1 /*
2 * Copyright (c) 1997, 2022, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "ci/bcEscapeAnalyzer.hpp"
28 #include "compiler/oopMap.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/c2/barrierSetC2.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/escape.hpp"
37 #include "opto/locknode.hpp"
38 #include "opto/machnode.hpp"
39 #include "opto/matcher.hpp"
40 #include "opto/parse.hpp"
41 #include "opto/regalloc.hpp"
42 #include "opto/regmask.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "runtime/sharedRuntime.hpp"
46 #include "utilities/powerOfTwo.hpp"
47 #include "code/vmreg.hpp"
48
49 // Portions of code courtesy of Clifford Click
50
51 // Optimization - Graph Style
52
53 //=============================================================================
54 uint StartNode::size_of() const { return sizeof(*this); }
55 bool StartNode::cmp( const Node &n ) const
56 { return _domain == ((StartNode&)n)._domain; }
57 const Type *StartNode::bottom_type() const { return _domain; }
58 const Type* StartNode::Value(PhaseGVN* phase) const { return _domain; }
59 #ifndef PRODUCT
60 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);}
61 void StartNode::dump_compact_spec(outputStream *st) const { /* empty */ }
62 #endif
63
64 //------------------------------Ideal------------------------------------------
65 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){
66 return remove_dead_region(phase, can_reshape) ? this : NULL;
67 }
68
69 //------------------------------calling_convention-----------------------------
70 void StartNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
71 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt);
72 }
73
74 //------------------------------Registers--------------------------------------
75 const RegMask &StartNode::in_RegMask(uint) const {
76 return RegMask::Empty;
77 }
78
79 //------------------------------match------------------------------------------
80 // Construct projections for incoming parameters, and their RegMask info
81 Node *StartNode::match( const ProjNode *proj, const Matcher *match ) {
82 switch (proj->_con) {
83 case TypeFunc::Control:
84 case TypeFunc::I_O:
85 case TypeFunc::Memory:
86 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
87 case TypeFunc::FramePtr:
88 return new MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP);
89 case TypeFunc::ReturnAdr:
90 return new MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP);
91 case TypeFunc::Parms:
92 default: {
93 uint parm_num = proj->_con - TypeFunc::Parms;
94 const Type *t = _domain->field_at(proj->_con);
95 if (t->base() == Type::Half) // 2nd half of Longs and Doubles
96 return new ConNode(Type::TOP);
97 uint ideal_reg = t->ideal_reg();
98 RegMask &rm = match->_calling_convention_mask[parm_num];
99 return new MachProjNode(this,proj->_con,rm,ideal_reg);
100 }
101 }
102 return NULL;
103 }
104
105 //------------------------------StartOSRNode----------------------------------
106 // The method start node for an on stack replacement adapter
107
108 //------------------------------osr_domain-----------------------------
109 const TypeTuple *StartOSRNode::osr_domain() {
110 const Type **fields = TypeTuple::fields(2);
111 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer
112
113 return TypeTuple::make(TypeFunc::Parms+1, fields);
114 }
115
116 //=============================================================================
117 const char * const ParmNode::names[TypeFunc::Parms+1] = {
118 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms"
119 };
120
121 #ifndef PRODUCT
122 void ParmNode::dump_spec(outputStream *st) const {
123 if( _con < TypeFunc::Parms ) {
124 st->print("%s", names[_con]);
125 } else {
126 st->print("Parm%d: ",_con-TypeFunc::Parms);
127 // Verbose and WizardMode dump bottom_type for all nodes
128 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st);
129 }
130 }
131
132 void ParmNode::dump_compact_spec(outputStream *st) const {
133 if (_con < TypeFunc::Parms) {
134 st->print("%s", names[_con]);
135 } else {
136 st->print("%d:", _con-TypeFunc::Parms);
137 // unconditionally dump bottom_type
138 bottom_type()->dump_on(st);
139 }
140 }
141
142 // For a ParmNode, all immediate inputs and outputs are considered relevant
143 // both in compact and standard representation.
144 void ParmNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
145 this->collect_nodes(in_rel, 1, false, false);
146 this->collect_nodes(out_rel, -1, false, false);
147 }
148 #endif
149
150 uint ParmNode::ideal_reg() const {
151 switch( _con ) {
152 case TypeFunc::Control : // fall through
153 case TypeFunc::I_O : // fall through
154 case TypeFunc::Memory : return 0;
155 case TypeFunc::FramePtr : // fall through
156 case TypeFunc::ReturnAdr: return Op_RegP;
157 default : assert( _con > TypeFunc::Parms, "" );
158 // fall through
159 case TypeFunc::Parms : {
160 // Type of argument being passed
161 const Type *t = in(0)->as_Start()->_domain->field_at(_con);
162 return t->ideal_reg();
163 }
164 }
165 ShouldNotReachHere();
166 return 0;
167 }
168
169 //=============================================================================
170 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) {
171 init_req(TypeFunc::Control,cntrl);
172 init_req(TypeFunc::I_O,i_o);
173 init_req(TypeFunc::Memory,memory);
174 init_req(TypeFunc::FramePtr,frameptr);
175 init_req(TypeFunc::ReturnAdr,retadr);
176 }
177
178 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){
179 return remove_dead_region(phase, can_reshape) ? this : NULL;
180 }
181
182 const Type* ReturnNode::Value(PhaseGVN* phase) const {
183 return ( phase->type(in(TypeFunc::Control)) == Type::TOP)
184 ? Type::TOP
185 : Type::BOTTOM;
186 }
187
188 // Do we Match on this edge index or not? No edges on return nodes
189 uint ReturnNode::match_edge(uint idx) const {
190 return 0;
191 }
192
193
194 #ifndef PRODUCT
195 void ReturnNode::dump_req(outputStream *st, DumpConfig* dc) const {
196 // Dump the required inputs, after printing "returns"
197 uint i; // Exit value of loop
198 for (i = 0; i < req(); i++) { // For all required inputs
199 if (i == TypeFunc::Parms) st->print("returns ");
200 Node* p = in(i);
201 if (p != nullptr) {
202 p->dump_idx(false, st, dc);
203 st->print(" ");
204 } else {
205 st->print("_ ");
206 }
207 }
208 }
209 #endif
210
211 //=============================================================================
212 RethrowNode::RethrowNode(
213 Node* cntrl,
214 Node* i_o,
215 Node* memory,
216 Node* frameptr,
217 Node* ret_adr,
218 Node* exception
219 ) : Node(TypeFunc::Parms + 1) {
220 init_req(TypeFunc::Control , cntrl );
221 init_req(TypeFunc::I_O , i_o );
222 init_req(TypeFunc::Memory , memory );
223 init_req(TypeFunc::FramePtr , frameptr );
224 init_req(TypeFunc::ReturnAdr, ret_adr);
225 init_req(TypeFunc::Parms , exception);
226 }
227
228 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){
229 return remove_dead_region(phase, can_reshape) ? this : NULL;
230 }
231
232 const Type* RethrowNode::Value(PhaseGVN* phase) const {
233 return (phase->type(in(TypeFunc::Control)) == Type::TOP)
234 ? Type::TOP
235 : Type::BOTTOM;
236 }
237
238 uint RethrowNode::match_edge(uint idx) const {
239 return 0;
240 }
241
242 #ifndef PRODUCT
243 void RethrowNode::dump_req(outputStream *st, DumpConfig* dc) const {
244 // Dump the required inputs, after printing "exception"
245 uint i; // Exit value of loop
246 for (i = 0; i < req(); i++) { // For all required inputs
247 if (i == TypeFunc::Parms) st->print("exception ");
248 Node* p = in(i);
249 if (p != nullptr) {
250 p->dump_idx(false, st, dc);
251 st->print(" ");
252 } else {
253 st->print("_ ");
254 }
255 }
256 }
257 #endif
258
259 //=============================================================================
260 // Do we Match on this edge index or not? Match only target address & method
261 uint TailCallNode::match_edge(uint idx) const {
262 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
263 }
264
265 //=============================================================================
266 // Do we Match on this edge index or not? Match only target address & oop
267 uint TailJumpNode::match_edge(uint idx) const {
268 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
269 }
270
271 //=============================================================================
272 JVMState::JVMState(ciMethod* method, JVMState* caller) :
273 _method(method) {
274 assert(method != NULL, "must be valid call site");
275 _bci = InvocationEntryBci;
276 _reexecute = Reexecute_Undefined;
277 debug_only(_bci = -99); // random garbage value
278 debug_only(_map = (SafePointNode*)-1);
279 _caller = caller;
280 _depth = 1 + (caller == NULL ? 0 : caller->depth());
281 _locoff = TypeFunc::Parms;
282 _stkoff = _locoff + _method->max_locals();
283 _monoff = _stkoff + _method->max_stack();
284 _scloff = _monoff;
285 _endoff = _monoff;
286 _sp = 0;
287 }
288 JVMState::JVMState(int stack_size) :
289 _method(NULL) {
290 _bci = InvocationEntryBci;
291 _reexecute = Reexecute_Undefined;
292 debug_only(_map = (SafePointNode*)-1);
293 _caller = NULL;
294 _depth = 1;
295 _locoff = TypeFunc::Parms;
296 _stkoff = _locoff;
297 _monoff = _stkoff + stack_size;
298 _scloff = _monoff;
299 _endoff = _monoff;
300 _sp = 0;
301 }
302
303 //--------------------------------of_depth-------------------------------------
304 JVMState* JVMState::of_depth(int d) const {
305 const JVMState* jvmp = this;
306 assert(0 < d && (uint)d <= depth(), "oob");
307 for (int skip = depth() - d; skip > 0; skip--) {
308 jvmp = jvmp->caller();
309 }
310 assert(jvmp->depth() == (uint)d, "found the right one");
311 return (JVMState*)jvmp;
312 }
313
314 //-----------------------------same_calls_as-----------------------------------
315 bool JVMState::same_calls_as(const JVMState* that) const {
316 if (this == that) return true;
317 if (this->depth() != that->depth()) return false;
318 const JVMState* p = this;
319 const JVMState* q = that;
320 for (;;) {
321 if (p->_method != q->_method) return false;
322 if (p->_method == NULL) return true; // bci is irrelevant
323 if (p->_bci != q->_bci) return false;
324 if (p->_reexecute != q->_reexecute) return false;
325 p = p->caller();
326 q = q->caller();
327 if (p == q) return true;
328 assert(p != NULL && q != NULL, "depth check ensures we don't run off end");
329 }
330 }
331
332 //------------------------------debug_start------------------------------------
333 uint JVMState::debug_start() const {
334 debug_only(JVMState* jvmroot = of_depth(1));
335 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
336 return of_depth(1)->locoff();
337 }
338
339 //-------------------------------debug_end-------------------------------------
340 uint JVMState::debug_end() const {
341 debug_only(JVMState* jvmroot = of_depth(1));
342 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
343 return endoff();
344 }
345
346 //------------------------------debug_depth------------------------------------
347 uint JVMState::debug_depth() const {
348 uint total = 0;
349 for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) {
350 total += jvmp->debug_size();
351 }
352 return total;
353 }
354
355 #ifndef PRODUCT
356
357 //------------------------------format_helper----------------------------------
358 // Given an allocation (a Chaitin object) and a Node decide if the Node carries
359 // any defined value or not. If it does, print out the register or constant.
