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) const {
196 // Dump the required inputs, enclosed in '(' and ')'
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 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
201 else st->print("_ ");
202 }
203 }
204 #endif
205
206 //=============================================================================
207 RethrowNode::RethrowNode(
208 Node* cntrl,
209 Node* i_o,
210 Node* memory,
211 Node* frameptr,
212 Node* ret_adr,
213 Node* exception
214 ) : Node(TypeFunc::Parms + 1) {
215 init_req(TypeFunc::Control , cntrl );
216 init_req(TypeFunc::I_O , i_o );
217 init_req(TypeFunc::Memory , memory );
218 init_req(TypeFunc::FramePtr , frameptr );
219 init_req(TypeFunc::ReturnAdr, ret_adr);
220 init_req(TypeFunc::Parms , exception);
221 }
222
223 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){
224 return remove_dead_region(phase, can_reshape) ? this : NULL;
225 }
226
227 const Type* RethrowNode::Value(PhaseGVN* phase) const {
228 return (phase->type(in(TypeFunc::Control)) == Type::TOP)
229 ? Type::TOP
230 : Type::BOTTOM;
231 }
232
233 uint RethrowNode::match_edge(uint idx) const {
234 return 0;
235 }
236
237 #ifndef PRODUCT
238 void RethrowNode::dump_req(outputStream *st) const {
239 // Dump the required inputs, enclosed in '(' and ')'
240 uint i; // Exit value of loop
241 for (i = 0; i < req(); i++) { // For all required inputs
242 if (i == TypeFunc::Parms) st->print("exception");
243 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
244 else st->print("_ ");
245 }
246 }
247 #endif
248
249 //=============================================================================
250 // Do we Match on this edge index or not? Match only target address & method
251 uint TailCallNode::match_edge(uint idx) const {
252 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
253 }
254
255 //=============================================================================
256 // Do we Match on this edge index or not? Match only target address & oop
257 uint TailJumpNode::match_edge(uint idx) const {
258 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
259 }
260
261 //=============================================================================
262 JVMState::JVMState(ciMethod* method, JVMState* caller) :
263 _method(method) {
264 assert(method != NULL, "must be valid call site");
265 _bci = InvocationEntryBci;
266 _reexecute = Reexecute_Undefined;
267 debug_only(_bci = -99); // random garbage value
268 debug_only(_map = (SafePointNode*)-1);
269 _caller = caller;
270 _depth = 1 + (caller == NULL ? 0 : caller->depth());
271 _locoff = TypeFunc::Parms;
272 _stkoff = _locoff + _method->max_locals();
273 _monoff = _stkoff + _method->max_stack();
274 _scloff = _monoff;
275 _endoff = _monoff;
276 _sp = 0;
277 }
278 JVMState::JVMState(int stack_size) :
279 _method(NULL) {
280 _bci = InvocationEntryBci;
281 _reexecute = Reexecute_Undefined;
282 debug_only(_map = (SafePointNode*)-1);
283 _caller = NULL;
284 _depth = 1;
285 _locoff = TypeFunc::Parms;
286 _stkoff = _locoff;
287 _monoff = _stkoff + stack_size;
288 _scloff = _monoff;
289 _endoff = _monoff;
290 _sp = 0;
291 }
292
293 //--------------------------------of_depth-------------------------------------
294 JVMState* JVMState::of_depth(int d) const {
295 const JVMState* jvmp = this;
296 assert(0 < d && (uint)d <= depth(), "oob");
297 for (int skip = depth() - d; skip > 0; skip--) {
298 jvmp = jvmp->caller();
299 }
300 assert(jvmp->depth() == (uint)d, "found the right one");
301 return (JVMState*)jvmp;
302 }
303
304 //-----------------------------same_calls_as-----------------------------------
305 bool JVMState::same_calls_as(const JVMState* that) const {
306 if (this == that) return true;
307 if (this->depth() != that->depth()) return false;
308 const JVMState* p = this;
309 const JVMState* q = that;
310 for (;;) {
311 if (p->_method != q->_method) return false;
312 if (p->_method == NULL) return true; // bci is irrelevant
313 if (p->_bci != q->_bci) return false;
314 if (p->_reexecute != q->_reexecute) return false;
315 p = p->caller();
316 q = q->caller();
317 if (p == q) return true;
318 assert(p != NULL && q != NULL, "depth check ensures we don't run off end");
319 }
320 }
321
322 //------------------------------debug_start------------------------------------
323 uint JVMState::debug_start() const {
324 debug_only(JVMState* jvmroot = of_depth(1));
325 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
326 return of_depth(1)->locoff();
327 }
328
329 //-------------------------------debug_end-------------------------------------
330 uint JVMState::debug_end() const {
331 debug_only(JVMState* jvmroot = of_depth(1));
332 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
333 return endoff();
334 }
335
336 //------------------------------debug_depth------------------------------------
337 uint JVMState::debug_depth() const {
338 uint total = 0;
339 for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) {
340 total += jvmp->debug_size();
341 }
342 return total;
343 }
344
345 #ifndef PRODUCT
346
347 //------------------------------format_helper----------------------------------
348 // Given an allocation (a Chaitin object) and a Node decide if the Node carries
349 // any defined value or not. If it does, print out the register or constant.
