appropriate regmask taking into account the whole range
803 if(ideal_reg >= Op_VecA && ideal_reg <= Op_VecZ) {
804 if(OptoReg::is_valid(regs.second())) {
805 for (OptoReg::Name r = regs.first(); r <= regs.second(); r = OptoReg::add(r, 1)) {
806 rm.insert(r);
807 }
808 }
809 }
810 }
811
812 if (OptoReg::is_valid(regs.second())) {
813 rm.insert(regs.second());
814 }
815 return new MachProjNode(this,con,rm,ideal_reg);
816 } else {
817 assert(con == TypeFunc::Parms+1, "only one return value");
818 assert(range_cc->field_at(TypeFunc::Parms+1) == Type::HALF, "");
819 return new MachProjNode(this,con, RegMask::EMPTY, (uint)OptoReg::Bad);
820 }
821 }
822 }
823
824 switch (con) {
825 case TypeFunc::Control:
826 case TypeFunc::I_O:
827 case TypeFunc::Memory:
828 return new MachProjNode(this,proj->_con,RegMask::EMPTY,MachProjNode::unmatched_proj);
829
830 case TypeFunc::ReturnAdr:
831 case TypeFunc::FramePtr:
832 default:
833 ShouldNotReachHere();
834 }
835 return nullptr;
836 }
837
838 // Do we Match on this edge index or not? Match no edges
839 uint CallNode::match_edge(uint idx) const {
840 return 0;
841 }
842
843 //
844 // Determine whether the call could modify the field of the specified
845 // instance at the specified offset.
846 //
847 bool CallNode::may_modify(const TypeOopPtr* t_oop, PhaseValues* phase) {
848 assert((t_oop != nullptr), "sanity");
849 if (is_call_to_arraycopystub() && strcmp(_name, "unsafe_arraycopy") != 0) {
850 const TypeTuple* args = _tf->domain_sig();
851 Node* dest = nullptr;
852 // Stubs that can be called once an ArrayCopyNode is expanded have
853 // different signatures. Look for the second pointer argument,
854 // that is the destination of the copy.
855 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
856 if (args->field_at(i)->isa_ptr()) {
857 j++;
858 if (j == 2) {
859 dest = in(i);
860 break;
861 }
862 }
863 }
864 guarantee(dest != nullptr, "Call had only one ptr in, broken IR!");
865 if (phase->type(dest)->isa_rawptr()) {
866 // may happen for an arraycopy that initializes a newly allocated object. Conservatively return true;
867 return true;
868 }
869 if (!dest->is_top() && may_modify_arraycopy_helper(phase->type(dest)->is_oopptr(), t_oop, phase)) {
870 return true;
871 }
872 return false;
873 }
874 if (t_oop->is_known_instance()) {
875 // The instance_id is set only for scalar-replaceable allocations which
876 // are not passed as arguments according to Escape Analysis.
877 return false;
878 }
879 if (t_oop->is_ptr_to_boxed_value()) {
880 ciKlass* boxing_klass = t_oop->is_instptr()->instance_klass();
881 if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) {
882 // Skip unrelated boxing methods.
883 Node* proj = proj_out_or_null(TypeFunc::Parms);
884 if ((proj == nullptr) || (phase->type(proj)->is_instptr()->instance_klass() != boxing_klass)) {
885 return false;
886 }
887 }
888 if (is_CallJava() && as_CallJava()->method() != nullptr) {
889 ciMethod* meth = as_CallJava()->method();
890 if (meth->is_getter()) {
891 return false;
892 }
893 // May modify (by reflection) if an boxing object is passed
894 // as argument or returned.
895 Node* proj = returns_pointer() ? proj_out_or_null(TypeFunc::Parms) : nullptr;
896 if (proj != nullptr) {
897 const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr();
898 if ((inst_t != nullptr) && (!inst_t->klass_is_exact() ||
899 (inst_t->instance_klass() == boxing_klass))) {
900 return true;
901 }
902 }
903 const TypeTuple* d = tf()->domain_cc();
904 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
905 const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr();
906 if ((inst_t != nullptr) && (!inst_t->klass_is_exact() ||
907 (inst_t->instance_klass() == boxing_klass))) {
908 return true;
909 }
910 }
911 return false;
912 }
913 }
914 return true;
915 }
916
917 // Does this call have a direct reference to n other than debug information?
918 bool CallNode::has_non_debug_use(Node* n) {
919 const TypeTuple* d = tf()->domain_cc();
920 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
921 if (in(i) == n) {
922 return true;
923 }
924 }
925 return false;
926 }
927
928 bool CallNode::has_debug_use(Node* n) {
929 if (jvms() != nullptr) {
930 for (uint i = jvms()->debug_start(); i < jvms()->debug_end(); i++) {
931 if (in(i) == n) {
932 return true;
933 }
934 }
935 }
936 return false;
937 }
938
939 // Returns the unique CheckCastPP of a call
940 // or 'this' if there are several CheckCastPP or unexpected uses
941 // or returns null if there is no one.
942 Node *CallNode::result_cast() {
943 Node *cast = nullptr;
944
945 Node *p = proj_out_or_null(TypeFunc::Parms);
946 if (p == nullptr)
947 return nullptr;
948
949 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
950 Node *use = p->fast_out(i);
951 if (use->is_CheckCastPP()) {
952 if (cast != nullptr) {
953 return this; // more than 1 CheckCastPP
954 }
955 cast = use;
956 } else if (!use->is_Initialize() &&
957 !use->is_AddP() &&
958 use->Opcode() != Op_MemBarStoreStore) {
959 // Expected uses are restricted to a CheckCastPP, an Initialize
960 // node, a MemBarStoreStore (clone) and AddP nodes. If we
961 // encounter any other use (a Phi node can be seen in rare
962 // cases) return this to prevent incorrect optimizations.
963 return this;
964 }
965 }
966 return cast;
967 }
968
969
970 CallProjections* CallNode::extract_projections(bool separate_io_proj, bool do_asserts) const {
971 uint max_res = TypeFunc::Parms-1;
972 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
973 ProjNode *pn = fast_out(i)->as_Proj();
974 max_res = MAX2(max_res, pn->_con);
975 }
976
977 assert(max_res < _tf->range_cc()->cnt(), "result out of bounds");
978
979 uint projs_size = sizeof(CallProjections);
980 if (max_res > TypeFunc::Parms) {
981 projs_size += (max_res-TypeFunc::Parms)*sizeof(Node*);
982 }
983 char* projs_storage = resource_allocate_bytes(projs_size);
984 CallProjections* projs = new(projs_storage)CallProjections(max_res - TypeFunc::Parms + 1);
985
986 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
987 ProjNode *pn = fast_out(i)->as_Proj();
988 if (pn->outcnt() == 0) continue;
989 switch (pn->_con) {
990 case TypeFunc::Control:
991 {
992 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
993 projs->fallthrough_proj = pn;
994 const Node* cn = pn->unique_ctrl_out_or_null();
995 if (cn != nullptr && cn->is_Catch()) {
996 ProjNode *cpn = nullptr;
997 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
998 cpn = cn->fast_out(k)->as_Proj();
999 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
1000 if (cpn->_con == CatchProjNode::fall_through_index)
1001 projs->fallthrough_catchproj = cpn;
1002 else {
1003 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
1004 projs->catchall_catchproj = cpn;
1005 }
1006 }
1007 }
1008 break;
1009 }
1010 case TypeFunc::I_O:
1011 if (pn->_is_io_use)
1012 projs->catchall_ioproj = pn;
1013 else
1014 projs->fallthrough_ioproj = pn;
1015 for (DUIterator j = pn->outs(); pn->has_out(j); j++) {
1016 Node* e = pn->out(j);
1017 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) {
1018 assert(projs->exobj == nullptr, "only one");
1019 projs->exobj = e;
1020 }
1021 }
1022 break;
1023 case TypeFunc::Memory:
1024 if (pn->_is_io_use)
1025 projs->catchall_memproj = pn;
1026 else
1027 projs->fallthrough_memproj = pn;
1028 break;
1029 case TypeFunc::Parms:
1030 projs->resproj[0] = pn;
1031 break;
1032 default:
1033 assert(pn->_con <= max_res, "unexpected projection from allocation node.");
1034 projs->resproj[pn->_con-TypeFunc::Parms] = pn;
1035 break;
1036 }
1037 }
1038
1039 // The resproj may not exist because the result could be ignored
1040 // and the exception object may not exist if an exception handler
1041 // swallows the exception but all the other must exist and be found.
1042 do_asserts = do_asserts && !Compile::current()->inlining_incrementally();
1043 assert(!do_asserts || projs->fallthrough_proj != nullptr, "must be found");
1044 assert(!do_asserts || projs->fallthrough_catchproj != nullptr, "must be found");
1045 assert(!do_asserts || projs->fallthrough_memproj != nullptr, "must be found");
1046 assert(!do_asserts || projs->fallthrough_ioproj != nullptr, "must be found");
1047 assert(!do_asserts || projs->catchall_catchproj != nullptr, "must be found");
1048 if (separate_io_proj) {
1049 assert(!do_asserts || projs->catchall_memproj != nullptr, "must be found");
1050 assert(!do_asserts || projs->catchall_ioproj != nullptr, "must be found");
1051 }
1052 return projs;
1053 }
1054
1055 Node* CallNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1056 #ifdef ASSERT
1057 // Validate attached generator
1058 CallGenerator* cg = generator();
1059 if (cg != nullptr) {
1060 assert((is_CallStaticJava() && cg->is_mh_late_inline()) ||
1061 (is_CallDynamicJava() && cg->is_virtual_late_inline()), "mismatch");
1062 }
1063 #endif // ASSERT
1064 return SafePointNode::Ideal(phase, can_reshape);
1065 }
1066
1067 bool CallNode::is_call_to_arraycopystub() const {
1068 if (_name != nullptr && strstr(_name, "arraycopy") != nullptr) {
1069 return true;
1070 }
1071 return false;
1072 }
1073
1074 bool CallNode::is_call_to_multianewarray_stub() const {
1075 if (_name != nullptr &&
1076 strstr(_name, "multianewarray") != nullptr &&
1077 strstr(_name, "C2 runtime") != nullptr) {
1078 return true;
1079 }
1080 return false;
1081 }
1082
1083 //=============================================================================
1084 uint CallJavaNode::size_of() const { return sizeof(*this); }
1085 bool CallJavaNode::cmp( const Node &n ) const {
1086 CallJavaNode &call = (CallJavaNode&)n;
1087 return CallNode::cmp(call) && _method == call._method &&
1088 _override_symbolic_info == call._override_symbolic_info;
1089 }
1090
1091 void CallJavaNode::copy_call_debug_info(PhaseIterGVN* phase, SafePointNode* sfpt) {
1092 // Copy debug information and adjust JVMState information
1093 uint old_dbg_start = sfpt->is_Call() ? sfpt->as_Call()->tf()->domain_sig()->cnt() : (uint)TypeFunc::Parms+1;
1094 uint new_dbg_start = tf()->domain_sig()->cnt();
1095 int jvms_adj = new_dbg_start - old_dbg_start;
1096 assert (new_dbg_start == req(), "argument count mismatch");
1097 Compile* C = phase->C;
1098
1099 // SafePointScalarObject node could be referenced several times in debug info.