360 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
361 if (n == NULL) { st->print(" NULL"); return; }
362 if (n->is_SafePointScalarObject()) {
363 // Scalar replacement.
364 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
365 scobjs->append_if_missing(spobj);
366 int sco_n = scobjs->find(spobj);
367 assert(sco_n >= 0, "");
368 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
369 return;
370 }
371 if (regalloc->node_regs_max_index() > 0 &&
372 OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
373 char buf[50];
374 regalloc->dump_register(n,buf);
375 st->print(" %s%d]=%s",msg,i,buf);
376 } else { // No register, but might be constant
377 const Type *t = n->bottom_type();
378 switch (t->base()) {
379 case Type::Int:
380 st->print(" %s%d]=#" INT32_FORMAT,msg,i,t->is_int()->get_con());
381 break;
382 case Type::AnyPtr:
383 assert( t == TypePtr::NULL_PTR || n->in_dump(), "" );
384 st->print(" %s%d]=#NULL",msg,i);
385 break;
386 case Type::AryPtr:
387 case Type::InstPtr:
388 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->isa_oopptr()->const_oop()));
389 break;
390 case Type::KlassPtr:
391 case Type::AryKlassPtr:
392 case Type::InstKlassPtr:
393 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_klassptr()->exact_klass()));
394 break;
395 case Type::MetadataPtr:
396 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_metadataptr()->metadata()));
397 break;
398 case Type::NarrowOop:
399 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_oopptr()->const_oop()));
400 break;
401 case Type::RawPtr:
402 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,p2i(t->is_rawptr()));
403 break;
404 case Type::DoubleCon:
405 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
406 break;
407 case Type::FloatCon:
408 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
409 break;
410 case Type::Long:
411 st->print(" %s%d]=#" INT64_FORMAT,msg,i,(int64_t)(t->is_long()->get_con()));
412 break;
413 case Type::Half:
414 case Type::Top:
415 st->print(" %s%d]=_",msg,i);
416 break;
417 default: ShouldNotReachHere();
418 }
419 }
420 }
421
422 //---------------------print_method_with_lineno--------------------------------
423 void JVMState::print_method_with_lineno(outputStream* st, bool show_name) const {
424 if (show_name) _method->print_short_name(st);
425
426 int lineno = _method->line_number_from_bci(_bci);
427 if (lineno != -1) {
428 st->print(" @ bci:%d (line %d)", _bci, lineno);
429 } else {
430 st->print(" @ bci:%d", _bci);
431 }
432 }
433
434 //------------------------------format-----------------------------------------
435 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
436 st->print(" #");
437 if (_method) {
438 print_method_with_lineno(st, true);
439 } else {
440 st->print_cr(" runtime stub ");
441 return;
442 }
443 if (n->is_MachSafePoint()) {
444 GrowableArray<SafePointScalarObjectNode*> scobjs;
445 MachSafePointNode *mcall = n->as_MachSafePoint();
446 uint i;
447 // Print locals
448 for (i = 0; i < (uint)loc_size(); i++)
449 format_helper(regalloc, st, mcall->local(this, i), "L[", i, &scobjs);
450 // Print stack
451 for (i = 0; i < (uint)stk_size(); i++) {
452 if ((uint)(_stkoff + i) >= mcall->len())
453 st->print(" oob ");
454 else
455 format_helper(regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs);
456 }
457 for (i = 0; (int)i < nof_monitors(); i++) {
458 Node *box = mcall->monitor_box(this, i);
459 Node *obj = mcall->monitor_obj(this, i);
460 if (regalloc->node_regs_max_index() > 0 &&
461 OptoReg::is_valid(regalloc->get_reg_first(box))) {
462 box = BoxLockNode::box_node(box);
463 format_helper(regalloc, st, box, "MON-BOX[", i, &scobjs);
464 } else {
465 OptoReg::Name box_reg = BoxLockNode::reg(box);
466 st->print(" MON-BOX%d=%s+%d",
467 i,
468 OptoReg::regname(OptoReg::c_frame_pointer),
469 regalloc->reg2offset(box_reg));
470 }
471 const char* obj_msg = "MON-OBJ[";
472 if (EliminateLocks) {
473 if (BoxLockNode::box_node(box)->is_eliminated())
474 obj_msg = "MON-OBJ(LOCK ELIMINATED)[";
475 }
476 format_helper(regalloc, st, obj, obj_msg, i, &scobjs);
477 }
478
479 for (i = 0; i < (uint)scobjs.length(); i++) {
480 // Scalar replaced objects.
481 st->cr();
482 st->print(" # ScObj" INT32_FORMAT " ", i);
483 SafePointScalarObjectNode* spobj = scobjs.at(i);
484 ciKlass* cik = spobj->bottom_type()->is_oopptr()->exact_klass();
485 assert(cik->is_instance_klass() ||
486 cik->is_array_klass(), "Not supported allocation.");
487 ciInstanceKlass *iklass = NULL;
488 if (cik->is_instance_klass()) {
489 cik->print_name_on(st);
490 iklass = cik->as_instance_klass();
491 } else if (cik->is_type_array_klass()) {
492 cik->as_array_klass()->base_element_type()->print_name_on(st);
493 st->print("[%d]", spobj->n_fields());
494 } else if (cik->is_obj_array_klass()) {
495 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass();
496 if (cie->is_instance_klass()) {
497 cie->print_name_on(st);
498 } else if (cie->is_type_array_klass()) {
499 cie->as_array_klass()->base_element_type()->print_name_on(st);
500 } else {
501 ShouldNotReachHere();
502 }
503 st->print("[%d]", spobj->n_fields());
504 int ndim = cik->as_array_klass()->dimension() - 1;
505 while (ndim-- > 0) {
506 st->print("[]");
507 }
508 }
509 st->print("={");
510 uint nf = spobj->n_fields();
511 if (nf > 0) {
512 uint first_ind = spobj->first_index(mcall->jvms());
513 Node* fld_node = mcall->in(first_ind);
514 ciField* cifield;
515 if (iklass != NULL) {
516 st->print(" [");
517 cifield = iklass->nonstatic_field_at(0);
518 cifield->print_name_on(st);
519 format_helper(regalloc, st, fld_node, ":", 0, &scobjs);
520 } else {
521 format_helper(regalloc, st, fld_node, "[", 0, &scobjs);
522 }
523 for (uint j = 1; j < nf; j++) {
524 fld_node = mcall->in(first_ind+j);
525 if (iklass != NULL) {
526 st->print(", [");
527 cifield = iklass->nonstatic_field_at(j);
528 cifield->print_name_on(st);
529 format_helper(regalloc, st, fld_node, ":", j, &scobjs);
530 } else {
531 format_helper(regalloc, st, fld_node, ", [", j, &scobjs);
532 }
533 }
534 }
535 st->print(" }");
536 }
537 }
538 st->cr();
539 if (caller() != NULL) caller()->format(regalloc, n, st);
540 }
541
542
543 void JVMState::dump_spec(outputStream *st) const {
544 if (_method != NULL) {
545 bool printed = false;
546 if (!Verbose) {
547 // The JVMS dumps make really, really long lines.
548 // Take out the most boring parts, which are the package prefixes.
549 char buf[500];
550 stringStream namest(buf, sizeof(buf));
551 _method->print_short_name(&namest);
552 if (namest.count() < sizeof(buf)) {
553 const char* name = namest.base();
554 if (name[0] == ' ') ++name;
555 const char* endcn = strchr(name, ':'); // end of class name
556 if (endcn == NULL) endcn = strchr(name, '(');
557 if (endcn == NULL) endcn = name + strlen(name);
558 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
559 --endcn;
560 st->print(" %s", endcn);
561 printed = true;
562 }
563 }
564 print_method_with_lineno(st, !printed);
565 if(_reexecute == Reexecute_True)
566 st->print(" reexecute");
567 } else {
568 st->print(" runtime stub");
569 }
570 if (caller() != NULL) caller()->dump_spec(st);
571 }
572
573
574 void JVMState::dump_on(outputStream* st) const {
575 bool print_map = _map && !((uintptr_t)_map & 1) &&
576 ((caller() == NULL) || (caller()->map() != _map));
577 if (print_map) {
578 if (_map->len() > _map->req()) { // _map->has_exceptions()
579 Node* ex = _map->in(_map->req()); // _map->next_exception()
580 // skip the first one; it's already being printed
581 while (ex != NULL && ex->len() > ex->req()) {
582 ex = ex->in(ex->req()); // ex->next_exception()
583 ex->dump(1);
584 }
585 }
586 _map->dump(Verbose ? 2 : 1);
587 }
588 if (caller() != NULL) {
589 caller()->dump_on(st);
590 }
591 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=",
592 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false");
593 if (_method == NULL) {
594 st->print_cr("(none)");
595 } else {
596 _method->print_name(st);
597 st->cr();
598 if (bci() >= 0 && bci() < _method->code_size()) {
599 st->print(" bc: ");
600 _method->print_codes_on(bci(), bci()+1, st);
601 }
602 }
603 }
604
605 // Extra way to dump a jvms from the debugger,
606 // to avoid a bug with C++ member function calls.