350 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
351 if (n == NULL) { st->print(" NULL"); return; }
352 if (n->is_SafePointScalarObject()) {
353 // Scalar replacement.
354 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
355 scobjs->append_if_missing(spobj);
356 int sco_n = scobjs->find(spobj);
357 assert(sco_n >= 0, "");
358 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
359 return;
360 }
361 if (regalloc->node_regs_max_index() > 0 &&
362 OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
363 char buf[50];
364 regalloc->dump_register(n,buf);
365 st->print(" %s%d]=%s",msg,i,buf);
366 } else { // No register, but might be constant
367 const Type *t = n->bottom_type();
368 switch (t->base()) {
369 case Type::Int:
370 st->print(" %s%d]=#" INT32_FORMAT,msg,i,t->is_int()->get_con());
371 break;
372 case Type::AnyPtr:
373 assert( t == TypePtr::NULL_PTR || n->in_dump(), "" );
374 st->print(" %s%d]=#NULL",msg,i);
375 break;
376 case Type::AryPtr:
377 case Type::InstPtr:
378 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->isa_oopptr()->const_oop()));
379 break;
380 case Type::KlassPtr:
381 case Type::AryKlassPtr:
382 case Type::InstKlassPtr:
383 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_klassptr()->klass()));
384 break;
385 case Type::MetadataPtr:
386 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_metadataptr()->metadata()));
387 break;
388 case Type::NarrowOop:
389 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_oopptr()->const_oop()));
390 break;
391 case Type::RawPtr:
392 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,p2i(t->is_rawptr()));
393 break;
394 case Type::DoubleCon:
395 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
396 break;
397 case Type::FloatCon:
398 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
399 break;
400 case Type::Long:
401 st->print(" %s%d]=#" INT64_FORMAT,msg,i,(int64_t)(t->is_long()->get_con()));
402 break;
403 case Type::Half:
404 case Type::Top:
405 st->print(" %s%d]=_",msg,i);
406 break;
407 default: ShouldNotReachHere();
408 }
409 }
410 }
411
412 //---------------------print_method_with_lineno--------------------------------
413 void JVMState::print_method_with_lineno(outputStream* st, bool show_name) const {
414 if (show_name) _method->print_short_name(st);
415
416 int lineno = _method->line_number_from_bci(_bci);
417 if (lineno != -1) {
418 st->print(" @ bci:%d (line %d)", _bci, lineno);
419 } else {
420 st->print(" @ bci:%d", _bci);
421 }
422 }
423
424 //------------------------------format-----------------------------------------
425 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
426 st->print(" #");
427 if (_method) {
428 print_method_with_lineno(st, true);
429 } else {
430 st->print_cr(" runtime stub ");
431 return;
432 }
433 if (n->is_MachSafePoint()) {
434 GrowableArray<SafePointScalarObjectNode*> scobjs;
435 MachSafePointNode *mcall = n->as_MachSafePoint();
436 uint i;
437 // Print locals
438 for (i = 0; i < (uint)loc_size(); i++)
439 format_helper(regalloc, st, mcall->local(this, i), "L[", i, &scobjs);
440 // Print stack
441 for (i = 0; i < (uint)stk_size(); i++) {
442 if ((uint)(_stkoff + i) >= mcall->len())
443 st->print(" oob ");
444 else
445 format_helper(regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs);
446 }
447 for (i = 0; (int)i < nof_monitors(); i++) {
448 Node *box = mcall->monitor_box(this, i);
449 Node *obj = mcall->monitor_obj(this, i);
450 if (regalloc->node_regs_max_index() > 0 &&
451 OptoReg::is_valid(regalloc->get_reg_first(box))) {
452 box = BoxLockNode::box_node(box);
453 format_helper(regalloc, st, box, "MON-BOX[", i, &scobjs);
454 } else {
455 OptoReg::Name box_reg = BoxLockNode::reg(box);
456 st->print(" MON-BOX%d=%s+%d",
457 i,
458 OptoReg::regname(OptoReg::c_frame_pointer),
459 regalloc->reg2offset(box_reg));
460 }
461 const char* obj_msg = "MON-OBJ[";
462 if (EliminateLocks) {
463 if (BoxLockNode::box_node(box)->is_eliminated())
464 obj_msg = "MON-OBJ(LOCK ELIMINATED)[";
465 }
466 format_helper(regalloc, st, obj, obj_msg, i, &scobjs);
467 }
468
469 for (i = 0; i < (uint)scobjs.length(); i++) {
470 // Scalar replaced objects.