1100 // Use Dict to record cloned nodes.
1101 Dict* sosn_map = new Dict(cmpkey,hashkey);
1102 for (uint i = old_dbg_start; i < sfpt->req(); i++) {
1103 Node* old_in = sfpt->in(i);
1104 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
1105 if (old_in != nullptr && old_in->is_SafePointScalarObject()) {
1106 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
1107 bool new_node;
1108 Node* new_in = old_sosn->clone(sosn_map, new_node);
1109 if (new_node) { // New node?
1110 new_in->set_req(0, C->root()); // reset control edge
1111 new_in = phase->transform(new_in); // Register new node.
1112 }
1113 old_in = new_in;
1114 }
1115 add_req(old_in);
1116 }
1117
1118 // JVMS may be shared so clone it before we modify it
1119 set_jvms(sfpt->jvms() != nullptr ? sfpt->jvms()->clone_deep(C) : nullptr);
1120 for (JVMState *jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1121 jvms->set_map(this);
1122 jvms->set_locoff(jvms->locoff()+jvms_adj);
1123 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
1124 jvms->set_monoff(jvms->monoff()+jvms_adj);
1125 jvms->set_scloff(jvms->scloff()+jvms_adj);
1126 jvms->set_endoff(jvms->endoff()+jvms_adj);
1127 }
1128 }
1129
1130 #ifdef ASSERT
1131 bool CallJavaNode::validate_symbolic_info() const {
1132 if (method() == nullptr) {
1133 return true; // call into runtime or uncommon trap
1134 }
1135 Bytecodes::Code bc = jvms()->method()->java_code_at_bci(jvms()->bci());
1136 if (Arguments::is_valhalla_enabled() && (bc == Bytecodes::_if_acmpeq || bc == Bytecodes::_if_acmpne)) {
1137 return true;
1138 }
1139 ciMethod* symbolic_info = jvms()->method()->get_method_at_bci(jvms()->bci());
1140 ciMethod* callee = method();
1141 if (symbolic_info->is_method_handle_intrinsic() && !callee->is_method_handle_intrinsic()) {
1142 assert(override_symbolic_info(), "should be set");
1143 }
1144 assert(ciMethod::is_consistent_info(symbolic_info, callee), "inconsistent info");
1145 return true;
1146 }
1147 #endif
1148
1149 #ifndef PRODUCT
1150 void CallJavaNode::dump_spec(outputStream* st) const {
1151 if( _method ) _method->print_short_name(st);
1152 CallNode::dump_spec(st);
1153 }
1154
1155 void CallJavaNode::dump_compact_spec(outputStream* st) const {
1156 if (_method) {
1157 _method->print_short_name(st);
1158 } else {
1159 st->print("<?>");
1160 }
1161 }
1162 #endif
1163
1164 void CallJavaNode::register_for_late_inline() {
1165 if (generator() != nullptr) {
1166 Compile::current()->prepend_late_inline(generator());
1167 set_generator(nullptr);
1168 } else {
1169 assert(false, "repeated inline attempt");
1170 }
1171 }
1172
1173 //=============================================================================
1174 uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
1175 bool CallStaticJavaNode::cmp( const Node &n ) const {
1176 CallStaticJavaNode &call = (CallStaticJavaNode&)n;
1177 return CallJavaNode::cmp(call);
1178 }
1179
1180 Node* CallStaticJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1181 if (can_reshape && uncommon_trap_request() != 0) {
1182 PhaseIterGVN* igvn = phase->is_IterGVN();
1183 if (remove_unknown_flat_array_load(igvn, control(), memory(), in(TypeFunc::Parms))) {
1184 if (!control()->is_Region()) {
1185 igvn->replace_input_of(this, 0, phase->C->top());
1186 }
1187 return this;
1188 }
1189 }
1190
1191 // Try to replace the runtime call to the substitutability test emitted by acmp if (at least) one operand is a known type
1192 if (can_reshape && !control()->is_top() && method() != nullptr && method()->holder() == phase->C->env()->ValueObjectMethods_klass() &&
1193 (method()->name() == ciSymbols::isSubstitutable_name())) {
1194 Node* left = in(TypeFunc::Parms);
1195 Node* right = in(TypeFunc::Parms + 1);
1196 if (!left->is_top() && !right->is_top() && (left->is_InlineType() || right->is_InlineType())) {
1197 if (!left->is_InlineType()) {
1198 swap(left, right);
1199 }
1200 InlineTypeNode* vt = left->as_InlineType();
1201
1202 // Check if the field layout can be optimized
1203 if (vt->can_emit_substitutability_check(right)) {
1204 PhaseIterGVN* igvn = phase->is_IterGVN();
1205
1206 Node* ctrl = control();
1207 RegionNode* region = new RegionNode(1);
1208 Node* phi = new PhiNode(region, TypeInt::POS);
1209
1210 Node* base = right;
1211 Node* ptr = right;
1212 if (!base->is_InlineType()) {
1213 // Parse time checks guarantee that both operands are non-null and have the same type
1214 base = igvn->register_new_node_with_optimizer(new CheckCastPPNode(ctrl, base, vt->bottom_type()));
1215 ptr = base;
1216 }
1217 // Emit IR for field-wise comparison
1218 vt->check_substitutability(igvn, region, phi, &ctrl, in(TypeFunc::Memory), base, ptr);
1219
1220 // Equals
1221 region->add_req(ctrl);
1222 phi->add_req(igvn->intcon(1));
1223
1224 ctrl = igvn->register_new_node_with_optimizer(region);
1225 Node* res = igvn->register_new_node_with_optimizer(phi);
1226
1227 // Kill exception projections and return a tuple that will replace the call
1228 CallProjections* projs = extract_projections(false /*separate_io_proj*/);
1229 if (projs->fallthrough_catchproj != nullptr) {
1230 igvn->replace_node(projs->fallthrough_catchproj, ctrl);
1231 }
1232 if (projs->catchall_memproj != nullptr) {
1233 igvn->replace_node(projs->catchall_memproj, igvn->C->top());
1234 }
1235 if (projs->catchall_ioproj != nullptr) {
1236 igvn->replace_node(projs->catchall_ioproj, igvn->C->top());
1237 }
1238 if (projs->catchall_catchproj != nullptr) {
1239 igvn->replace_node(projs->catchall_catchproj, igvn->C->top());
1240 }
1241 return TupleNode::make(tf()->range_cc(), ctrl, i_o(), memory(), frameptr(), returnadr(), res);
1242 }
1243 }
1244 }
1245
1246 CallGenerator* cg = generator();
1247 if (can_reshape && cg != nullptr) {
1248 if (cg->is_mh_late_inline()) {
1249 assert(IncrementalInlineMH, "required");
1250 assert(cg->call_node() == this, "mismatch");
1251 assert(cg->method()->is_method_handle_intrinsic(), "required");
1252
1253 // Check whether this MH handle call becomes a candidate for inlining.
1254 ciMethod* callee = cg->method();
1255 vmIntrinsics::ID iid = callee->intrinsic_id();
1256 if (iid == vmIntrinsics::_invokeBasic) {
1257 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) {
1258 register_for_late_inline();
1259 }
1260 } else if (iid == vmIntrinsics::_linkToNative) {
1261 // never retry
1262 } else {
1263 assert(callee->has_member_arg(), "wrong type of call?");
1264 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) {
1265 register_for_late_inline();
1266 }
1267 }
1268 } else {
1269 assert(IncrementalInline, "required");
1270 assert(!cg->method()->is_method_handle_intrinsic(), "required");
1271 if (phase->C->print_inlining()) {
1272 phase->C->inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE,
1273 "static call node changed: trying again");
1274 }
1275 register_for_late_inline();
1276 }
1277 }
1278 return CallNode::Ideal(phase, can_reshape);
1279 }
1280
1281 //----------------------------is_uncommon_trap----------------------------
1282 // Returns true if this is an uncommon trap.
1283 bool CallStaticJavaNode::is_uncommon_trap() const {
1284 return (_name != nullptr && !strcmp(_name, "uncommon_trap"));
1285 }
1286
1287 //----------------------------uncommon_trap_request----------------------------
1288 // If this is an uncommon trap, return the request code, else zero.
1289 int CallStaticJavaNode::uncommon_trap_request() const {
1290 return is_uncommon_trap() ? extract_uncommon_trap_request(this) : 0;
1291 }
1292 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
1293 #ifndef PRODUCT
1294 if (!(call->req() > TypeFunc::Parms &&
1295 call->in(TypeFunc::Parms) != nullptr &&
1296 call->in(TypeFunc::Parms)->is_Con() &&
1297 call->in(TypeFunc::Parms)->bottom_type()->isa_int())) {
1298 assert(in_dump() != 0, "OK if dumping");
1299 tty->print("[bad uncommon trap]");
1300 return 0;
1301 }
1302 #endif
1303 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
1304 }
1305
1306 // Split if can cause the flat array branch of an array load with unknown type (see
1307 // Parse::array_load) to end in an uncommon trap. In that case, the call to
1308 // 'load_unknown_inline' is useless. Replace it with an uncommon trap with the same JVMState.