607 void dump_jvms(JVMState* jvms) {
608 jvms->dump();
609 }
610 #endif
611
612 //--------------------------clone_shallow--------------------------------------
613 JVMState* JVMState::clone_shallow(Compile* C) const {
614 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
615 n->set_bci(_bci);
616 n->_reexecute = _reexecute;
617 n->set_locoff(_locoff);
618 n->set_stkoff(_stkoff);
619 n->set_monoff(_monoff);
620 n->set_scloff(_scloff);
621 n->set_endoff(_endoff);
622 n->set_sp(_sp);
623 n->set_map(_map);
624 return n;
625 }
626
627 //---------------------------clone_deep----------------------------------------
628 JVMState* JVMState::clone_deep(Compile* C) const {
629 JVMState* n = clone_shallow(C);
630 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) {
631 p->_caller = p->_caller->clone_shallow(C);
632 }
633 assert(n->depth() == depth(), "sanity");
634 assert(n->debug_depth() == debug_depth(), "sanity");
635 return n;
636 }
637
638 /**
639 * Reset map for all callers
640 */
641 void JVMState::set_map_deep(SafePointNode* map) {
642 for (JVMState* p = this; p != NULL; p = p->_caller) {
643 p->set_map(map);
644 }
645 }
646
647 // unlike set_map(), this is two-way setting.
648 void JVMState::bind_map(SafePointNode* map) {
649 set_map(map);
650 _map->set_jvms(this);
651 }
652
653 // Adapt offsets in in-array after adding or removing an edge.
654 // Prerequisite is that the JVMState is used by only one node.
655 void JVMState::adapt_position(int delta) {
656 for (JVMState* jvms = this; jvms != NULL; jvms = jvms->caller()) {
657 jvms->set_locoff(jvms->locoff() + delta);
658 jvms->set_stkoff(jvms->stkoff() + delta);
659 jvms->set_monoff(jvms->monoff() + delta);
660 jvms->set_scloff(jvms->scloff() + delta);
661 jvms->set_endoff(jvms->endoff() + delta);
662 }
663 }
664
665 // Mirror the stack size calculation in the deopt code
666 // How much stack space would we need at this point in the program in
667 // case of deoptimization?
668 int JVMState::interpreter_frame_size() const {
669 const JVMState* jvms = this;
670 int size = 0;
671 int callee_parameters = 0;
672 int callee_locals = 0;
673 int extra_args = method()->max_stack() - stk_size();
674
675 while (jvms != NULL) {
676 int locks = jvms->nof_monitors();
677 int temps = jvms->stk_size();
678 bool is_top_frame = (jvms == this);
679 ciMethod* method = jvms->method();
680
681 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(),
682 temps + callee_parameters,
683 extra_args,
684 locks,
685 callee_parameters,
686 callee_locals,
687 is_top_frame);
688 size += frame_size;
689
690 callee_parameters = method->size_of_parameters();
691 callee_locals = method->max_locals();
692 extra_args = 0;
693 jvms = jvms->caller();
694 }
695 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord;
696 }
697
698 //=============================================================================
699 bool CallNode::cmp( const Node &n ) const
700 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
701 #ifndef PRODUCT
702 void CallNode::dump_req(outputStream *st, DumpConfig* dc) const {
703 // Dump the required inputs, enclosed in '(' and ')'
704 uint i; // Exit value of loop
705 for (i = 0; i < req(); i++) { // For all required inputs
706 if (i == TypeFunc::Parms) st->print("(");
707 Node* p = in(i);
708 if (p != nullptr) {
709 p->dump_idx(false, st, dc);
710 st->print(" ");
711 } else {
712 st->print("_ ");
713 }
714 }
715 st->print(")");
716 }
717
718 void CallNode::dump_spec(outputStream *st) const {
719 st->print(" ");
720 if (tf() != NULL) tf()->dump_on(st);
721 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt);
722 if (jvms() != NULL) jvms()->dump_spec(st);
723 }
724 #endif
725
726 const Type *CallNode::bottom_type() const { return tf()->range(); }
727 const Type* CallNode::Value(PhaseGVN* phase) const {
728 if (phase->type(in(0)) == Type::TOP) return Type::TOP;
729 return tf()->range();
730 }
731
732 //------------------------------calling_convention-----------------------------
733 void CallNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
734 // Use the standard compiler calling convention
735 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt);
736 }
737
738
739 //------------------------------match------------------------------------------
740 // Construct projections for control, I/O, memory-fields, ..., and
741 // return result(s) along with their RegMask info
742 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
743 switch (proj->_con) {
744 case TypeFunc::Control:
745 case TypeFunc::I_O:
746 case TypeFunc::Memory:
747 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
748
749 case TypeFunc::Parms+1: // For LONG & DOUBLE returns
750 assert(tf()->range()->field_at(TypeFunc::Parms+1) == Type::HALF, "");
751 // 2nd half of doubles and longs
752 return new MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad);
753
754 case TypeFunc::Parms: { // Normal returns
755 uint ideal_reg = tf()->range()->field_at(TypeFunc::Parms)->ideal_reg();
756 OptoRegPair regs = Opcode() == Op_CallLeafVector
757 ? match->vector_return_value(ideal_reg) // Calls into assembly vector routine
758 : is_CallRuntime()
759 ? match->c_return_value(ideal_reg) // Calls into C runtime
760 : match-> return_value(ideal_reg); // Calls into compiled Java code
761 RegMask rm = RegMask(regs.first());
762
763 if (Opcode() == Op_CallLeafVector) {
764 // If the return is in vector, compute appropriate regmask taking into account the whole range
765 if(ideal_reg >= Op_VecS && ideal_reg <= Op_VecZ) {
766 if(OptoReg::is_valid(regs.second())) {
767 for (OptoReg::Name r = regs.first(); r <= regs.second(); r = OptoReg::add(r, 1)) {
768 rm.Insert(r);
769 }
770 }
771 }
772 }
773
774 if( OptoReg::is_valid(regs.second()) )
775 rm.Insert( regs.second() );
776 return new MachProjNode(this,proj->_con,rm,ideal_reg);
777 }
778
779 case TypeFunc::ReturnAdr:
780 case TypeFunc::FramePtr:
781 default:
782 ShouldNotReachHere();
783 }
784 return NULL;
785 }
786
787 // Do we Match on this edge index or not? Match no edges
788 uint CallNode::match_edge(uint idx) const {
789 return 0;
790 }
791
792 //
793 // Determine whether the call could modify the field of the specified
794 // instance at the specified offset.
795 //
796 bool CallNode::may_modify(const TypeOopPtr *t_oop, PhaseTransform *phase) {
797 assert((t_oop != NULL), "sanity");
798 if (is_call_to_arraycopystub() && strcmp(_name, "unsafe_arraycopy") != 0) {
799 const TypeTuple* args = _tf->domain();
800 Node* dest = NULL;
801 // Stubs that can be called once an ArrayCopyNode is expanded have
802 // different signatures. Look for the second pointer argument,
803 // that is the destination of the copy.
804 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
805 if (args->field_at(i)->isa_ptr()) {
806 j++;
807 if (j == 2) {
808 dest = in(i);
809 break;
810 }
811 }
812 }
813 guarantee(dest != NULL, "Call had only one ptr in, broken IR!");
814 if (!dest->is_top() && may_modify_arraycopy_helper(phase->type(dest)->is_oopptr(), t_oop, phase)) {
815 return true;
816 }
817 return false;
818 }
819 if (t_oop->is_known_instance()) {
820 // The instance_id is set only for scalar-replaceable allocations which
821 // are not passed as arguments according to Escape Analysis.
822 return false;
823 }
824 if (t_oop->is_ptr_to_boxed_value()) {
825 ciKlass* boxing_klass = t_oop->is_instptr()->instance_klass();
826 if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) {
827 // Skip unrelated boxing methods.
828 Node* proj = proj_out_or_null(TypeFunc::Parms);
829 if ((proj == NULL) || (phase->type(proj)->is_instptr()->instance_klass() != boxing_klass)) {
830 return false;
831 }
832 }
833 if (is_CallJava() && as_CallJava()->method() != NULL) {
834 ciMethod* meth = as_CallJava()->method();
835 if (meth->is_getter()) {
836 return false;
837 }
838 // May modify (by reflection) if an boxing object is passed
839 // as argument or returned.
840 Node* proj = returns_pointer() ? proj_out_or_null(TypeFunc::Parms) : NULL;
841 if (proj != NULL) {
842 const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr();
843 if ((inst_t != NULL) && (!inst_t->klass_is_exact() ||
844 (inst_t->instance_klass() == boxing_klass))) {
845 return true;
846 }
847 }
848 const TypeTuple* d = tf()->domain();
849 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
850 const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr();
851 if ((inst_t != NULL) && (!inst_t->klass_is_exact() ||
852 (inst_t->instance_klass() == boxing_klass))) {
853 return true;
854 }
855 }
856 return false;
857 }
858 }
859 return true;
860 }
861
862 // Does this call have a direct reference to n other than debug information?
863 bool CallNode::has_non_debug_use(Node *n) {
864 const TypeTuple * d = tf()->domain();
865 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
866 Node *arg = in(i);
867 if (arg == n) {
868 return true;
869 }
870 }
871 return false;
872 }
873
874 // Returns the unique CheckCastPP of a call
875 // or 'this' if there are several CheckCastPP or unexpected uses
876 // or returns NULL if there is no one.
877 Node *CallNode::result_cast() {
878 Node *cast = NULL;
879
880 Node *p = proj_out_or_null(TypeFunc::Parms);
881 if (p == NULL)
882 return NULL;
883
884 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
885 Node *use = p->fast_out(i);
886 if (use->is_CheckCastPP()) {
887 if (cast != NULL) {
888 return this; // more than 1 CheckCastPP
889 }
890 cast = use;
891 } else if (!use->is_Initialize() &&
892 !use->is_AddP() &&
893 use->Opcode() != Op_MemBarStoreStore) {
894 // Expected uses are restricted to a CheckCastPP, an Initialize
895 // node, a MemBarStoreStore (clone) and AddP nodes. If we
896 // encounter any other use (a Phi node can be seen in rare
897 // cases) return this to prevent incorrect optimizations.