471 st->cr();
472 st->print(" # ScObj" INT32_FORMAT " ", i);
473 SafePointScalarObjectNode* spobj = scobjs.at(i);
474 ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass();
475 assert(cik->is_instance_klass() ||
476 cik->is_array_klass(), "Not supported allocation.");
477 ciInstanceKlass *iklass = NULL;
478 if (cik->is_instance_klass()) {
479 cik->print_name_on(st);
480 iklass = cik->as_instance_klass();
481 } else if (cik->is_type_array_klass()) {
482 cik->as_array_klass()->base_element_type()->print_name_on(st);
483 st->print("[%d]", spobj->n_fields());
484 } else if (cik->is_obj_array_klass()) {
485 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass();
486 if (cie->is_instance_klass()) {
487 cie->print_name_on(st);
488 } else if (cie->is_type_array_klass()) {
489 cie->as_array_klass()->base_element_type()->print_name_on(st);
490 } else {
491 ShouldNotReachHere();
492 }
493 st->print("[%d]", spobj->n_fields());
494 int ndim = cik->as_array_klass()->dimension() - 1;
495 while (ndim-- > 0) {
496 st->print("[]");
497 }
498 }
499 st->print("={");
500 uint nf = spobj->n_fields();
501 if (nf > 0) {
502 uint first_ind = spobj->first_index(mcall->jvms());
503 Node* fld_node = mcall->in(first_ind);
504 ciField* cifield;
505 if (iklass != NULL) {
506 st->print(" [");
507 cifield = iklass->nonstatic_field_at(0);
508 cifield->print_name_on(st);
509 format_helper(regalloc, st, fld_node, ":", 0, &scobjs);
510 } else {
511 format_helper(regalloc, st, fld_node, "[", 0, &scobjs);
512 }
513 for (uint j = 1; j < nf; j++) {
514 fld_node = mcall->in(first_ind+j);
515 if (iklass != NULL) {
516 st->print(", [");
517 cifield = iklass->nonstatic_field_at(j);
518 cifield->print_name_on(st);
519 format_helper(regalloc, st, fld_node, ":", j, &scobjs);
520 } else {
521 format_helper(regalloc, st, fld_node, ", [", j, &scobjs);
522 }
523 }
524 }
525 st->print(" }");
526 }
527 }
528 st->cr();
529 if (caller() != NULL) caller()->format(regalloc, n, st);
530 }
531
532
533 void JVMState::dump_spec(outputStream *st) const {
534 if (_method != NULL) {
535 bool printed = false;
536 if (!Verbose) {
537 // The JVMS dumps make really, really long lines.
538 // Take out the most boring parts, which are the package prefixes.