1309 bool CallStaticJavaNode::remove_unknown_flat_array_load(PhaseIterGVN* igvn, Node* ctl, Node* mem, Node* unc_arg) {
1310 if (ctl == nullptr || ctl->is_top() || mem == nullptr || mem->is_top() || !mem->is_MergeMem()) {
1311 return false;
1312 }
1313 if (ctl->is_Region()) {
1314 bool res = false;
1315 for (uint i = 1; i < ctl->req(); i++) {
1316 MergeMemNode* mm = mem->clone()->as_MergeMem();
1317 for (MergeMemStream mms(mm); mms.next_non_empty(); ) {
1318 Node* m = mms.memory();
1319 if (m->is_Phi() && m->in(0) == ctl) {
1320 mms.set_memory(m->in(i));
1321 }
1322 }
1323 if (remove_unknown_flat_array_load(igvn, ctl->in(i), mm, unc_arg)) {
1324 res = true;
1325 if (!ctl->in(i)->is_Region()) {
1326 igvn->replace_input_of(ctl, i, igvn->C->top());
1327 }
1328 }
1329 igvn->remove_dead_node(mm, PhaseIterGVN::NodeOrigin::Speculative);
1330 }
1331 return res;
1332 }
1333 // Verify the control flow is ok
1334 Node* call = ctl;
1335 MemBarNode* membar = nullptr;
1336 for (;;) {
1337 if (call == nullptr || call->is_top()) {
1338 return false;
1339 }
1340 if (call->is_Proj() || call->is_Catch() || call->is_MemBar()) {
1341 call = call->in(0);
1342 } else if (call->Opcode() == Op_CallStaticJava && !call->in(0)->is_top() &&
1343 call->as_Call()->entry_point() == OptoRuntime::load_unknown_inline_Java()) {
1344 // If there is no explicit flat array accesses in the compilation unit, there would be no
1345 // membar here
1346 if (call->in(0)->is_Proj() && call->in(0)->in(0)->is_MemBar()) {
1347 membar = call->in(0)->in(0)->as_MemBar();
1348 }
1349 break;
1350 } else {
1351 return false;
1352 }
1353 }
1354
1355 JVMState* jvms = call->jvms();
1356 if (igvn->C->too_many_traps(jvms->method(), jvms->bci(), Deoptimization::trap_request_reason(uncommon_trap_request()))) {
1357 return false;
1358 }
1359
1360 Node* call_mem = call->in(TypeFunc::Memory);
1361 if (call_mem == nullptr || call_mem->is_top()) {
1362 return false;
1363 }
1364 if (!call_mem->is_MergeMem()) {
1365 call_mem = MergeMemNode::make(call_mem);
1366 igvn->register_new_node_with_optimizer(call_mem);
1367 }
1368
1369 // Verify that there's no unexpected side effect
1370 for (MergeMemStream mms2(mem->as_MergeMem(), call_mem->as_MergeMem()); mms2.next_non_empty2(); ) {
1371 Node* m1 = mms2.is_empty() ? mms2.base_memory() : mms2.memory();
1372 Node* m2 = mms2.memory2();
1373
1374 for (uint i = 0; i < 100; i++) {
1375 if (m1 == m2) {
1376 break;
1377 } else if (m1->is_Proj()) {
1378 m1 = m1->in(0);
1379 } else if (m1->is_MemBar()) {
1380 m1 = m1->in(TypeFunc::Memory);
1381 } else if (m1->Opcode() == Op_CallStaticJava &&
1382 m1->as_Call()->entry_point() == OptoRuntime::load_unknown_inline_Java()) {
1383 if (m1 != call) {
1384 if (call_mem->outcnt() == 0) {
1385 igvn->remove_dead_node(call_mem, PhaseIterGVN::NodeOrigin::Speculative);
1386 }
1387 return false;
1388 }
1389 break;
1390 } else if (m1->is_MergeMem()) {
1391 MergeMemNode* mm = m1->as_MergeMem();
1392 int idx = mms2.alias_idx();
1393 if (idx == Compile::AliasIdxBot) {
1394 m1 = mm->base_memory();
1395 } else {
1396 m1 = mm->memory_at(idx);
1397 }
1398 } else {
1399 if (call_mem->outcnt() == 0) {
1400 igvn->remove_dead_node(call_mem, PhaseIterGVN::NodeOrigin::Speculative);
1401 }
1402 return false;
1403 }
1404 }
1405 }
1406 if (call_mem->outcnt() == 0) {
1407 igvn->remove_dead_node(call_mem, PhaseIterGVN::NodeOrigin::Speculative);
1408 }
1409
1410 // Remove membar preceding the call
1411 if (membar != nullptr) {
1412 membar->remove(igvn);
1413 }
1414
1415 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
1416 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap", nullptr);
1417 unc->init_req(TypeFunc::Control, call->in(0));
1418 unc->init_req(TypeFunc::I_O, call->in(TypeFunc::I_O));
1419 unc->init_req(TypeFunc::Memory, call->in(TypeFunc::Memory));
1420 unc->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
1421 unc->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
1422 unc->init_req(TypeFunc::Parms+0, unc_arg);
1423 unc->set_cnt(PROB_UNLIKELY_MAG(4));
1424 unc->copy_call_debug_info(igvn, call->as_CallStaticJava());
1425
1426 // Replace the call with an uncommon trap
1427 igvn->replace_input_of(call, 0, igvn->C->top());
1428
1429 igvn->register_new_node_with_optimizer(unc);
1430
1431 Node* ctrl = igvn->transform(new ProjNode(unc, TypeFunc::Control));
1432 Node* halt = igvn->transform(new HaltNode(ctrl, call->in(TypeFunc::FramePtr), "uncommon trap returned which should never happen"));
1433 igvn->add_input_to(igvn->C->root(), halt);
1434
1435 return true;
1436 }
1437
1438
1439 #ifndef PRODUCT
1440 void CallStaticJavaNode::dump_spec(outputStream *st) const {
1441 st->print("# Static ");
1442 if (_name != nullptr) {
1443 st->print("%s", _name);
1444 int trap_req = uncommon_trap_request();
1445 if (trap_req != 0) {
1446 char buf[100];
1447 st->print("(%s)",
1448 Deoptimization::format_trap_request(buf, sizeof(buf),
1449 trap_req));
1450 }
1451 st->print(" ");
1452 }
1453 CallJavaNode::dump_spec(st);
1454 }
1455
1456 void CallStaticJavaNode::dump_compact_spec(outputStream* st) const {
1457 if (_method) {
1458 _method->print_short_name(st);
1459 } else if (_name) {
1460 st->print("%s", _name);
1461 } else {
1462 st->print("<?>");
1463 }
1464 }
1465 #endif
1466
1467 //=============================================================================
1468 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
1469 bool CallDynamicJavaNode::cmp( const Node &n ) const {
1470 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
1471 return CallJavaNode::cmp(call);
1472 }
1473
1474 Node* CallDynamicJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1475 CallGenerator* cg = generator();
1476 if (can_reshape && cg != nullptr) {
1477 if (cg->is_virtual_late_inline()) {
1478 assert(IncrementalInlineVirtual, "required");
1479 assert(cg->call_node() == this, "mismatch");
1480
1481 if (cg->callee_method() == nullptr) {
1482 // Recover symbolic info for method resolution.
1483 ciMethod* caller = jvms()->method();
1484 ciBytecodeStream iter(caller);
1485 iter.force_bci(jvms()->bci());
1486
1487 bool not_used1;
1488 ciSignature* not_used2;
1489 ciMethod* orig_callee = iter.get_method(not_used1, ¬_used2); // callee in the bytecode
1490 ciKlass* holder = iter.get_declared_method_holder();
1491 if (orig_callee->is_method_handle_intrinsic()) {
1492 assert(_override_symbolic_info, "required");
1493 orig_callee = method();
1494 holder = method()->holder();
1495 }
1496
1497 ciInstanceKlass* klass = ciEnv::get_instance_klass_for_declared_method_holder(holder);
1498
1499 Node* receiver_node = in(TypeFunc::Parms);
1500 const TypeOopPtr* receiver_type = phase->type(receiver_node)->isa_oopptr();
1501
1502 int not_used3;
1503 bool call_does_dispatch;
1504 ciMethod* callee = phase->C->optimize_virtual_call(caller, klass, holder, orig_callee, receiver_type, true /*is_virtual*/,
1505 call_does_dispatch, not_used3); // out-parameters
1506 if (!call_does_dispatch) {
1507 cg->set_callee_method(callee);
1508 }
1509 }
1510 if (cg->callee_method() != nullptr) {
1511 // Register for late inlining.
1512 register_for_late_inline(); // MH late inlining prepends to the list, so do the same
1513 }
1514 } else {
1515 assert(IncrementalInline, "required");
1516 if (phase->C->print_inlining()) {
1517 phase->C->inline_printer()->record(cg->method(), cg->call_node()->jvms(), InliningResult::FAILURE,
1518 "dynamic call node changed: trying again");
1519 }
1520 register_for_late_inline();
1521 }
1522 }
1523 return CallNode::Ideal(phase, can_reshape);
1524 }
1525
1526 #ifndef PRODUCT
1527 void CallDynamicJavaNode::dump_spec(outputStream *st) const {
1528 st->print("# Dynamic ");
1529 CallJavaNode::dump_spec(st);
1530 }
1531 #endif
1532
1533 //=============================================================================
1534 uint CallRuntimeNode::size_of() const { return sizeof(*this); }
1535 bool CallRuntimeNode::cmp( const Node &n ) const {
1536 CallRuntimeNode &call = (CallRuntimeNode&)n;
1537 return CallNode::cmp(call) && !strcmp(_name,call._name);
1538 }
1539 #ifndef PRODUCT
1540 void CallRuntimeNode::dump_spec(outputStream *st) const {
1541 st->print("# ");
1542 st->print("%s", _name);
1543 CallNode::dump_spec(st);
1544 }
1545 #endif
1546 uint CallLeafVectorNode::size_of() const { return sizeof(*this); }
1547 bool CallLeafVectorNode::cmp( const Node &n ) const {
1548 CallLeafVectorNode &call = (CallLeafVectorNode&)n;
1549 return CallLeafNode::cmp(call) && _num_bits == call._num_bits;
1550 }
1551
1552 //------------------------------calling_convention-----------------------------
1553 void CallRuntimeNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
1554 if (_entry_point == nullptr) {
1555 // The call to that stub is a special case: its inputs are
1556 // multiple values returned from a call and so it should follow
1557 // the return convention.