898 return this;
899 }
900 }
901 return cast;
902 }
903
904
905 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts) {
906 projs->fallthrough_proj = NULL;
907 projs->fallthrough_catchproj = NULL;
908 projs->fallthrough_ioproj = NULL;
909 projs->catchall_ioproj = NULL;
910 projs->catchall_catchproj = NULL;
911 projs->fallthrough_memproj = NULL;
912 projs->catchall_memproj = NULL;
913 projs->resproj = NULL;
914 projs->exobj = NULL;
915
916 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
917 ProjNode *pn = fast_out(i)->as_Proj();
918 if (pn->outcnt() == 0) continue;
919 switch (pn->_con) {
920 case TypeFunc::Control:
921 {
922 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
923 projs->fallthrough_proj = pn;
924 const Node* cn = pn->unique_ctrl_out_or_null();
925 if (cn != NULL && cn->is_Catch()) {
926 ProjNode *cpn = NULL;
927 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
928 cpn = cn->fast_out(k)->as_Proj();
929 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
930 if (cpn->_con == CatchProjNode::fall_through_index)
931 projs->fallthrough_catchproj = cpn;
932 else {
933 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
934 projs->catchall_catchproj = cpn;
935 }
936 }
937 }
938 break;
939 }
940 case TypeFunc::I_O:
941 if (pn->_is_io_use)
942 projs->catchall_ioproj = pn;
943 else
944 projs->fallthrough_ioproj = pn;
945 for (DUIterator j = pn->outs(); pn->has_out(j); j++) {
946 Node* e = pn->out(j);
947 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) {
948 assert(projs->exobj == NULL, "only one");
949 projs->exobj = e;
950 }
951 }
952 break;
953 case TypeFunc::Memory:
954 if (pn->_is_io_use)
955 projs->catchall_memproj = pn;
956 else
957 projs->fallthrough_memproj = pn;
958 break;
959 case TypeFunc::Parms:
960 projs->resproj = pn;
961 break;
962 default:
963 assert(false, "unexpected projection from allocation node.");
964 }
965 }
966
967 // The resproj may not exist because the result could be ignored
968 // and the exception object may not exist if an exception handler
969 // swallows the exception but all the other must exist and be found.
970 assert(projs->fallthrough_proj != NULL, "must be found");
971 do_asserts = do_asserts && !Compile::current()->inlining_incrementally();
972 assert(!do_asserts || projs->fallthrough_catchproj != NULL, "must be found");
973 assert(!do_asserts || projs->fallthrough_memproj != NULL, "must be found");
974 assert(!do_asserts || projs->fallthrough_ioproj != NULL, "must be found");
975 assert(!do_asserts || projs->catchall_catchproj != NULL, "must be found");
976 if (separate_io_proj) {
977 assert(!do_asserts || projs->catchall_memproj != NULL, "must be found");
978 assert(!do_asserts || projs->catchall_ioproj != NULL, "must be found");
979 }
980 }
981
982 Node* CallNode::Ideal(PhaseGVN* phase, bool can_reshape) {
983 #ifdef ASSERT
984 // Validate attached generator
985 CallGenerator* cg = generator();
986 if (cg != NULL) {
987 assert(is_CallStaticJava() && cg->is_mh_late_inline() ||
988 is_CallDynamicJava() && cg->is_virtual_late_inline(), "mismatch");
989 }
990 #endif // ASSERT
991 return SafePointNode::Ideal(phase, can_reshape);
992 }
993
994 bool CallNode::is_call_to_arraycopystub() const {
995 if (_name != NULL && strstr(_name, "arraycopy") != 0) {
996 return true;
997 }
998 return false;
999 }
1000
1001 //=============================================================================
1002 uint CallJavaNode::size_of() const { return sizeof(*this); }
1003 bool CallJavaNode::cmp( const Node &n ) const {
1004 CallJavaNode &call = (CallJavaNode&)n;
1005 return CallNode::cmp(call) && _method == call._method &&
1006 _override_symbolic_info == call._override_symbolic_info;
1007 }
1008
1009 void CallJavaNode::copy_call_debug_info(PhaseIterGVN* phase, SafePointNode* sfpt) {
1010 // Copy debug information and adjust JVMState information
1011 uint old_dbg_start = sfpt->is_Call() ? sfpt->as_Call()->tf()->domain()->cnt() : (uint)TypeFunc::Parms+1;
1012 uint new_dbg_start = tf()->domain()->cnt();
1013 int jvms_adj = new_dbg_start - old_dbg_start;
1014 assert (new_dbg_start == req(), "argument count mismatch");
1015 Compile* C = phase->C;
1016
1017 // SafePointScalarObject node could be referenced several times in debug info.
1018 // Use Dict to record cloned nodes.
1019 Dict* sosn_map = new Dict(cmpkey,hashkey);
1020 for (uint i = old_dbg_start; i < sfpt->req(); i++) {
1021 Node* old_in = sfpt->in(i);
1022 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
1023 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
1024 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
1025 bool new_node;
1026 Node* new_in = old_sosn->clone(sosn_map, new_node);
1027 if (new_node) { // New node?
1028 new_in->set_req(0, C->root()); // reset control edge
1029 new_in = phase->transform(new_in); // Register new node.
1030 }
1031 old_in = new_in;
1032 }
1033 add_req(old_in);
1034 }
1035
1036 // JVMS may be shared so clone it before we modify it
1037 set_jvms(sfpt->jvms() != NULL ? sfpt->jvms()->clone_deep(C) : NULL);
1038 for (JVMState *jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
1039 jvms->set_map(this);
1040 jvms->set_locoff(jvms->locoff()+jvms_adj);
1041 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
1042 jvms->set_monoff(jvms->monoff()+jvms_adj);
1043 jvms->set_scloff(jvms->scloff()+jvms_adj);
1044 jvms->set_endoff(jvms->endoff()+jvms_adj);
1045 }
1046 }
1047
1048 #ifdef ASSERT
1049 bool CallJavaNode::validate_symbolic_info() const {
1050 if (method() == NULL) {
1051 return true; // call into runtime or uncommon trap
1052 }
1053 ciMethod* symbolic_info = jvms()->method()->get_method_at_bci(jvms()->bci());
1054 ciMethod* callee = method();
1055 if (symbolic_info->is_method_handle_intrinsic() && !callee->is_method_handle_intrinsic()) {
1056 assert(override_symbolic_info(), "should be set");
1057 }
1058 assert(ciMethod::is_consistent_info(symbolic_info, callee), "inconsistent info");
1059 return true;
1060 }
1061 #endif
1062
1063 #ifndef PRODUCT
1064 void CallJavaNode::dump_spec(outputStream* st) const {
1065 if( _method ) _method->print_short_name(st);
1066 CallNode::dump_spec(st);
1067 }
1068
1069 void CallJavaNode::dump_compact_spec(outputStream* st) const {
1070 if (_method) {
1071 _method->print_short_name(st);
1072 } else {
1073 st->print("<?>");
1074 }
1075 }
1076 #endif
1077
1078 //=============================================================================
1079 uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
1080 bool CallStaticJavaNode::cmp( const Node &n ) const {
1081 CallStaticJavaNode &call = (CallStaticJavaNode&)n;
1082 return CallJavaNode::cmp(call);
1083 }
1084
1085 Node* CallStaticJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1086 CallGenerator* cg = generator();
1087 if (can_reshape && cg != NULL) {
1088 assert(IncrementalInlineMH, "required");
1089 assert(cg->call_node() == this, "mismatch");
1090 assert(cg->is_mh_late_inline(), "not virtual");
1091
1092 // Check whether this MH handle call becomes a candidate for inlining.
1093 ciMethod* callee = cg->method();
1094 vmIntrinsics::ID iid = callee->intrinsic_id();
1095 if (iid == vmIntrinsics::_invokeBasic) {
1096 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) {
1097 phase->C->prepend_late_inline(cg);
1098 set_generator(NULL);
1099 }
1100 } else if (iid == vmIntrinsics::_linkToNative) {
1101 // never retry
1102 } else {
1103 assert(callee->has_member_arg(), "wrong type of call?");
1104 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) {
1105 phase->C->prepend_late_inline(cg);
1106 set_generator(NULL);
1107 }
1108 }
1109 }
1110 return CallNode::Ideal(phase, can_reshape);
1111 }
1112
1113 //----------------------------uncommon_trap_request----------------------------
1114 // If this is an uncommon trap, return the request code, else zero.
1115 int CallStaticJavaNode::uncommon_trap_request() const {
1116 if (_name != NULL && !strcmp(_name, "uncommon_trap")) {
1117 return extract_uncommon_trap_request(this);
1118 }
1119 return 0;
1120 }
1121 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
1122 #ifndef PRODUCT
1123 if (!(call->req() > TypeFunc::Parms &&
1124 call->in(TypeFunc::Parms) != NULL &&
1125 call->in(TypeFunc::Parms)->is_Con() &&
1126 call->in(TypeFunc::Parms)->bottom_type()->isa_int())) {
1127 assert(in_dump() != 0, "OK if dumping");
1128 tty->print("[bad uncommon trap]");
1129 return 0;
1130 }
1131 #endif
1132 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
1133 }
1134
1135 #ifndef PRODUCT
1136 void CallStaticJavaNode::dump_spec(outputStream *st) const {
1137 st->print("# Static ");
1138 if (_name != NULL) {
1139 st->print("%s", _name);
1140 int trap_req = uncommon_trap_request();
1141 if (trap_req != 0) {
1142 char buf[100];
1143 st->print("(%s)",
1144 Deoptimization::format_trap_request(buf, sizeof(buf),
1145 trap_req));
1146 }
1147 st->print(" ");
1148 }
1149 CallJavaNode::dump_spec(st);
1150 }
1151
1152 void CallStaticJavaNode::dump_compact_spec(outputStream* st) const {
1153 if (_method) {
1154 _method->print_short_name(st);
1155 } else if (_name) {
1156 st->print("%s", _name);
1157 } else {
1158 st->print("<?>");
1159 }
1160 }
1161 #endif
1162
1163 //=============================================================================
1164 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
1165 bool CallDynamicJavaNode::cmp( const Node &n ) const {
1166 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
1167 return CallJavaNode::cmp(call);
1168 }
1169
1170 Node* CallDynamicJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1171 CallGenerator* cg = generator();
1172 if (can_reshape && cg != NULL) {
1173 assert(IncrementalInlineVirtual, "required");
1174 assert(cg->call_node() == this, "mismatch");
1175 assert(cg->is_virtual_late_inline(), "not virtual");
1176
1177 // Recover symbolic info for method resolution.