539 char buf[500];
540 stringStream namest(buf, sizeof(buf));
541 _method->print_short_name(&namest);
542 if (namest.count() < sizeof(buf)) {
543 const char* name = namest.base();
544 if (name[0] == ' ') ++name;
545 const char* endcn = strchr(name, ':'); // end of class name
546 if (endcn == NULL) endcn = strchr(name, '(');
547 if (endcn == NULL) endcn = name + strlen(name);
548 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
549 --endcn;
550 st->print(" %s", endcn);
551 printed = true;
552 }
553 }
554 print_method_with_lineno(st, !printed);
555 if(_reexecute == Reexecute_True)
556 st->print(" reexecute");
557 } else {
558 st->print(" runtime stub");
559 }
560 if (caller() != NULL) caller()->dump_spec(st);
561 }
562
563
564 void JVMState::dump_on(outputStream* st) const {
565 bool print_map = _map && !((uintptr_t)_map & 1) &&
566 ((caller() == NULL) || (caller()->map() != _map));
567 if (print_map) {
568 if (_map->len() > _map->req()) { // _map->has_exceptions()
569 Node* ex = _map->in(_map->req()); // _map->next_exception()
570 // skip the first one; it's already being printed
571 while (ex != NULL && ex->len() > ex->req()) {
572 ex = ex->in(ex->req()); // ex->next_exception()
573 ex->dump(1);
574 }
575 }
576 _map->dump(Verbose ? 2 : 1);
577 }
578 if (caller() != NULL) {
579 caller()->dump_on(st);
580 }
581 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=",
582 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false");
583 if (_method == NULL) {
584 st->print_cr("(none)");
585 } else {
586 _method->print_name(st);
587 st->cr();
588 if (bci() >= 0 && bci() < _method->code_size()) {
589 st->print(" bc: ");
590 _method->print_codes_on(bci(), bci()+1, st);
591 }
592 }
593 }
594
595 // Extra way to dump a jvms from the debugger,
596 // to avoid a bug with C++ member function calls.
597 void dump_jvms(JVMState* jvms) {
598 jvms->dump();
599 }
600 #endif
601
602 //--------------------------clone_shallow--------------------------------------
603 JVMState* JVMState::clone_shallow(Compile* C) const {
604 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
605 n->set_bci(_bci);
606 n->_reexecute = _reexecute;
607 n->set_locoff(_locoff);
608 n->set_stkoff(_stkoff);
609 n->set_monoff(_monoff);
610 n->set_scloff(_scloff);
611 n->set_endoff(_endoff);
612 n->set_sp(_sp);
613 n->set_map(_map);
614 return n;
615 }
616
617 //---------------------------clone_deep----------------------------------------
618 JVMState* JVMState::clone_deep(Compile* C) const {
619 JVMState* n = clone_shallow(C);
620 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) {
621 p->_caller = p->_caller->clone_shallow(C);
622 }
623 assert(n->depth() == depth(), "sanity");
624 assert(n->debug_depth() == debug_depth(), "sanity");
625 return n;
626 }
627
628 /**
629 * Reset map for all callers
630 */
631 void JVMState::set_map_deep(SafePointNode* map) {
632 for (JVMState* p = this; p != NULL; p = p->_caller) {
633 p->set_map(map);
634 }
635 }
636
637 // unlike set_map(), this is two-way setting.
638 void JVMState::bind_map(SafePointNode* map) {
639 set_map(map);
640 _map->set_jvms(this);
641 }
642
643 // Adapt offsets in in-array after adding or removing an edge.
644 // Prerequisite is that the JVMState is used by only one node.
645 void JVMState::adapt_position(int delta) {
646 for (JVMState* jvms = this; jvms != NULL; jvms = jvms->caller()) {
647 jvms->set_locoff(jvms->locoff() + delta);
648 jvms->set_stkoff(jvms->stkoff() + delta);
649 jvms->set_monoff(jvms->monoff() + delta);
650 jvms->set_scloff(jvms->scloff() + delta);
651 jvms->set_endoff(jvms->endoff() + delta);
652 }
653 }
654
655 // Mirror the stack size calculation in the deopt code
656 // How much stack space would we need at this point in the program in
657 // case of deoptimization?