1558 SharedRuntime::java_return_convention(sig_bt, parm_regs, argcnt);
1559 return;
1560 }
1561 SharedRuntime::c_calling_convention(sig_bt, parm_regs, argcnt);
1562 }
1563
1564 void CallLeafVectorNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
1565 #ifdef ASSERT
1566 assert(tf()->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1567 "return vector size must match");
1568 const TypeTuple* d = tf()->domain_sig();
1569 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1570 Node* arg = in(i);
1571 assert(arg->bottom_type()->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1572 "vector argument size must match");
1573 }
1574 #endif
1575
1576 SharedRuntime::vector_calling_convention(parm_regs, _num_bits, argcnt);
1577 }
1578
1579 //=============================================================================
1580 //------------------------------calling_convention-----------------------------
1581
1582
1583 //=============================================================================
1584 bool CallLeafPureNode::is_unused() const {
1585 return proj_out_or_null(TypeFunc::Parms) == nullptr;
1586 }
1587
1588 bool CallLeafPureNode::is_dead() const {
1589 return proj_out_or_null(TypeFunc::Control) == nullptr;
1590 }
1591
1592 /* We make a tuple of the global input state + TOP for the output values.
1593 * We use this to delete a pure function that is not used: by replacing the call with
1594 * such a tuple, we let output Proj's idealization pick the corresponding input of the
1595 * pure call, so jumping over it, and effectively, removing the call from the graph.
1596 * This avoids doing the graph surgery manually, but leaves that to IGVN
1597 * that is specialized for doing that right. We need also tuple components for output
1598 * values of the function to respect the return arity, and in case there is a projection
1599 * that would pick an output (which shouldn't happen at the moment).
1600 */
1601 TupleNode* CallLeafPureNode::make_tuple_of_input_state_and_top_return_values(const Compile* C) const {
1602 // Transparently propagate input state but parameters
1603 TupleNode* tuple = TupleNode::make(
1604 tf()->range_cc(),
1605 in(TypeFunc::Control),
1606 in(TypeFunc::I_O),
1607 in(TypeFunc::Memory),
1608 in(TypeFunc::FramePtr),
1609 in(TypeFunc::ReturnAdr));
1610
1611 // And add TOPs for the return values
1612 for (uint i = TypeFunc::Parms; i < tf()->range_cc()->cnt(); i++) {
1613 tuple->set_req(i, C->top());
1614 }
1615
1616 return tuple;
1617 }
1618
1619 CallLeafPureNode* CallLeafPureNode::inline_call_leaf_pure_node(Node* control) const {
1620 Node* top = Compile::current()->top();
1621 if (control == nullptr) {
1622 control = in(TypeFunc::Control);
1623 }
1624
1625 CallLeafPureNode* call = new CallLeafPureNode(tf(), entry_point(), _name);
1626 call->init_req(TypeFunc::Control, control);
1627 call->init_req(TypeFunc::I_O, top);
1628 call->init_req(TypeFunc::Memory, top);
1629 call->init_req(TypeFunc::ReturnAdr, top);
1630 call->init_req(TypeFunc::FramePtr, top);
1631 for (unsigned int i = 0; i < tf()->domain_cc()->cnt() - TypeFunc::Parms; i++) {
1632 call->init_req(TypeFunc::Parms + i, in(TypeFunc::Parms + i));
1633 }
1634
1635 return call;
1636 }
1637
1638 Node* CallLeafPureNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1639 if (is_dead()) {
1640 return nullptr;
1641 }
1642
1643 // We need to wait until IGVN because during parsing, usages might still be missing
1644 // and we would remove the call immediately.
1645 if (can_reshape && is_unused()) {
1646 // The result is not used. We remove the call by replacing it with a tuple, that
1647 // is later disintegrated by the projections.
1648 return make_tuple_of_input_state_and_top_return_values(phase->C);
1649 }
1650
1651 return CallRuntimeNode::Ideal(phase, can_reshape);
1652 }
1653
1654 #ifndef PRODUCT
1655 void CallLeafNode::dump_spec(outputStream *st) const {
1656 st->print("# ");
1657 st->print("%s", _name);
1658 CallNode::dump_spec(st);
1659 }
1660 #endif
1661
1662 uint CallLeafNoFPNode::match_edge(uint idx) const {
1663 // Null entry point is a special case for which the target is in a
1664 // register. Need to match that edge.
1665 return entry_point() == nullptr && idx == TypeFunc::Parms;
1666 }
1667
1668 //=============================================================================
1669
1670 void SafePointNode::set_local(const JVMState* jvms, uint idx, Node *c) {
1671 assert(verify_jvms(jvms), "jvms must match");
1672 int loc = jvms->locoff() + idx;
1673 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
1674 // If current local idx is top then local idx - 1 could
1675 // be a long/double that needs to be killed since top could
1676 // represent the 2nd half of the long/double.
1677 uint ideal = in(loc -1)->ideal_reg();
1678 if (ideal == Op_RegD || ideal == Op_RegL) {
1679 // set other (low index) half to top
1680 set_req(loc - 1, in(loc));
1681 }
1682 }
1683 set_req(loc, c);
1684 }
1685
1686 uint SafePointNode::size_of() const { return sizeof(*this); }
1687 bool SafePointNode::cmp( const Node &n ) const {
1688 return (&n == this); // Always fail except on self
1689 }
1690
1691 //-------------------------set_next_exception----------------------------------
1692 void SafePointNode::set_next_exception(SafePointNode* n) {
1693 assert(n == nullptr || n->Opcode() == Op_SafePoint, "correct value for next_exception");
1694 if (len() == req()) {
1695 if (n != nullptr) add_prec(n);
1696 } else {
1697 set_prec(req(), n);
1698 }
1699 }
1700
1701
1702 //----------------------------next_exception-----------------------------------
1703 SafePointNode* SafePointNode::next_exception() const {
1704 if (len() == req()) {
1705 return nullptr;
1706 } else {
1707 Node* n = in(req());
1708 assert(n == nullptr || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
1709 return (SafePointNode*) n;
1710 }
1711 }
1712
1713
1714 //------------------------------Ideal------------------------------------------
1715 // Skip over any collapsed Regions
1716 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1717 assert(_jvms == nullptr || ((uintptr_t)_jvms->map() & 1) || _jvms->map() == this, "inconsistent JVMState");
1718 if (remove_dead_region(phase, can_reshape)) {
1719 return this;
1720 }
1721 // Scalarize inline types in safepoint debug info.
1722 // Delay this until all inlining is over to avoid getting inconsistent debug info.
1723 if (phase->C->scalarize_in_safepoints() && can_reshape && jvms() != nullptr) {
1724 for (uint i = jvms()->debug_start(); i < jvms()->debug_end(); i++) {
1725 Node* n = in(i)->uncast();
1726 if (n->is_InlineType()) {
1727 n->as_InlineType()->make_scalar_in_safepoints(phase->is_IterGVN());
1728 }
1729 }
1730 }
1731 return nullptr;
1732 }
1733
1734 //------------------------------Identity---------------------------------------
1735 // Remove obviously duplicate safepoints
1736 Node* SafePointNode::Identity(PhaseGVN* phase) {
1737
1738 // If you have back to back safepoints, remove one
1739 if (in(TypeFunc::Control)->is_SafePoint()) {
1740 Node* out_c = unique_ctrl_out_or_null();
1741 // This can be the safepoint of an outer strip mined loop if the inner loop's backedge was removed. Replacing the
1742 // outer loop's safepoint could confuse removal of the outer loop.
1743 if (out_c != nullptr && !out_c->is_OuterStripMinedLoopEnd()) {
1744 return in(TypeFunc::Control);
1745 }
1746 }
1747
1748 // Transforming long counted loops requires a safepoint node. Do not
1749 // eliminate a safepoint until loop opts are over.
1750 if (in(0)->is_Proj() && !phase->C->major_progress()) {
1751 Node *n0 = in(0)->in(0);
1752 // Check if he is a call projection (except Leaf Call)
1753 if( n0->is_Catch() ) {
1754 n0 = n0->in(0)->in(0);
1755 assert( n0->is_Call(), "expect a call here" );
1756 }
1757 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
1758 // Don't remove a safepoint belonging to an OuterStripMinedLoopEndNode.
1759 // If the loop dies, they will be removed together.
1760 if (has_out_with(Op_OuterStripMinedLoopEnd)) {
1761 return this;
1762 }
1763 // Useless Safepoint, so remove it
1764 return in(TypeFunc::Control);
1765 }
1766 }
1767
1768 return this;
1769 }
1770
1771 //------------------------------Value------------------------------------------
1772 const Type* SafePointNode::Value(PhaseGVN* phase) const {
1773 if (phase->type(in(0)) == Type::TOP) {
1774 return Type::TOP;
1775 }
1776 if (in(0) == this) {
1777 return Type::TOP; // Dead infinite loop
1778 }
1779 return Type::CONTROL;
1780 }
1781
1782 #ifndef PRODUCT
1783 void SafePointNode::dump_spec(outputStream *st) const {
1784 st->print(" SafePoint ");
1785 _replaced_nodes.dump(st);
1786 }
1787 #endif
1788
1789 const RegMask &SafePointNode::in_RegMask(uint idx) const {
1790 if (idx < TypeFunc::Parms) {
1791 return RegMask::EMPTY;
1792 }
1793 // Values outside the domain represent debug info
1794 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1795 }
1796 const RegMask &SafePointNode::out_RegMask() const {
1797 return RegMask::EMPTY;
1798 }
1799
1800
1801 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
1802 assert((int)grow_by > 0, "sanity");
1803 int monoff = jvms->monoff();
1804 int scloff = jvms->scloff();
1805 int endoff = jvms->endoff();
1806 assert(endoff == (int)req(), "no other states or debug info after me");
1807 Node* top = Compile::current()->top();
1808 for (uint i = 0; i < grow_by; i++) {
1809 ins_req(monoff, top);
1810 }
1811 jvms->set_monoff(monoff + grow_by);
1812 jvms->set_scloff(scloff + grow_by);
1813 jvms->set_endoff(endoff + grow_by);
1814 }
1815
1816 void SafePointNode::push_monitor(const FastLockNode *lock) {
1817 // Add a LockNode, which points to both the original BoxLockNode (the
1818 // stack space for the monitor) and the Object being locked.