1178 ciMethod* caller = jvms()->method();
1179 ciBytecodeStream iter(caller);
1180 iter.force_bci(jvms()->bci());
1181
1182 bool not_used1;
1183 ciSignature* not_used2;
1184 ciMethod* orig_callee = iter.get_method(not_used1, ¬_used2); // callee in the bytecode
1185 ciKlass* holder = iter.get_declared_method_holder();
1186 if (orig_callee->is_method_handle_intrinsic()) {
1187 assert(_override_symbolic_info, "required");
1188 orig_callee = method();
1189 holder = method()->holder();
1190 }
1191
1192 ciInstanceKlass* klass = ciEnv::get_instance_klass_for_declared_method_holder(holder);
1193
1194 Node* receiver_node = in(TypeFunc::Parms);
1195 const TypeOopPtr* receiver_type = phase->type(receiver_node)->isa_oopptr();
1196
1197 int not_used3;
1198 bool call_does_dispatch;
1199 ciMethod* callee = phase->C->optimize_virtual_call(caller, klass, holder, orig_callee, receiver_type, true /*is_virtual*/,
1200 call_does_dispatch, not_used3); // out-parameters
1201 if (!call_does_dispatch) {
1202 // Register for late inlining.
1203 cg->set_callee_method(callee);
1204 phase->C->prepend_late_inline(cg); // MH late inlining prepends to the list, so do the same
1205 set_generator(NULL);
1206 }
1207 }
1208 return CallNode::Ideal(phase, can_reshape);
1209 }
1210
1211 #ifndef PRODUCT
1212 void CallDynamicJavaNode::dump_spec(outputStream *st) const {
1213 st->print("# Dynamic ");
1214 CallJavaNode::dump_spec(st);
1215 }
1216 #endif
1217
1218 //=============================================================================
1219 uint CallRuntimeNode::size_of() const { return sizeof(*this); }
1220 bool CallRuntimeNode::cmp( const Node &n ) const {
1221 CallRuntimeNode &call = (CallRuntimeNode&)n;
1222 return CallNode::cmp(call) && !strcmp(_name,call._name);
1223 }
1224 #ifndef PRODUCT
1225 void CallRuntimeNode::dump_spec(outputStream *st) const {
1226 st->print("# ");
1227 st->print("%s", _name);
1228 CallNode::dump_spec(st);
1229 }
1230 #endif
1231 uint CallLeafVectorNode::size_of() const { return sizeof(*this); }
1232 bool CallLeafVectorNode::cmp( const Node &n ) const {
1233 CallLeafVectorNode &call = (CallLeafVectorNode&)n;
1234 return CallLeafNode::cmp(call) && _num_bits == call._num_bits;
1235 }
1236
1237 //------------------------------calling_convention-----------------------------
1238 void CallRuntimeNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
1239 SharedRuntime::c_calling_convention(sig_bt, parm_regs, /*regs2=*/nullptr, argcnt);
1240 }
1241
1242 void CallLeafVectorNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
1243 #ifdef ASSERT
1244 assert(tf()->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1245 "return vector size must match");
1246 const TypeTuple* d = tf()->domain();
1247 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1248 Node* arg = in(i);
1249 assert(arg->bottom_type()->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1250 "vector argument size must match");
1251 }
1252 #endif
1253
1254 SharedRuntime::vector_calling_convention(parm_regs, _num_bits, argcnt);
1255 }
1256
1257 //=============================================================================
1258 //------------------------------calling_convention-----------------------------
1259
1260
1261 //=============================================================================
1262 #ifndef PRODUCT
1263 void CallLeafNode::dump_spec(outputStream *st) const {
1264 st->print("# ");
1265 st->print("%s", _name);
1266 CallNode::dump_spec(st);
1267 }
1268 #endif
1269
1270 //=============================================================================
1271
1272 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) {
1273 assert(verify_jvms(jvms), "jvms must match");
1274 int loc = jvms->locoff() + idx;
1275 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
1276 // If current local idx is top then local idx - 1 could
1277 // be a long/double that needs to be killed since top could
1278 // represent the 2nd half of the long/double.
1279 uint ideal = in(loc -1)->ideal_reg();
1280 if (ideal == Op_RegD || ideal == Op_RegL) {
1281 // set other (low index) half to top
1282 set_req(loc - 1, in(loc));
1283 }
1284 }
1285 set_req(loc, c);
1286 }
1287
1288 uint SafePointNode::size_of() const { return sizeof(*this); }
1289 bool SafePointNode::cmp( const Node &n ) const {
1290 return (&n == this); // Always fail except on self
1291 }
1292
1293 //-------------------------set_next_exception----------------------------------
1294 void SafePointNode::set_next_exception(SafePointNode* n) {
1295 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception");
1296 if (len() == req()) {
1297 if (n != NULL) add_prec(n);
1298 } else {
1299 set_prec(req(), n);
1300 }
1301 }
1302
1303
1304 //----------------------------next_exception-----------------------------------
1305 SafePointNode* SafePointNode::next_exception() const {
1306 if (len() == req()) {
1307 return NULL;
1308 } else {
1309 Node* n = in(req());
1310 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
1311 return (SafePointNode*) n;
1312 }
1313 }
1314
1315
1316 //------------------------------Ideal------------------------------------------
1317 // Skip over any collapsed Regions
1318 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1319 assert(_jvms == NULL || ((uintptr_t)_jvms->map() & 1) || _jvms->map() == this, "inconsistent JVMState");
1320 return remove_dead_region(phase, can_reshape) ? this : NULL;
1321 }
1322
1323 //------------------------------Identity---------------------------------------
1324 // Remove obviously duplicate safepoints
1325 Node* SafePointNode::Identity(PhaseGVN* phase) {
1326
1327 // If you have back to back safepoints, remove one
1328 if (in(TypeFunc::Control)->is_SafePoint()) {
1329 Node* out_c = unique_ctrl_out_or_null();
1330 // This can be the safepoint of an outer strip mined loop if the inner loop's backedge was removed. Replacing the
1331 // outer loop's safepoint could confuse removal of the outer loop.
1332 if (out_c != NULL && !out_c->is_OuterStripMinedLoopEnd()) {
1333 return in(TypeFunc::Control);
1334 }
1335 }
1336
1337 // Transforming long counted loops requires a safepoint node. Do not
1338 // eliminate a safepoint until loop opts are over.
1339 if (in(0)->is_Proj() && !phase->C->major_progress()) {
1340 Node *n0 = in(0)->in(0);
1341 // Check if he is a call projection (except Leaf Call)
1342 if( n0->is_Catch() ) {
1343 n0 = n0->in(0)->in(0);
1344 assert( n0->is_Call(), "expect a call here" );
1345 }
1346 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
1347 // Don't remove a safepoint belonging to an OuterStripMinedLoopEndNode.
1348 // If the loop dies, they will be removed together.
1349 if (has_out_with(Op_OuterStripMinedLoopEnd)) {
1350 return this;
1351 }
1352 // Useless Safepoint, so remove it
1353 return in(TypeFunc::Control);
1354 }
1355 }
1356
1357 return this;
1358 }
1359
1360 //------------------------------Value------------------------------------------
1361 const Type* SafePointNode::Value(PhaseGVN* phase) const {
1362 if (phase->type(in(0)) == Type::TOP) {
1363 return Type::TOP;
1364 }
1365 if (in(0) == this) {
1366 return Type::TOP; // Dead infinite loop
1367 }
1368 return Type::CONTROL;
1369 }
1370
1371 #ifndef PRODUCT
1372 void SafePointNode::dump_spec(outputStream *st) const {
1373 st->print(" SafePoint ");
1374 _replaced_nodes.dump(st);
1375 }
1376
1377 // The related nodes of a SafepointNode are all data inputs, excluding the
1378 // control boundary, as well as all outputs till level 2 (to include projection
1379 // nodes and targets). In compact mode, just include inputs till level 1 and
1380 // outputs as before.
1381 void SafePointNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
1382 if (compact) {
1383 this->collect_nodes(in_rel, 1, false, false);
1384 } else {
1385 this->collect_nodes_in_all_data(in_rel, false);
1386 }
1387 this->collect_nodes(out_rel, -2, false, false);
1388 }
1389 #endif
1390
1391 const RegMask &SafePointNode::in_RegMask(uint idx) const {
1392 if( idx < TypeFunc::Parms ) return RegMask::Empty;
1393 // Values outside the domain represent debug info
1394 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1395 }
1396 const RegMask &SafePointNode::out_RegMask() const {
1397 return RegMask::Empty;
1398 }
1399
1400
1401 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
1402 assert((int)grow_by > 0, "sanity");
1403 int monoff = jvms->monoff();
1404 int scloff = jvms->scloff();
1405 int endoff = jvms->endoff();
1406 assert(endoff == (int)req(), "no other states or debug info after me");
1407 Node* top = Compile::current()->top();
1408 for (uint i = 0; i < grow_by; i++) {
1409 ins_req(monoff, top);
1410 }
1411 jvms->set_monoff(monoff + grow_by);
1412 jvms->set_scloff(scloff + grow_by);
1413 jvms->set_endoff(endoff + grow_by);
1414 }
1415
1416 void SafePointNode::push_monitor(const FastLockNode *lock) {
1417 // Add a LockNode, which points to both the original BoxLockNode (the
1418 // stack space for the monitor) and the Object being locked.
1419 const int MonitorEdges = 2;
1420 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1421 assert(req() == jvms()->endoff(), "correct sizing");
1422 int nextmon = jvms()->scloff();
1423 if (GenerateSynchronizationCode) {
1424 ins_req(nextmon, lock->box_node());
1425 ins_req(nextmon+1, lock->obj_node());
1426 } else {
1427 Node* top = Compile::current()->top();
1428 ins_req(nextmon, top);
1429 ins_req(nextmon, top);
1430 }
1431 jvms()->set_scloff(nextmon + MonitorEdges);
1432 jvms()->set_endoff(req());
1433 }
1434
1435 void SafePointNode::pop_monitor() {
1436 // Delete last monitor from debug info
1437 debug_only(int num_before_pop = jvms()->nof_monitors());
1438 const int MonitorEdges = 2;
1439 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1440 int scloff = jvms()->scloff();
1441 int endoff = jvms()->endoff();
1442 int new_scloff = scloff - MonitorEdges;
1443 int new_endoff = endoff - MonitorEdges;
1444 jvms()->set_scloff(new_scloff);
1445 jvms()->set_endoff(new_endoff);
1446 while (scloff > new_scloff) del_req_ordered(--scloff);
1447 assert(jvms()->nof_monitors() == num_before_pop-1, "");
1448 }
1449
1450 Node *SafePointNode::peek_monitor_box() const {
1451 int mon = jvms()->nof_monitors() - 1;
1452 assert(mon >= 0, "must have a monitor");
1453 return monitor_box(jvms(), mon);
1454 }
1455
1456 Node *SafePointNode::peek_monitor_obj() const {
1457 int mon = jvms()->nof_monitors() - 1;
1458 assert(mon >= 0, "must have a monitor");
1459 return monitor_obj(jvms(), mon);
1460 }
1461
1462 Node* SafePointNode::peek_operand(uint off) const {
1463 assert(jvms()->sp() > 0, "must have an operand");
1464 assert(off < jvms()->sp(), "off is out-of-range");
1465 return stack(jvms(), jvms()->sp() - off - 1);
1466 }
1467
1468 // Do we Match on this edge index or not? Match no edges
1469 uint SafePointNode::match_edge(uint idx) const {
1470 return (TypeFunc::Parms == idx);
1471 }
1472
1473 void SafePointNode::disconnect_from_root(PhaseIterGVN *igvn) {
1474 assert(Opcode() == Op_SafePoint, "only value for safepoint in loops");
1475 int nb = igvn->C->root()->find_prec_edge(this);
1476 if (nb != -1) {
1477 igvn->C->root()->rm_prec(nb);
1478 }
1479 }
1480
1481 //============== SafePointScalarObjectNode ==============
1482
1483 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp,
1484 #ifdef ASSERT
1485 Node* alloc,
1486 #endif
1487 uint first_index,
1488 uint n_fields) :
1489 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1490 _first_index(first_index),
1491 _n_fields(n_fields)
1492 #ifdef ASSERT
1493 , _alloc(alloc)
1494 #endif
1495 {
1496 #ifdef ASSERT
1497 if (!alloc->is_Allocate()
1498 && !(alloc->Opcode() == Op_VectorBox)) {
1499 alloc->dump();
1500 assert(false, "unexpected call node");
1501 }
1502 #endif
1503 init_class_id(Class_SafePointScalarObject);
1504 }
1505
1506 // Do not allow value-numbering for SafePointScalarObject node.
1507 uint SafePointScalarObjectNode::hash() const { return NO_HASH; }
1508 bool SafePointScalarObjectNode::cmp( const Node &n ) const {
1509 return (&n == this); // Always fail except on self
1510 }
1511
1512 uint SafePointScalarObjectNode::ideal_reg() const {
1513 return 0; // No matching to machine instruction
1514 }
1515
1516 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1517 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1518 }
1519
1520 const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1521 return RegMask::Empty;
1522 }
1523
1524 uint SafePointScalarObjectNode::match_edge(uint idx) const {
1525 return 0;
1526 }
1527
1528 SafePointScalarObjectNode*
1529 SafePointScalarObjectNode::clone(Dict* sosn_map, bool& new_node) const {
1530 void* cached = (*sosn_map)[(void*)this];
1531 if (cached != NULL) {
1532 new_node = false;
1533 return (SafePointScalarObjectNode*)cached;
1534 }
1535 new_node = true;
1536 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1537 sosn_map->Insert((void*)this, (void*)res);
1538 return res;
1539 }
1540
1541
1542 #ifndef PRODUCT
1543 void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1544 st->print(" # fields@[%d..%d]", first_index(),
1545 first_index() + n_fields() - 1);
1546 }
1547
1548 #endif
1549
1550 //=============================================================================
1551 uint AllocateNode::size_of() const { return sizeof(*this); }
1552
1553 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1554 Node *ctrl, Node *mem, Node *abio,
1555 Node *size, Node *klass_node, Node *initial_test)
1556 : CallNode(atype, NULL, TypeRawPtr::BOTTOM)
1557 {
1558 init_class_id(Class_Allocate);
1559 init_flags(Flag_is_macro);
1560 _is_scalar_replaceable = false;
1561 _is_non_escaping = false;
1562 _is_allocation_MemBar_redundant = false;
1563 Node *topnode = C->top();
1564
1565 init_req( TypeFunc::Control , ctrl );
1566 init_req( TypeFunc::I_O , abio );
1567 init_req( TypeFunc::Memory , mem );
1568 init_req( TypeFunc::ReturnAdr, topnode );
1569 init_req( TypeFunc::FramePtr , topnode );
1570 init_req( AllocSize , size);
1571 init_req( KlassNode , klass_node);
1572 init_req( InitialTest , initial_test);
1573 init_req( ALength , topnode);
1574 init_req( ValidLengthTest , topnode);
1575 C->add_macro_node(this);
1576 }
1577
1578 void AllocateNode::compute_MemBar_redundancy(ciMethod* initializer)
1579 {
1580 assert(initializer != NULL &&
1581 initializer->is_initializer() &&
1582 !initializer->is_static(),
1583 "unexpected initializer method");
1584 BCEscapeAnalyzer* analyzer = initializer->get_bcea();
1585 if (analyzer == NULL) {
1586 return;
1587 }
1588
1589 // Allocation node is first parameter in its initializer
1590 if (analyzer->is_arg_stack(0) || analyzer->is_arg_local(0)) {
1591 _is_allocation_MemBar_redundant = true;
1592 }
1593 }
1594 Node *AllocateNode::make_ideal_mark(PhaseGVN *phase, Node* obj, Node* control, Node* mem) {
1595 Node* mark_node = NULL;
1596 // For now only enable fast locking for non-array types
1597 mark_node = phase->MakeConX(markWord::prototype().value());
1598 return mark_node;
1599 }
1600
1601 // Retrieve the length from the AllocateArrayNode. Narrow the type with a
1602 // CastII, if appropriate. If we are not allowed to create new nodes, and
1603 // a CastII is appropriate, return NULL.
1604 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) {
1605 Node *length = in(AllocateNode::ALength);
1606 assert(length != NULL, "length is not null");
1607
1608 const TypeInt* length_type = phase->find_int_type(length);
1609 const TypeAryPtr* ary_type = oop_type->isa_aryptr();
1610
1611 if (ary_type != NULL && length_type != NULL) {
1612 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type);
1613 if (narrow_length_type != length_type) {
1614 // Assert one of:
1615 // - the narrow_length is 0
1616 // - the narrow_length is not wider than length
1617 assert(narrow_length_type == TypeInt::ZERO ||
1618 length_type->is_con() && narrow_length_type->is_con() &&
1619 (narrow_length_type->_hi <= length_type->_lo) ||
1620 (narrow_length_type->_hi <= length_type->_hi &&
1621 narrow_length_type->_lo >= length_type->_lo),
1622 "narrow type must be narrower than length type");
1623
1624 // Return NULL if new nodes are not allowed
1625 if (!allow_new_nodes) {
1626 return NULL;
1627 }
1628 // Create a cast which is control dependent on the initialization to
1629 // propagate the fact that the array length must be positive.
1630 InitializeNode* init = initialization();
1631 if (init != NULL) {
1632 length = new CastIINode(length, narrow_length_type);
1633 length->set_req(TypeFunc::Control, init->proj_out_or_null(TypeFunc::Control));
1634 }
1635 }
1636 }
1637
1638 return length;
1639 }
1640
1641 //=============================================================================
1642 uint LockNode::size_of() const { return sizeof(*this); }
1643
1644 // Redundant lock elimination
1645 //
1646 // There are various patterns of locking where we release and
1647 // immediately reacquire a lock in a piece of code where no operations
1648 // occur in between that would be observable. In those cases we can
1649 // skip releasing and reacquiring the lock without violating any
1650 // fairness requirements. Doing this around a loop could cause a lock
1651 // to be held for a very long time so we concentrate on non-looping
1652 // control flow. We also require that the operations are fully
1653 // redundant meaning that we don't introduce new lock operations on
1654 // some paths so to be able to eliminate it on others ala PRE. This
1655 // would probably require some more extensive graph manipulation to
1656 // guarantee that the memory edges were all handled correctly.
1657 //
1658 // Assuming p is a simple predicate which can't trap in any way and s
1659 // is a synchronized method consider this code:
1660 //
1661 // s();
1662 // if (p)
1663 // s();
1664 // else
1665 // s();
1666 // s();
1667 //
1668 // 1. The unlocks of the first call to s can be eliminated if the
1669 // locks inside the then and else branches are eliminated.
1670 //
1671 // 2. The unlocks of the then and else branches can be eliminated if
1672 // the lock of the final call to s is eliminated.
1673 //
1674 // Either of these cases subsumes the simple case of sequential control flow
1675 //
1676 // Additionally we can eliminate versions without the else case:
1677 //
1678 // s();
1679 // if (p)
1680 // s();
1681 // s();
1682 //
1683 // 3. In this case we eliminate the unlock of the first s, the lock
1684 // and unlock in the then case and the lock in the final s.
1685 //
1686 // Note also that in all these cases the then/else pieces don't have
1687 // to be trivial as long as they begin and end with synchronization
1688 // operations.
1689 //
1690 // s();
1691 // if (p)
1692 // s();
1693 // f();
1694 // s();
1695 // s();
1696 //
1697 // The code will work properly for this case, leaving in the unlock
1698 // before the call to f and the relock after it.
1699 //
1700 // A potentially interesting case which isn't handled here is when the
1701 // locking is partially redundant.
1702 //
1703 // s();
1704 // if (p)
1705 // s();
1706 //
1707 // This could be eliminated putting unlocking on the else case and
1708 // eliminating the first unlock and the lock in the then side.