1819 const int MonitorEdges = 2;
1820 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1821 assert(req() == jvms()->endoff(), "correct sizing");
1822 int nextmon = jvms()->scloff();
1823 ins_req(nextmon, lock->box_node());
1824 ins_req(nextmon+1, lock->obj_node());
1825 jvms()->set_scloff(nextmon + MonitorEdges);
1826 jvms()->set_endoff(req());
1827 }
1828
1829 void SafePointNode::pop_monitor() {
1830 // Delete last monitor from debug info
1831 DEBUG_ONLY(int num_before_pop = jvms()->nof_monitors());
1832 const int MonitorEdges = 2;
1833 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1834 int scloff = jvms()->scloff();
1835 int endoff = jvms()->endoff();
1836 int new_scloff = scloff - MonitorEdges;
1837 int new_endoff = endoff - MonitorEdges;
1838 jvms()->set_scloff(new_scloff);
1839 jvms()->set_endoff(new_endoff);
1840 while (scloff > new_scloff) del_req_ordered(--scloff);
1841 assert(jvms()->nof_monitors() == num_before_pop-1, "");
1842 }
1843
1844 Node *SafePointNode::peek_monitor_box() const {
1845 int mon = jvms()->nof_monitors() - 1;
1846 assert(mon >= 0, "must have a monitor");
1847 return monitor_box(jvms(), mon);
1848 }
1849
1850 Node *SafePointNode::peek_monitor_obj() const {
1851 int mon = jvms()->nof_monitors() - 1;
1852 assert(mon >= 0, "must have a monitor");
1853 return monitor_obj(jvms(), mon);
1854 }
1855
1856 Node* SafePointNode::peek_operand(uint off) const {
1857 assert(jvms()->sp() > 0, "must have an operand");
1858 assert(off < jvms()->sp(), "off is out-of-range");
1859 return stack(jvms(), jvms()->sp() - off - 1);
1860 }
1861
1862 // Do we Match on this edge index or not? Match no edges
1863 uint SafePointNode::match_edge(uint idx) const {
1864 return (TypeFunc::Parms == idx);
1865 }
1866
1867 void SafePointNode::disconnect_from_root(PhaseIterGVN *igvn) {
1868 assert(Opcode() == Op_SafePoint, "only value for safepoint in loops");
1869 int nb = igvn->C->root()->find_prec_edge(this);
1870 if (nb != -1) {
1871 igvn->delete_precedence_of(igvn->C->root(), nb);
1872 }
1873 }
1874
1875 //============== SafePointScalarObjectNode ==============
1876
1877 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp, Node* alloc, uint first_index, uint depth, uint n_fields) :
1878 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1879 _first_index(first_index),
1880 _depth(depth),
1881 _n_fields(n_fields),
1882 _alloc(alloc)
1883 {
1884 #ifdef ASSERT
1885 if (alloc != nullptr && !alloc->is_Allocate() && !(alloc->Opcode() == Op_VectorBox)) {
1886 alloc->dump();
1887 assert(false, "unexpected call node");
1888 }
1889 #endif
1890 init_class_id(Class_SafePointScalarObject);
1891 }
1892
1893 // Do not allow value-numbering for SafePointScalarObject node.
1894 uint SafePointScalarObjectNode::hash() const { return NO_HASH; }
1895 bool SafePointScalarObjectNode::cmp( const Node &n ) const {
1896 return (&n == this); // Always fail except on self
1897 }
1898
1899 uint SafePointScalarObjectNode::ideal_reg() const {
1900 return 0; // No matching to machine instruction
1901 }
1902
1903 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1904 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1905 }
1906
1907 const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1908 return RegMask::EMPTY;
1909 }
1910
1911 uint SafePointScalarObjectNode::match_edge(uint idx) const {
1912 return 0;
1913 }
1914
1915 SafePointScalarObjectNode*
1916 SafePointScalarObjectNode::clone(Dict* sosn_map, bool& new_node) const {
1917 void* cached = (*sosn_map)[(void*)this];
1918 if (cached != nullptr) {
1919 new_node = false;
1920 return (SafePointScalarObjectNode*)cached;
1921 }
1922 new_node = true;
1923 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1924 sosn_map->Insert((void*)this, (void*)res);
1925 return res;
1926 }
1927
1928
1929 #ifndef PRODUCT
1930 void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1931 st->print(" # fields@[%d..%d]", first_index(), first_index() + n_fields() - 1);
1932 }
1933 #endif
1934
1935 //============== SafePointScalarMergeNode ==============
1936
1937 SafePointScalarMergeNode::SafePointScalarMergeNode(const TypeOopPtr* tp, int merge_pointer_idx) :
1938 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1939 _merge_pointer_idx(merge_pointer_idx)
1940 {
1941 init_class_id(Class_SafePointScalarMerge);
1942 }
1943
1944 // Do not allow value-numbering for SafePointScalarMerge node.
1945 uint SafePointScalarMergeNode::hash() const { return NO_HASH; }
1946 bool SafePointScalarMergeNode::cmp( const Node &n ) const {
1947 return (&n == this); // Always fail except on self
1948 }
1949
1950 uint SafePointScalarMergeNode::ideal_reg() const {
1951 return 0; // No matching to machine instruction
1952 }
1953
1954 const RegMask &SafePointScalarMergeNode::in_RegMask(uint idx) const {
1955 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1956 }
1957
1958 const RegMask &SafePointScalarMergeNode::out_RegMask() const {
1959 return RegMask::EMPTY;
1960 }
1961
1962 uint SafePointScalarMergeNode::match_edge(uint idx) const {
1963 return 0;
1964 }
1965
1966 SafePointScalarMergeNode*
1967 SafePointScalarMergeNode::clone(Dict* sosn_map, bool& new_node) const {
1968 void* cached = (*sosn_map)[(void*)this];
1969 if (cached != nullptr) {
1970 new_node = false;
1971 return (SafePointScalarMergeNode*)cached;
1972 }
1973 new_node = true;
1974 SafePointScalarMergeNode* res = (SafePointScalarMergeNode*)Node::clone();
1975 sosn_map->Insert((void*)this, (void*)res);
1976 return res;
1977 }
1978
1979 #ifndef PRODUCT
1980 void SafePointScalarMergeNode::dump_spec(outputStream *st) const {
1981 st->print(" # merge_pointer_idx=%d, scalarized_objects=%d", _merge_pointer_idx, req()-1);
1982 }
1983 #endif
1984
1985 //=============================================================================
1986 uint AllocateNode::size_of() const { return sizeof(*this); }
1987
1988 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1989 Node *ctrl, Node *mem, Node *abio,
1990 Node *size, Node *klass_node,
1991 Node* initial_test,
1992 InlineTypeNode* inline_type_node)
1993 : CallNode(atype, nullptr, TypeRawPtr::BOTTOM)
1994 {
1995 init_class_id(Class_Allocate);
1996 init_flags(Flag_is_macro);
1997 _is_scalar_replaceable = false;
1998 _is_non_escaping = false;
1999 _is_allocation_MemBar_redundant = false;
2000 Node *topnode = C->top();
2001
2002 init_req( TypeFunc::Control , ctrl );
2003 init_req( TypeFunc::I_O , abio );
2004 init_req( TypeFunc::Memory , mem );
2005 init_req( TypeFunc::ReturnAdr, topnode );
2006 init_req( TypeFunc::FramePtr , topnode );
2007 init_req( AllocSize , size);
2008 init_req( KlassNode , klass_node);
2009 init_req( InitialTest , initial_test);
2010 init_req( ALength , topnode);
2011 init_req( ValidLengthTest , topnode);
2012 init_req( InlineType , inline_type_node);
2013 // DefaultValue defaults to nullptr
2014 // RawDefaultValue defaults to nullptr
2015 C->add_macro_node(this);
2016 }
2017
2018 void AllocateNode::compute_MemBar_redundancy(ciMethod* initializer)
2019 {
2020 assert(initializer != nullptr &&
2021 (initializer->is_object_constructor() || initializer->is_class_initializer()),
2022 "unexpected initializer method");
2023 BCEscapeAnalyzer* analyzer = initializer->get_bcea();
2024 if (analyzer == nullptr) {
2025 return;
2026 }
2027
2028 // Allocation node is first parameter in its initializer
2029 if (analyzer->is_arg_stack(0) || analyzer->is_arg_local(0)) {
2030 _is_allocation_MemBar_redundant = true;
2031 }
2032 }
2033
2034 Node* AllocateNode::make_ideal_mark(PhaseGVN* phase, Node* control, Node* mem) {
2035 Node* mark_node = nullptr;
2036 if (UseCompactObjectHeaders || Arguments::is_valhalla_enabled()) {
2037 Node* klass_node = in(AllocateNode::KlassNode);
2038 Node* proto_adr = phase->transform(AddPNode::make_with_base(phase->C->top(), klass_node, phase->MakeConX(in_bytes(Klass::prototype_header_offset()))));
2039 mark_node = LoadNode::make(*phase, control, mem, proto_adr, phase->type(proto_adr)->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
2040 } else {
2041 // For now only enable fast locking for non-array types
2042 mark_node = phase->MakeConX(markWord::prototype().value());
2043 }
2044 return mark_node;
2045 }
2046
2047 // Retrieve the length from the AllocateArrayNode. Narrow the type with a
2048 // CastII, if appropriate. If we are not allowed to create new nodes, and
2049 // a CastII is appropriate, return null.
2050 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseValues* phase, bool allow_new_nodes) {
2051 Node *length = in(AllocateNode::ALength);
2052 assert(length != nullptr, "length is not null");
2053
2054 const TypeInt* length_type = phase->find_int_type(length);
2055 const TypeAryPtr* ary_type = oop_type->isa_aryptr();
2056
2057 if (ary_type != nullptr && length_type != nullptr) {
2058 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type);
2059 if (narrow_length_type != length_type) {
2060 // Assert one of:
2061 // - the narrow_length is 0
2062 // - the narrow_length is not wider than length
2063 assert(narrow_length_type == TypeInt::ZERO ||
2064 (length_type->is_con() && narrow_length_type->is_con() &&
2065 (narrow_length_type->_hi <= length_type->_lo)) ||
2066 (narrow_length_type->_hi <= length_type->_hi &&
2067 narrow_length_type->_lo >= length_type->_lo),
2068 "narrow type must be narrower than length type");
2069
2070 // Return null if new nodes are not allowed
2071 if (!allow_new_nodes) {
2072 return nullptr;
2073 }
2074 // Create a cast which is control dependent on the initialization to
2075 // propagate the fact that the array length must be positive.