1709 // Alternatively the unlock could be moved out of the then side so it
1710 // was after the merge and the first unlock and second lock
1711 // eliminated. This might require less manipulation of the memory
1712 // state to get correct.
1713 //
1714 // Additionally we might allow work between a unlock and lock before
1715 // giving up eliminating the locks. The current code disallows any
1716 // conditional control flow between these operations. A formulation
1717 // similar to partial redundancy elimination computing the
1718 // availability of unlocking and the anticipatability of locking at a
1719 // program point would allow detection of fully redundant locking with
1720 // some amount of work in between. I'm not sure how often I really
1721 // think that would occur though. Most of the cases I've seen
1722 // indicate it's likely non-trivial work would occur in between.
1723 // There may be other more complicated constructs where we could
1724 // eliminate locking but I haven't seen any others appear as hot or
1725 // interesting.
1726 //
1727 // Locking and unlocking have a canonical form in ideal that looks
1728 // roughly like this:
1729 //
1730 // <obj>
1731 // | \\------+
1732 // | \ \
1733 // | BoxLock \
1734 // | | | \
1735 // | | \ \
1736 // | | FastLock
1737 // | | /
1738 // | | /
1739 // | | |
1740 //
1741 // Lock
1742 // |
1743 // Proj #0
1744 // |
1745 // MembarAcquire
1746 // |
1747 // Proj #0
1748 //
1749 // MembarRelease
1750 // |
1751 // Proj #0
1752 // |
1753 // Unlock
1754 // |
1755 // Proj #0
1756 //
1757 //
1758 // This code proceeds by processing Lock nodes during PhaseIterGVN
1759 // and searching back through its control for the proper code
1760 // patterns. Once it finds a set of lock and unlock operations to
1761 // eliminate they are marked as eliminatable which causes the
1762 // expansion of the Lock and Unlock macro nodes to make the operation a NOP
1763 //
1764 //=============================================================================
1765
1766 //
1767 // Utility function to skip over uninteresting control nodes. Nodes skipped are:
1768 // - copy regions. (These may not have been optimized away yet.)
1769 // - eliminated locking nodes
1770 //
1771 static Node *next_control(Node *ctrl) {
1772 if (ctrl == NULL)
1773 return NULL;
1774 while (1) {
1775 if (ctrl->is_Region()) {
1776 RegionNode *r = ctrl->as_Region();
1777 Node *n = r->is_copy();
1778 if (n == NULL)
1779 break; // hit a region, return it
1780 else
1781 ctrl = n;
1782 } else if (ctrl->is_Proj()) {
1783 Node *in0 = ctrl->in(0);
1784 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
1785 ctrl = in0->in(0);
1786 } else {
1787 break;
1788 }
1789 } else {
1790 break; // found an interesting control
1791 }
1792 }
1793 return ctrl;
1794 }
1795 //
1796 // Given a control, see if it's the control projection of an Unlock which
1797 // operating on the same object as lock.
1798 //
1799 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
1800 GrowableArray<AbstractLockNode*> &lock_ops) {
1801 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL;
1802 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) {
1803 Node *n = ctrl_proj->in(0);
1804 if (n != NULL && n->is_Unlock()) {
1805 UnlockNode *unlock = n->as_Unlock();
1806 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1807 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1808 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
1809 if (lock_obj->eqv_uncast(unlock_obj) &&
1810 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) &&
1811 !unlock->is_eliminated()) {
1812 lock_ops.append(unlock);
1813 return true;
1814 }
1815 }
1816 }
1817 return false;
1818 }
1819
1820 //
1821 // Find the lock matching an unlock. Returns null if a safepoint
1822 // or complicated control is encountered first.
1823 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
1824 LockNode *lock_result = NULL;
1825 // find the matching lock, or an intervening safepoint
1826 Node *ctrl = next_control(unlock->in(0));
1827 while (1) {
1828 assert(ctrl != NULL, "invalid control graph");
1829 assert(!ctrl->is_Start(), "missing lock for unlock");
1830 if (ctrl->is_top()) break; // dead control path
1831 if (ctrl->is_Proj()) ctrl = ctrl->in(0);
1832 if (ctrl->is_SafePoint()) {
1833 break; // found a safepoint (may be the lock we are searching for)
1834 } else if (ctrl->is_Region()) {
1835 // Check for a simple diamond pattern. Punt on anything more complicated
1836 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) {
1837 Node *in1 = next_control(ctrl->in(1));
1838 Node *in2 = next_control(ctrl->in(2));
1839 if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
1840 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
1841 ctrl = next_control(in1->in(0)->in(0));
1842 } else {
1843 break;
1844 }
1845 } else {
1846 break;
1847 }
1848 } else {
1849 ctrl = next_control(ctrl->in(0)); // keep searching
1850 }
1851 }
1852 if (ctrl->is_Lock()) {
1853 LockNode *lock = ctrl->as_Lock();
1854 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1855 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1856 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
1857 if (lock_obj->eqv_uncast(unlock_obj) &&
1858 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) {
1859 lock_result = lock;
1860 }
1861 }
1862 return lock_result;
1863 }
1864
1865 // This code corresponds to case 3 above.
1866
1867 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
1868 GrowableArray<AbstractLockNode*> &lock_ops) {
1869 Node* if_node = node->in(0);
1870 bool if_true = node->is_IfTrue();
1871
1872 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
1873 Node *lock_ctrl = next_control(if_node->in(0));
1874 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
1875 Node* lock1_node = NULL;
1876 ProjNode* proj = if_node->as_If()->proj_out(!if_true);
1877 if (if_true) {
1878 if (proj->is_IfFalse() && proj->outcnt() == 1) {
1879 lock1_node = proj->unique_out();
1880 }
1881 } else {
1882 if (proj->is_IfTrue() && proj->outcnt() == 1) {
1883 lock1_node = proj->unique_out();
1884 }
1885 }
1886 if (lock1_node != NULL && lock1_node->is_Lock()) {
1887 LockNode *lock1 = lock1_node->as_Lock();
1888 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1889 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1890 Node* lock1_obj = bs->step_over_gc_barrier(lock1->obj_node());
1891 if (lock_obj->eqv_uncast(lock1_obj) &&
1892 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) &&
1893 !lock1->is_eliminated()) {
1894 lock_ops.append(lock1);
1895 return true;
1896 }
1897 }
1898 }
1899 }
1900
1901 lock_ops.trunc_to(0);
1902 return false;
1903 }
1904
1905 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
1906 GrowableArray<AbstractLockNode*> &lock_ops) {
1907 // check each control merging at this point for a matching unlock.
1908 // in(0) should be self edge so skip it.
1909 for (int i = 1; i < (int)region->req(); i++) {
1910 Node *in_node = next_control(region->in(i));
1911 if (in_node != NULL) {
1912 if (find_matching_unlock(in_node, lock, lock_ops)) {
1913 // found a match so keep on checking.
1914 continue;
1915 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
1916 continue;
1917 }
1918
1919 // If we fall through to here then it was some kind of node we
1920 // don't understand or there wasn't a matching unlock, so give
1921 // up trying to merge locks.
1922 lock_ops.trunc_to(0);
1923 return false;
1924 }
1925 }
1926 return true;
1927
1928 }
1929
1930 const char* AbstractLockNode::_kind_names[] = {"Regular", "NonEscObj", "Coarsened", "Nested"};
1931
1932 const char * AbstractLockNode::kind_as_string() const {
1933 return _kind_names[_kind];
1934 }
1935
1936 #ifndef PRODUCT
1937 //
1938 // Create a counter which counts the number of times this lock is acquired
1939 //
1940 void AbstractLockNode::create_lock_counter(JVMState* state) {
1941 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
1942 }
1943
1944 void AbstractLockNode::set_eliminated_lock_counter() {
1945 if (_counter) {
1946 // Update the counter to indicate that this lock was eliminated.
1947 // The counter update code will stay around even though the
1948 // optimizer will eliminate the lock operation itself.
1949 _counter->set_tag(NamedCounter::EliminatedLockCounter);
1950 }
1951 }
1952
1953 void AbstractLockNode::dump_spec(outputStream* st) const {
1954 st->print("%s ", _kind_names[_kind]);
1955 CallNode::dump_spec(st);
1956 }
1957
1958 void AbstractLockNode::dump_compact_spec(outputStream* st) const {
1959 st->print("%s", _kind_names[_kind]);
1960 }
1961
1962 // The related set of lock nodes includes the control boundary.
1963 void AbstractLockNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
1964 if (compact) {
1965 this->collect_nodes(in_rel, 1, false, false);
1966 } else {
1967 this->collect_nodes_in_all_data(in_rel, true);
1968 }
1969 this->collect_nodes(out_rel, -2, false, false);
1970 }
1971 #endif
1972
1973 //=============================================================================
1974 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1975
1976 // perform any generic optimizations first (returns 'this' or NULL)
1977 Node *result = SafePointNode::Ideal(phase, can_reshape);
1978 if (result != NULL) return result;
1979 // Don't bother trying to transform a dead node
1980 if (in(0) && in(0)->is_top()) return NULL;
1981
1982 // Now see if we can optimize away this lock. We don't actually
1983 // remove the locking here, we simply set the _eliminate flag which
1984 // prevents macro expansion from expanding the lock. Since we don't
1985 // modify the graph, the value returned from this function is the
1986 // one computed above.
1987 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
1988 //
1989 // If we are locking an non-escaped object, the lock/unlock is unnecessary
1990 //
1991 ConnectionGraph *cgr = phase->C->congraph();
1992 if (cgr != NULL && cgr->not_global_escape(obj_node())) {
1993 assert(!is_eliminated() || is_coarsened(), "sanity");
1994 // The lock could be marked eliminated by lock coarsening
1995 // code during first IGVN before EA. Replace coarsened flag
1996 // to eliminate all associated locks/unlocks.