2076 InitializeNode* init = initialization();
2077 if (init != nullptr) {
2078 length = new CastIINode(init->proj_out_or_null(TypeFunc::Control), length, narrow_length_type);
2079 }
2080 }
2081 }
2082
2083 return length;
2084 }
2085
2086 //=============================================================================
2087 const TypeFunc* LockNode::_lock_type_Type = nullptr;
2088
2089 uint LockNode::size_of() const { return sizeof(*this); }
2090
2091 // Redundant lock elimination
2092 //
2093 // There are various patterns of locking where we release and
2094 // immediately reacquire a lock in a piece of code where no operations
2095 // occur in between that would be observable. In those cases we can
2096 // skip releasing and reacquiring the lock without violating any
2097 // fairness requirements. Doing this around a loop could cause a lock
2098 // to be held for a very long time so we concentrate on non-looping
2099 // control flow. We also require that the operations are fully
2100 // redundant meaning that we don't introduce new lock operations on
2101 // some paths so to be able to eliminate it on others ala PRE. This
2102 // would probably require some more extensive graph manipulation to
2103 // guarantee that the memory edges were all handled correctly.
2104 //
2105 // Assuming p is a simple predicate which can't trap in any way and s
2106 // is a synchronized method consider this code:
2107 //
2108 // s();
2109 // if (p)
2110 // s();
2111 // else
2112 // s();
2113 // s();
2114 //
2115 // 1. The unlocks of the first call to s can be eliminated if the
2116 // locks inside the then and else branches are eliminated.
2117 //
2118 // 2. The unlocks of the then and else branches can be eliminated if
2119 // the lock of the final call to s is eliminated.
2120 //
2121 // Either of these cases subsumes the simple case of sequential control flow
2122 //
2123 // Additionally we can eliminate versions without the else case:
2124 //
2125 // s();
2126 // if (p)
2127 // s();
2128 // s();
2129 //
2130 // 3. In this case we eliminate the unlock of the first s, the lock
2131 // and unlock in the then case and the lock in the final s.
2132 //
2133 // Note also that in all these cases the then/else pieces don't have
2134 // to be trivial as long as they begin and end with synchronization
2135 // operations.
2136 //
2137 // s();
2138 // if (p)
2139 // s();
2140 // f();
2141 // s();
2142 // s();
2143 //
2144 // The code will work properly for this case, leaving in the unlock
2145 // before the call to f and the relock after it.
2146 //
2147 // A potentially interesting case which isn't handled here is when the
2148 // locking is partially redundant.
2149 //
2150 // s();
2151 // if (p)
2152 // s();
2153 //
2154 // This could be eliminated putting unlocking on the else case and
2155 // eliminating the first unlock and the lock in the then side.
2156 // Alternatively the unlock could be moved out of the then side so it
2157 // was after the merge and the first unlock and second lock
2158 // eliminated. This might require less manipulation of the memory
2159 // state to get correct.
2160 //
2161 // Additionally we might allow work between a unlock and lock before
2162 // giving up eliminating the locks. The current code disallows any
2163 // conditional control flow between these operations. A formulation
2164 // similar to partial redundancy elimination computing the
2165 // availability of unlocking and the anticipatability of locking at a
2166 // program point would allow detection of fully redundant locking with
2167 // some amount of work in between. I'm not sure how often I really
2168 // think that would occur though. Most of the cases I've seen
2169 // indicate it's likely non-trivial work would occur in between.
2170 // There may be other more complicated constructs where we could
2171 // eliminate locking but I haven't seen any others appear as hot or
2172 // interesting.
2173 //
2174 // Locking and unlocking have a canonical form in ideal that looks
2175 // roughly like this:
2176 //
2177 // <obj>
2178 // | \\------+
2179 // | \ \
2180 // | BoxLock \
2181 // | | | \
2182 // | | \ \
2183 // | | FastLock
2184 // | | /
2185 // | | /
2186 // | | |
2187 //
2188 // Lock
2189 // |
2190 // Proj #0
2191 // |
2192 // MembarAcquire
2193 // |
2194 // Proj #0
2195 //
2196 // MembarRelease
2197 // |
2198 // Proj #0
2199 // |
2200 // Unlock
2201 // |
2202 // Proj #0
2203 //
2204 //
2205 // This code proceeds by processing Lock nodes during PhaseIterGVN
2206 // and searching back through its control for the proper code
2207 // patterns. Once it finds a set of lock and unlock operations to
2208 // eliminate they are marked as eliminatable which causes the
2209 // expansion of the Lock and Unlock macro nodes to make the operation a NOP
2210 //
2211 //=============================================================================
2212
2213 //
2214 // Utility function to skip over uninteresting control nodes. Nodes skipped are:
2215 // - copy regions. (These may not have been optimized away yet.)
2216 // - eliminated locking nodes
2217 //
2218 static Node *next_control(Node *ctrl) {
2219 if (ctrl == nullptr)
2220 return nullptr;
2221 while (1) {
2222 if (ctrl->is_Region()) {
2223 RegionNode *r = ctrl->as_Region();
2224 Node *n = r->is_copy();
2225 if (n == nullptr)
2226 break; // hit a region, return it
2227 else
2228 ctrl = n;
2229 } else if (ctrl->is_Proj()) {
2230 Node *in0 = ctrl->in(0);
2231 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
2232 ctrl = in0->in(0);
2233 } else {
2234 break;
2235 }
2236 } else {
2237 break; // found an interesting control
2238 }
2239 }
2240 return ctrl;
2241 }
2242 //
2243 // Given a control, see if it's the control projection of an Unlock which
2244 // operating on the same object as lock.
2245 //
2246 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
2247 GrowableArray<AbstractLockNode*> &lock_ops) {
2248 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : nullptr;
2249 if (ctrl_proj != nullptr && ctrl_proj->_con == TypeFunc::Control) {
2250 Node *n = ctrl_proj->in(0);
2251 if (n != nullptr && n->is_Unlock()) {
2252 UnlockNode *unlock = n->as_Unlock();
2253 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2254 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
2255 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
2256 if (lock_obj->eqv_uncast(unlock_obj) &&
2257 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) &&
2258 !unlock->is_eliminated()) {
2259 lock_ops.append(unlock);
2260 return true;
2261 }
2262 }
2263 }
2264 return false;
2265 }
2266
2267 //
2268 // Find the lock matching an unlock. Returns null if a safepoint
2269 // or complicated control is encountered first.
2270 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
2271 LockNode *lock_result = nullptr;
2272 // find the matching lock, or an intervening safepoint
2273 Node *ctrl = next_control(unlock->in(0));
2274 while (1) {
2275 assert(ctrl != nullptr, "invalid control graph");
2276 assert(!ctrl->is_Start(), "missing lock for unlock");
2277 if (ctrl->is_top()) break; // dead control path
2278 if (ctrl->is_Proj()) ctrl = ctrl->in(0);
2279 if (ctrl->is_SafePoint()) {
2280 break; // found a safepoint (may be the lock we are searching for)
2281 } else if (ctrl->is_Region()) {
2282 // Check for a simple diamond pattern. Punt on anything more complicated
2283 if (ctrl->req() == 3 && ctrl->in(1) != nullptr && ctrl->in(2) != nullptr) {
2284 Node *in1 = next_control(ctrl->in(1));
2285 Node *in2 = next_control(ctrl->in(2));
2286 if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
2287 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
2288 ctrl = next_control(in1->in(0)->in(0));
2289 } else {
2290 break;
2291 }
2292 } else {
2293 break;
2294 }
2295 } else {
2296 ctrl = next_control(ctrl->in(0)); // keep searching
2297 }
2298 }
2299 if (ctrl->is_Lock()) {
2300 LockNode *lock = ctrl->as_Lock();
2301 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2302 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
2303 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
2304 if (lock_obj->eqv_uncast(unlock_obj) &&
2305 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) {
2306 lock_result = lock;
2307 }
2308 }
2309 return lock_result;
2310 }
2311
2312 // This code corresponds to case 3 above.
2313
2314 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
2315 GrowableArray<AbstractLockNode*> &lock_ops) {
2316 Node* if_node = node->in(0);
2317 bool if_true = node->is_IfTrue();
2318
2319 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
2320 Node *lock_ctrl = next_control(if_node->in(0));
2321 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
2322 Node* lock1_node = nullptr;
2323 ProjNode* proj = if_node->as_If()->proj_out(!if_true);
2324 if (if_true) {
2325 if (proj->is_IfFalse() && proj->outcnt() == 1) {
2326 lock1_node = proj->unique_out();
2327 }
2328 } else {
2329 if (proj->is_IfTrue() && proj->outcnt() == 1) {
2330 lock1_node = proj->unique_out();
2331 }
2332 }
2333 if (lock1_node != nullptr && lock1_node->is_Lock()) {
2334 LockNode *lock1 = lock1_node->as_Lock();
2335 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2336 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
2337 Node* lock1_obj = bs->step_over_gc_barrier(lock1->obj_node());
2338 if (lock_obj->eqv_uncast(lock1_obj) &&
2339 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) &&
2340 !lock1->is_eliminated()) {
2341 lock_ops.append(lock1);
2342 return true;
2343 }
2344 }
2345 }
2346 }
2347
2348 lock_ops.trunc_to(0);
2349 return false;
2350 }
2351
2352 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
2353 GrowableArray<AbstractLockNode*> &lock_ops) {
2354 // check each control merging at this point for a matching unlock.
2355 // in(0) should be self edge so skip it.
2356 for (int i = 1; i < (int)region->req(); i++) {
2357 Node *in_node = next_control(region->in(i));
2358 if (in_node != nullptr) {
2359 if (find_matching_unlock(in_node, lock, lock_ops)) {
2360 // found a match so keep on checking.
2361 continue;
2362 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
2363 continue;
2364 }
2365
2366 // If we fall through to here then it was some kind of node we
2367 // don't understand or there wasn't a matching unlock, so give
2368 // up trying to merge locks.
2369 lock_ops.trunc_to(0);
2370 return false;
2371 }
2372 }
2373 return true;
2374
2375 }
2376
2377 // Check that all locks/unlocks associated with object come from balanced regions.