1997 #ifdef ASSERT
1998 this->log_lock_optimization(phase->C,"eliminate_lock_set_non_esc1");
1999 #endif
2000 this->set_non_esc_obj();
2001 return result;
2002 }
2003
2004 if (!phase->C->do_locks_coarsening()) {
2005 return result; // Compiling without locks coarsening
2006 }
2007 //
2008 // Try lock coarsening
2009 //
2010 PhaseIterGVN* iter = phase->is_IterGVN();
2011 if (iter != NULL && !is_eliminated()) {
2012
2013 GrowableArray<AbstractLockNode*> lock_ops;
2014
2015 Node *ctrl = next_control(in(0));
2016
2017 // now search back for a matching Unlock
2018 if (find_matching_unlock(ctrl, this, lock_ops)) {
2019 // found an unlock directly preceding this lock. This is the
2020 // case of single unlock directly control dependent on a
2021 // single lock which is the trivial version of case 1 or 2.
2022 } else if (ctrl->is_Region() ) {
2023 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
2024 // found lock preceded by multiple unlocks along all paths
2025 // joining at this point which is case 3 in description above.
2026 }
2027 } else {
2028 // see if this lock comes from either half of an if and the
2029 // predecessors merges unlocks and the other half of the if
2030 // performs a lock.
2031 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
2032 // found unlock splitting to an if with locks on both branches.
2033 }
2034 }
2035
2036 if (lock_ops.length() > 0) {
2037 // add ourselves to the list of locks to be eliminated.
2038 lock_ops.append(this);
2039
2040 #ifndef PRODUCT
2041 if (PrintEliminateLocks) {
2042 int locks = 0;
2043 int unlocks = 0;
2044 if (Verbose) {
2045 tty->print_cr("=== Locks coarsening ===");
2046 }
2047 for (int i = 0; i < lock_ops.length(); i++) {
2048 AbstractLockNode* lock = lock_ops.at(i);
2049 if (lock->Opcode() == Op_Lock)
2050 locks++;
2051 else
2052 unlocks++;
2053 if (Verbose) {
2054 tty->print(" %d: ", i);
2055 lock->dump();
2056 }
2057 }
2058 tty->print_cr("=== Coarsened %d unlocks and %d locks", unlocks, locks);
2059 }
2060 #endif
2061
2062 // for each of the identified locks, mark them
2063 // as eliminatable
2064 for (int i = 0; i < lock_ops.length(); i++) {
2065 AbstractLockNode* lock = lock_ops.at(i);
2066
2067 // Mark it eliminated by coarsening and update any counters
2068 #ifdef ASSERT
2069 lock->log_lock_optimization(phase->C, "eliminate_lock_set_coarsened");
2070 #endif
2071 lock->set_coarsened();
2072 }
2073 // Record this coarsened group.
2074 phase->C->add_coarsened_locks(lock_ops);
2075 } else if (ctrl->is_Region() &&
2076 iter->_worklist.member(ctrl)) {
2077 // We weren't able to find any opportunities but the region this
2078 // lock is control dependent on hasn't been processed yet so put
2079 // this lock back on the worklist so we can check again once any
2080 // region simplification has occurred.
2081 iter->_worklist.push(this);
2082 }
2083 }
2084 }
2085
2086 return result;
2087 }
2088
2089 //=============================================================================
2090 bool LockNode::is_nested_lock_region() {
2091 return is_nested_lock_region(NULL);
2092 }
2093
2094 // p is used for access to compilation log; no logging if NULL
2095 bool LockNode::is_nested_lock_region(Compile * c) {
2096 BoxLockNode* box = box_node()->as_BoxLock();
2097 int stk_slot = box->stack_slot();
2098 if (stk_slot <= 0) {
2099 #ifdef ASSERT
2100 this->log_lock_optimization(c, "eliminate_lock_INLR_1");
2101 #endif
2102 return false; // External lock or it is not Box (Phi node).
2103 }
2104
2105 // Ignore complex cases: merged locks or multiple locks.
2106 Node* obj = obj_node();
2107 LockNode* unique_lock = NULL;
2108 Node* bad_lock = NULL;
2109 if (!box->is_simple_lock_region(&unique_lock, obj, &bad_lock)) {
2110 #ifdef ASSERT
2111 this->log_lock_optimization(c, "eliminate_lock_INLR_2a", bad_lock);
2112 #endif
2113 return false;
2114 }
2115 if (unique_lock != this) {
2116 #ifdef ASSERT
2117 this->log_lock_optimization(c, "eliminate_lock_INLR_2b", (unique_lock != NULL ? unique_lock : bad_lock));
2118 if (PrintEliminateLocks && Verbose) {
2119 tty->print_cr("=============== unique_lock != this ============");
2120 tty->print(" this: ");
2121 this->dump();
2122 tty->print(" box: ");
2123 box->dump();
2124 tty->print(" obj: ");
2125 obj->dump();
2126 if (unique_lock != NULL) {
2127 tty->print(" unique_lock: ");
2128 unique_lock->dump();
2129 }
2130 if (bad_lock != NULL) {
2131 tty->print(" bad_lock: ");
2132 bad_lock->dump();
2133 }
2134 tty->print_cr("===============");
2135 }
2136 #endif
2137 return false;
2138 }
2139
2140 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2141 obj = bs->step_over_gc_barrier(obj);
2142 // Look for external lock for the same object.
2143 SafePointNode* sfn = this->as_SafePoint();
2144 JVMState* youngest_jvms = sfn->jvms();
2145 int max_depth = youngest_jvms->depth();
2146 for (int depth = 1; depth <= max_depth; depth++) {
2147 JVMState* jvms = youngest_jvms->of_depth(depth);
2148 int num_mon = jvms->nof_monitors();
2149 // Loop over monitors
2150 for (int idx = 0; idx < num_mon; idx++) {
2151 Node* obj_node = sfn->monitor_obj(jvms, idx);
2152 obj_node = bs->step_over_gc_barrier(obj_node);
2153 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock();
2154 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) {
2155 return true;
2156 }
2157 }
2158 }
2159 #ifdef ASSERT
2160 this->log_lock_optimization(c, "eliminate_lock_INLR_3");
2161 #endif
2162 return false;
2163 }
2164
2165 //=============================================================================
2166 uint UnlockNode::size_of() const { return sizeof(*this); }
2167
2168 //=============================================================================
2169 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2170
2171 // perform any generic optimizations first (returns 'this' or NULL)
2172 Node *result = SafePointNode::Ideal(phase, can_reshape);
2173 if (result != NULL) return result;
2174 // Don't bother trying to transform a dead node
2175 if (in(0) && in(0)->is_top()) return NULL;
2176
2177 // Now see if we can optimize away this unlock. We don't actually
2178 // remove the unlocking here, we simply set the _eliminate flag which
2179 // prevents macro expansion from expanding the unlock. Since we don't
2180 // modify the graph, the value returned from this function is the
2181 // one computed above.
2182 // Escape state is defined after Parse phase.
2183 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
2184 //
2185 // If we are unlocking an non-escaped object, the lock/unlock is unnecessary.
2186 //
2187 ConnectionGraph *cgr = phase->C->congraph();
2188 if (cgr != NULL && cgr->not_global_escape(obj_node())) {
2189 assert(!is_eliminated() || is_coarsened(), "sanity");
2190 // The lock could be marked eliminated by lock coarsening
2191 // code during first IGVN before EA. Replace coarsened flag
2192 // to eliminate all associated locks/unlocks.
2193 #ifdef ASSERT
2194 this->log_lock_optimization(phase->C, "eliminate_lock_set_non_esc2");
2195 #endif
2196 this->set_non_esc_obj();
2197 }
2198 }
2199 return result;
2200 }
2201
2202 void AbstractLockNode::log_lock_optimization(Compile *C, const char * tag, Node* bad_lock) const {
2203 if (C == NULL) {
2204 return;
2205 }
2206 CompileLog* log = C->log();
2207 if (log != NULL) {
2208 Node* box = box_node();
2209 Node* obj = obj_node();
2210 int box_id = box != NULL ? box->_idx : -1;
2211 int obj_id = obj != NULL ? obj->_idx : -1;
2212
2213 log->begin_head("%s compile_id='%d' lock_id='%d' class='%s' kind='%s' box_id='%d' obj_id='%d' bad_id='%d'",
2214 tag, C->compile_id(), this->_idx,
2215 is_Unlock() ? "unlock" : is_Lock() ? "lock" : "?",
2216 kind_as_string(), box_id, obj_id, (bad_lock != NULL ? bad_lock->_idx : -1));
2217 log->stamp();
2218 log->end_head();
2219 JVMState* p = is_Unlock() ? (as_Unlock()->dbg_jvms()) : jvms();
2220 while (p != NULL) {
2221 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
2222 p = p->caller();
2223 }
2224 log->tail(tag);
2225 }
2226 }
2227
2228 bool CallNode::may_modify_arraycopy_helper(const TypeOopPtr* dest_t, const TypeOopPtr *t_oop, PhaseTransform *phase) {
2229 if (dest_t->is_known_instance() && t_oop->is_known_instance()) {
2230 return dest_t->instance_id() == t_oop->instance_id();
2231 }
2232
2233 if (dest_t->isa_instptr() && !dest_t->is_instptr()->instance_klass()->equals(phase->C->env()->Object_klass())) {
2234 // clone
2235 if (t_oop->isa_aryptr()) {
2236 return false;
2237 }
2238 if (!t_oop->isa_instptr()) {
2239 return true;
2240 }
2241 if (dest_t->maybe_java_subtype_of(t_oop) || t_oop->maybe_java_subtype_of(dest_t)) {
2242 return true;
2243 }
2244 // unrelated
2245 return false;
2246 }
2247
2248 if (dest_t->isa_aryptr()) {
2249 // arraycopy or array clone
2250 if (t_oop->isa_instptr()) {
2251 return false;
2252 }
2253 if (!t_oop->isa_aryptr()) {
2254 return true;
2255 }
2256
2257 const Type* elem = dest_t->is_aryptr()->elem();
2258 if (elem == Type::BOTTOM) {
2259 // An array but we don't know what elements are
2260 return true;
2261 }
2262
2263 dest_t = dest_t->add_offset(Type::OffsetBot)->is_oopptr();
2264 uint dest_alias = phase->C->get_alias_index(dest_t);
2265 uint t_oop_alias = phase->C->get_alias_index(t_oop);
2266
2267 return dest_alias == t_oop_alias;
2268 }
2269
2270 return true;
2271 }