2378 bool AbstractLockNode::is_balanced() {
2379 Node* obj = obj_node();
2380 for (uint j = 0; j < obj->outcnt(); j++) {
2381 Node* n = obj->raw_out(j);
2382 if (n->is_AbstractLock() &&
2383 n->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2384 BoxLockNode* n_box = n->as_AbstractLock()->box_node()->as_BoxLock();
2385 if (n_box->is_unbalanced()) {
2386 return false;
2387 }
2388 }
2389 }
2390 return true;
2391 }
2392
2393 const char* AbstractLockNode::_kind_names[] = {"Regular", "NonEscObj", "Coarsened", "Nested"};
2394
2395 const char * AbstractLockNode::kind_as_string() const {
2396 return _kind_names[_kind];
2397 }
2398
2399 #ifndef PRODUCT
2400 //
2401 // Create a counter which counts the number of times this lock is acquired
2402 //
2403 void AbstractLockNode::create_lock_counter(JVMState* state) {
2404 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
2405 }
2406
2407 void AbstractLockNode::set_eliminated_lock_counter() {
2408 if (_counter) {
2409 // Update the counter to indicate that this lock was eliminated.
2410 // The counter update code will stay around even though the
2411 // optimizer will eliminate the lock operation itself.
2412 _counter->set_tag(NamedCounter::EliminatedLockCounter);
2413 }
2414 }
2415
2416 void AbstractLockNode::dump_spec(outputStream* st) const {
2417 st->print("%s ", _kind_names[_kind]);
2418 CallNode::dump_spec(st);
2419 }
2420
2421 void AbstractLockNode::dump_compact_spec(outputStream* st) const {
2422 st->print("%s", _kind_names[_kind]);
2423 }
2424 #endif
2425
2426 //=============================================================================
2427 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2428
2429 // perform any generic optimizations first (returns 'this' or null)
2430 Node *result = SafePointNode::Ideal(phase, can_reshape);
2431 if (result != nullptr) return result;
2432 // Don't bother trying to transform a dead node
2433 if (in(0) && in(0)->is_top()) return nullptr;
2434
2435 // Now see if we can optimize away this lock. We don't actually
2436 // remove the locking here, we simply set the _eliminate flag which
2437 // prevents macro expansion from expanding the lock. Since we don't
2438 // modify the graph, the value returned from this function is the
2439 // one computed above.
2440 const Type* obj_type = phase->type(obj_node());
2441 if (can_reshape && EliminateLocks && !is_non_esc_obj() && !obj_type->is_inlinetypeptr()) {
2442 //
2443 // If we are locking an non-escaped object, the lock/unlock is unnecessary
2444 //
2445 ConnectionGraph *cgr = phase->C->congraph();
2446 if (cgr != nullptr && cgr->can_eliminate_lock(this)) {
2447 assert(!is_eliminated() || is_coarsened(), "sanity");
2448 // The lock could be marked eliminated by lock coarsening
2449 // code during first IGVN before EA. Replace coarsened flag
2450 // to eliminate all associated locks/unlocks.
2451 #ifdef ASSERT
2452 this->log_lock_optimization(phase->C,"eliminate_lock_set_non_esc1");
2453 #endif
2454 this->set_non_esc_obj();
2455 return result;
2456 }
2457
2458 if (!phase->C->do_locks_coarsening()) {
2459 return result; // Compiling without locks coarsening
2460 }
2461 //
2462 // Try lock coarsening
2463 //
2464 PhaseIterGVN* iter = phase->is_IterGVN();
2465 if (iter != nullptr && !is_eliminated()) {
2466
2467 GrowableArray<AbstractLockNode*> lock_ops;
2468
2469 Node *ctrl = next_control(in(0));
2470
2471 // now search back for a matching Unlock
2472 if (find_matching_unlock(ctrl, this, lock_ops)) {
2473 // found an unlock directly preceding this lock. This is the
2474 // case of single unlock directly control dependent on a
2475 // single lock which is the trivial version of case 1 or 2.
2476 } else if (ctrl->is_Region() ) {
2477 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
2478 // found lock preceded by multiple unlocks along all paths
2479 // joining at this point which is case 3 in description above.
2480 }
2481 } else {
2482 // see if this lock comes from either half of an if and the
2483 // predecessors merges unlocks and the other half of the if
2484 // performs a lock.
2485 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
2486 // found unlock splitting to an if with locks on both branches.
2487 }
2488 }
2489
2490 if (lock_ops.length() > 0) {
2491 // add ourselves to the list of locks to be eliminated.
2492 lock_ops.append(this);
2493
2494 #ifndef PRODUCT
2495 if (PrintEliminateLocks) {
2496 int locks = 0;
2497 int unlocks = 0;
2498 if (Verbose) {
2499 tty->print_cr("=== Locks coarsening ===");
2500 tty->print("Obj: ");
2501 obj_node()->dump();
2502 }
2503 for (int i = 0; i < lock_ops.length(); i++) {
2504 AbstractLockNode* lock = lock_ops.at(i);
2505 if (lock->Opcode() == Op_Lock)
2506 locks++;
2507 else
2508 unlocks++;
2509 if (Verbose) {
2510 tty->print("Box %d: ", i);
2511 box_node()->dump();
2512 tty->print(" %d: ", i);
2513 lock->dump();
2514 }
2515 }
2516 tty->print_cr("=== Coarsened %d unlocks and %d locks", unlocks, locks);
2517 }
2518 #endif
2519
2520 // for each of the identified locks, mark them
2521 // as eliminatable
2522 for (int i = 0; i < lock_ops.length(); i++) {
2523 AbstractLockNode* lock = lock_ops.at(i);
2524
2525 // Mark it eliminated by coarsening and update any counters
2526 #ifdef ASSERT
2527 lock->log_lock_optimization(phase->C, "eliminate_lock_set_coarsened");
2528 #endif
2529 lock->set_coarsened();
2530 }
2531 // Record this coarsened group.
2532 phase->C->add_coarsened_locks(lock_ops);
2533 } else if (ctrl->is_Region() &&
2534 iter->_worklist.member(ctrl)) {
2535 // We weren't able to find any opportunities but the region this
2536 // lock is control dependent on hasn't been processed yet so put
2537 // this lock back on the worklist so we can check again once any
2538 // region simplification has occurred.
2539 iter->_worklist.push(this);
2540 }
2541 }
2542 }
2543
2544 return result;
2545 }
2546
2547 //=============================================================================
2548 bool LockNode::is_nested_lock_region() {
2549 return is_nested_lock_region(nullptr);
2550 }
2551
2552 // p is used for access to compilation log; no logging if null
2553 bool LockNode::is_nested_lock_region(Compile * c) {
2554 BoxLockNode* box = box_node()->as_BoxLock();
2555 int stk_slot = box->stack_slot();
2556 if (stk_slot <= 0) {
2557 #ifdef ASSERT
2558 this->log_lock_optimization(c, "eliminate_lock_INLR_1");
2559 #endif
2560 return false; // External lock or it is not Box (Phi node).
2561 }
2562
2563 // Ignore complex cases: merged locks or multiple locks.
2564 Node* obj = obj_node();
2565 LockNode* unique_lock = nullptr;
2566 Node* bad_lock = nullptr;
2567 if (!box->is_simple_lock_region(&unique_lock, obj, &bad_lock)) {
2568 #ifdef ASSERT
2569 this->log_lock_optimization(c, "eliminate_lock_INLR_2a", bad_lock);
2570 #endif
2571 return false;
2572 }
2573 if (unique_lock != this) {
2574 #ifdef ASSERT
2575 this->log_lock_optimization(c, "eliminate_lock_INLR_2b", (unique_lock != nullptr ? unique_lock : bad_lock));
2576 if (PrintEliminateLocks && Verbose) {
2577 tty->print_cr("=============== unique_lock != this ============");
2578 tty->print(" this: ");
2579 this->dump();
2580 tty->print(" box: ");
2581 box->dump();
2582 tty->print(" obj: ");
2583 obj->dump();
2584 if (unique_lock != nullptr) {
2585 tty->print(" unique_lock: ");
2586 unique_lock->dump();
2587 }
2588 if (bad_lock != nullptr) {
2589 tty->print(" bad_lock: ");
2590 bad_lock->dump();
2591 }
2592 tty->print_cr("===============");
2593 }
2594 #endif
2595 return false;
2596 }
2597
2598 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2599 obj = bs->step_over_gc_barrier(obj);
2600 // Look for external lock for the same object.
2601 SafePointNode* sfn = this->as_SafePoint();
2602 JVMState* youngest_jvms = sfn->jvms();
2603 int max_depth = youngest_jvms->depth();
2604 for (int depth = 1; depth <= max_depth; depth++) {
2605 JVMState* jvms = youngest_jvms->of_depth(depth);
2606 int num_mon = jvms->nof_monitors();
2607 // Loop over monitors
2608 for (int idx = 0; idx < num_mon; idx++) {
2609 Node* obj_node = sfn->monitor_obj(jvms, idx);
2610 obj_node = bs->step_over_gc_barrier(obj_node);
2611 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock();
2612 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) {
2613 box->set_nested();
2614 return true;
2615 }
2616 }
2617 }
2618 #ifdef ASSERT
2619 this->log_lock_optimization(c, "eliminate_lock_INLR_3");
2620 #endif
2621 return false;
2622 }
2623
2624 //=============================================================================
2625 uint UnlockNode::size_of() const { return sizeof(*this); }
2626
2627 //=============================================================================
2628 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2629
2630 // perform any generic optimizations first (returns 'this' or null)
2631 Node *result = SafePointNode::Ideal(phase, can_reshape);
2632 if (result != nullptr) return result;
2633 // Don't bother trying to transform a dead node
2634 if (in(0) && in(0)->is_top()) return nullptr;
2635
2636 // Now see if we can optimize away this unlock. We don't actually
2637 // remove the unlocking here, we simply set the _eliminate flag which
2638 // prevents macro expansion from expanding the unlock. Since we don't
2639 // modify the graph, the value returned from this function is the
2640 // one computed above.
2641 // Escape state is defined after Parse phase.
2642 const Type* obj_type = phase->type(obj_node());
2643 if (can_reshape && EliminateLocks && !is_non_esc_obj() && !obj_type->is_inlinetypeptr()) {
2644 //
2645 // If we are unlocking an non-escaped object, the lock/unlock is unnecessary.
2646 //
2647 ConnectionGraph *cgr = phase->C->congraph();
2648 if (cgr != nullptr && cgr->can_eliminate_lock(this)) {
2649 assert(!is_eliminated() || is_coarsened(), "sanity");
2650 // The lock could be marked eliminated by lock coarsening
2651 // code during first IGVN before EA. Replace coarsened flag
2652 // to eliminate all associated locks/unlocks.
2653 #ifdef ASSERT
2654 this->log_lock_optimization(phase->C, "eliminate_lock_set_non_esc2");
2655 #endif
2656 this->set_non_esc_obj();
2657 }
2658 }
2659 return result;
2660 }
2661
2662 void AbstractLockNode::log_lock_optimization(Compile *C, const char * tag, Node* bad_lock) const {
2663 if (C == nullptr) {
2664 return;
2665 }
2666 CompileLog* log = C->log();
2667 if (log != nullptr) {
2668 Node* box = box_node();
2669 Node* obj = obj_node();
2670 int box_id = box != nullptr ? box->_idx : -1;
2671 int obj_id = obj != nullptr ? obj->_idx : -1;
2672
2673 log->begin_head("%s compile_id='%d' lock_id='%d' class='%s' kind='%s' box_id='%d' obj_id='%d' bad_id='%d'",
2674 tag, C->compile_id(), this->_idx,
2675 is_Unlock() ? "unlock" : is_Lock() ? "lock" : "?",
2676 kind_as_string(), box_id, obj_id, (bad_lock != nullptr ? bad_lock->_idx : -1));
2677 log->stamp();
2678 log->end_head();
2679 JVMState* p = is_Unlock() ? (as_Unlock()->dbg_jvms()) : jvms();
2680 while (p != nullptr) {
2681 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
2682 p = p->caller();
2683 }
2684 log->tail(tag);
2685 }
2686 }
2687
2688 bool CallNode::may_modify_arraycopy_helper(const TypeOopPtr* dest_t, const TypeOopPtr* t_oop, PhaseValues* phase) {
2689 if (dest_t->is_known_instance() && t_oop->is_known_instance()) {
2690 return dest_t->instance_id() == t_oop->instance_id();
2691 }
2692
2693 if (dest_t->isa_instptr() && !dest_t->is_instptr()->instance_klass()->equals(phase->C->env()->Object_klass())) {
2694 // clone
2695 if (t_oop->isa_aryptr()) {
2696 return false;
2697 }
2698 if (!t_oop->isa_instptr()) {
2699 return true;
2700 }
2701 if (dest_t->maybe_java_subtype_of(t_oop) || t_oop->maybe_java_subtype_of(dest_t)) {
2702 return true;
2703 }
2704 // unrelated
2705 return false;
2706 }
2707
2708 if (dest_t->isa_aryptr()) {
2709 // arraycopy or array clone
2710 if (t_oop->isa_instptr()) {
2711 return false;
2712 }
2713 if (!t_oop->isa_aryptr()) {
2714 return true;
2715 }
2716
2717 const Type* elem = dest_t->is_aryptr()->elem();
2718 if (elem == Type::BOTTOM) {
2719 // An array but we don't know what elements are
2720 return true;
2721 }
2722
2723 dest_t = dest_t->is_aryptr()->with_field_offset(Type::OffsetBot)->add_offset(Type::OffsetBot)->is_oopptr();
2724 t_oop = t_oop->is_aryptr()->with_field_offset(Type::OffsetBot);
2725 uint dest_alias = phase->C->get_alias_index(dest_t);
2726 uint t_oop_alias = phase->C->get_alias_index(t_oop);
2727
2728 return dest_alias == t_oop_alias;
2729 }
2730
2731 return true;
2732 }
2733
2734 PowDNode::PowDNode(Compile* C, Node* base, Node* exp)
2735 : CallLeafPureNode(
2736 OptoRuntime::Math_DD_D_Type(),
2737 StubRoutines::dpow() != nullptr ? StubRoutines::dpow() : CAST_FROM_FN_PTR(address, SharedRuntime::dpow),
2738 "pow") {
2739 add_flag(Flag_is_macro);
2740 C->add_macro_node(this);
2741
2742 init_req(TypeFunc::Parms + 0, base);
2743 init_req(TypeFunc::Parms + 1, C->top()); // double slot padding
2744 init_req(TypeFunc::Parms + 2, exp);
2745 init_req(TypeFunc::Parms + 3, C->top()); // double slot padding
2746 }
2747
2748 const Type* PowDNode::Value(PhaseGVN* phase) const {
2749 const Type* t_base = phase->type(base());
2750 const Type* t_exp = phase->type(exp());
2751
2752 if (t_base == Type::TOP || t_exp == Type::TOP) {
2753 return Type::TOP;
2754 }
2755
2756 const TypeD* base_con = t_base->isa_double_constant();
2757 const TypeD* exp_con = t_exp->isa_double_constant();
2758 const TypeD* result_t = nullptr;
2759
2760 // constant folding: both inputs are constants
2761 if (base_con != nullptr && exp_con != nullptr) {
2762 result_t = TypeD::make(SharedRuntime::dpow(base_con->getd(), exp_con->getd()));
2763 }
2764
2765 // Special cases when only the exponent is known:
2766 if (exp_con != nullptr) {
2767 double e = exp_con->getd();
2768
2769 // If the second argument is positive or negative zero, then the result is 1.0.
2770 // i.e., pow(x, +/-0.0D) => 1.0
2771 if (e == 0.0) { // true for both -0.0 and +0.0
2772 result_t = TypeD::ONE;
2773 }
2774
2775 // If the second argument is NaN, then the result is NaN.
2776 // i.e., pow(x, NaN) => NaN
2777 if (g_isnan(e)) {
2778 result_t = TypeD::make(NAN);
2779 }
2780 }
2781
2782 if (result_t != nullptr) {
2783 // We can't simply return a TypeD here, it must be a tuple type to be compatible with call nodes.
2784 const Type** fields = TypeTuple::fields(2);
2785 fields[TypeFunc::Parms + 0] = result_t;
2786 fields[TypeFunc::Parms + 1] = Type::HALF;
2787 return TypeTuple::make(TypeFunc::Parms + 2, fields);
2788 }
2789
2790 return tf()->range_cc();
2791 }
2792
2793 Node* PowDNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2794 if (!can_reshape) {
2795 return nullptr; // wait for igvn
2796 }
2797
2798 PhaseIterGVN* igvn = phase->is_IterGVN();
2799 Node* base = this->base();
2800 Node* exp = this->exp();
2801
2802 const Type* t_exp = phase->type(exp);
2803 const TypeD* exp_con = t_exp->isa_double_constant();
2804
2805 // Special cases when only the exponent is known:
2806 if (exp_con != nullptr) {
2807 double e = exp_con->getd();
2808
2809 // If the second argument is 1.0, then the result is the same as the first argument.
2810 // i.e., pow(x, 1.0) => x
2811 if (e == 1.0) {
2812 return make_tuple_of_input_state_and_result(igvn, base);
2813 }
2814
2815 // If the second argument is 2.0, then strength reduce to multiplications.
2816 // i.e., pow(x, 2.0) => x * x
2817 if (e == 2.0) {
2818 Node* mul = igvn->transform(new MulDNode(base, base));
2819 return make_tuple_of_input_state_and_result(igvn, mul);
2820 }
2821
2822 // If the second argument is 0.5, the strength reduce to square roots.
2823 // i.e., pow(x, 0.5) => sqrt(x) iff x > 0
2824 if (e == 0.5 && Matcher::match_rule_supported(Op_SqrtD)) {
2825 Node* ctrl = in(TypeFunc::Control);
2826 Node* zero = igvn->zerocon(T_DOUBLE);
2827
2828 // According to the API specs, pow(-0.0, 0.5) = 0.0 and sqrt(-0.0) = -0.0.
2829 // So pow(-0.0, 0.5) shouldn't be replaced with sqrt(-0.0).
2830 // -0.0/+0.0 are both excluded since floating-point comparison doesn't distinguish -0.0 from +0.0.
2831 Node* cmp = igvn->register_new_node_with_optimizer(new CmpDNode(base, zero));
2832 Node* test = igvn->register_new_node_with_optimizer(new BoolNode(cmp, BoolTest::le));
2833
2834 IfNode* iff = new IfNode(ctrl, test, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2835 igvn->register_new_node_with_optimizer(iff);
2836 Node* if_slow = igvn->register_new_node_with_optimizer(new IfTrueNode(iff)); // x <= 0
2837 Node* if_fast = igvn->register_new_node_with_optimizer(new IfFalseNode(iff)); // x > 0
2838
2839 // slow path: call pow(x, 0.5)
2840 Node* call = igvn->register_new_node_with_optimizer(inline_call_leaf_pure_node(if_slow));
2841 Node* call_ctrl = igvn->register_new_node_with_optimizer(new ProjNode(call, TypeFunc::Control));
2842 Node* call_result = igvn->register_new_node_with_optimizer(new ProjNode(call, TypeFunc::Parms + 0));
2843
2844 // fast path: sqrt(x)
2845 Node* sqrt = igvn->register_new_node_with_optimizer(new SqrtDNode(igvn->C, if_fast, base));
2846
2847 // merge paths
2848 RegionNode* region = new RegionNode(3);
2849 igvn->register_new_node_with_optimizer(region);
2850 region->init_req(1, call_ctrl); // slow path
2851 region->init_req(2, if_fast); // fast path
2852
2853 PhiNode* phi = new PhiNode(region, Type::DOUBLE);
2854 igvn->register_new_node_with_optimizer(phi);
2855 phi->init_req(1, call_result); // slow: pow() result
2856 phi->init_req(2, sqrt); // fast: sqrt() result
2857
2858 igvn->C->set_has_split_ifs(true); // Has chance for split-if optimization
2859
2860 return make_tuple_of_input_state_and_result(igvn, phi, region);
2861 }
2862 }
2863
2864 return CallLeafPureNode::Ideal(phase, can_reshape);
2865 }
2866
2867 // We can't simply have Ideal() returning a Con or MulNode since the users are still expecting a Call node, but we could
2868 // produce a tuple that follows the same pattern so users can still get control, io, memory, etc..
2869 TupleNode* PowDNode::make_tuple_of_input_state_and_result(PhaseIterGVN* phase, Node* result, Node* control) {
2870 if (control == nullptr) {
2871 control = in(TypeFunc::Control);
2872 }
2873
2874 Compile* C = phase->C;
2875 C->remove_macro_node(this);
2876 TupleNode* tuple = TupleNode::make(
2877 tf()->range_cc(),
2878 control,
2879 in(TypeFunc::I_O),
2880 in(TypeFunc::Memory),
2881 in(TypeFunc::FramePtr),
2882 in(TypeFunc::ReturnAdr),
2883 result,
2884 C->top());
2885 return tuple;
2886 }