658 int JVMState::interpreter_frame_size() const {
659 const JVMState* jvms = this;
660 int size = 0;
661 int callee_parameters = 0;
662 int callee_locals = 0;
663 int extra_args = method()->max_stack() - stk_size();
664
665 while (jvms != NULL) {
666 int locks = jvms->nof_monitors();
667 int temps = jvms->stk_size();
668 bool is_top_frame = (jvms == this);
669 ciMethod* method = jvms->method();
670
671 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(),
672 temps + callee_parameters,
673 extra_args,
674 locks,
675 callee_parameters,
676 callee_locals,
677 is_top_frame);
678 size += frame_size;
679
680 callee_parameters = method->size_of_parameters();
681 callee_locals = method->max_locals();
682 extra_args = 0;
683 jvms = jvms->caller();
684 }
685 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord;
686 }
687
688 //=============================================================================
689 bool CallNode::cmp( const Node &n ) const
690 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
691 #ifndef PRODUCT
692 void CallNode::dump_req(outputStream *st) const {
693 // Dump the required inputs, enclosed in '(' and ')'
694 uint i; // Exit value of loop
695 for (i = 0; i < req(); i++) { // For all required inputs
696 if (i == TypeFunc::Parms) st->print("(");
697 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
698 else st->print("_ ");
699 }
700 st->print(")");
701 }
702
703 void CallNode::dump_spec(outputStream *st) const {
704 st->print(" ");
705 if (tf() != NULL) tf()->dump_on(st);
706 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt);
707 if (jvms() != NULL) jvms()->dump_spec(st);
708 }
709 #endif
710
711 const Type *CallNode::bottom_type() const { return tf()->range(); }
712 const Type* CallNode::Value(PhaseGVN* phase) const {
713 if (phase->type(in(0)) == Type::TOP) return Type::TOP;
714 return tf()->range();
715 }
716
717 //------------------------------calling_convention-----------------------------
718 void CallNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
719 // Use the standard compiler calling convention
720 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt);
721 }
722
723
724 //------------------------------match------------------------------------------
725 // Construct projections for control, I/O, memory-fields, ..., and
726 // return result(s) along with their RegMask info
727 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
728 switch (proj->_con) {
729 case TypeFunc::Control:
730 case TypeFunc::I_O:
731 case TypeFunc::Memory:
732 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
733
734 case TypeFunc::Parms+1: // For LONG & DOUBLE returns
735 assert(tf()->range()->field_at(TypeFunc::Parms+1) == Type::HALF, "");
736 // 2nd half of doubles and longs
737 return new MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad);
738
739 case TypeFunc::Parms: { // Normal returns
740 uint ideal_reg = tf()->range()->field_at(TypeFunc::Parms)->ideal_reg();
741 OptoRegPair regs = Opcode() == Op_CallLeafVector
742 ? match->vector_return_value(ideal_reg) // Calls into assembly vector routine
743 : is_CallRuntime()
744 ? match->c_return_value(ideal_reg) // Calls into C runtime
745 : match-> return_value(ideal_reg); // Calls into compiled Java code
746 RegMask rm = RegMask(regs.first());
747
748 if (Opcode() == Op_CallLeafVector) {
749 // If the return is in vector, compute appropriate regmask taking into account the whole range
750 if(ideal_reg >= Op_VecS && ideal_reg <= Op_VecZ) {
751 if(OptoReg::is_valid(regs.second())) {
752 for (OptoReg::Name r = regs.first(); r <= regs.second(); r = OptoReg::add(r, 1)) {
753 rm.Insert(r);
754 }
755 }
756 }
757 }
758
759 if( OptoReg::is_valid(regs.second()) )
760 rm.Insert( regs.second() );
761 return new MachProjNode(this,proj->_con,rm,ideal_reg);
762 }
763
764 case TypeFunc::ReturnAdr:
765 case TypeFunc::FramePtr:
766 default:
767 ShouldNotReachHere();
768 }
769 return NULL;
770 }
771
772 // Do we Match on this edge index or not? Match no edges
773 uint CallNode::match_edge(uint idx) const {
774 return 0;
775 }
776
777 //
778 // Determine whether the call could modify the field of the specified
779 // instance at the specified offset.
780 //
781 bool CallNode::may_modify(const TypeOopPtr *t_oop, PhaseTransform *phase) {
782 assert((t_oop != NULL), "sanity");
783 if (is_call_to_arraycopystub() && strcmp(_name, "unsafe_arraycopy") != 0) {
784 const TypeTuple* args = _tf->domain();
785 Node* dest = NULL;
786 // Stubs that can be called once an ArrayCopyNode is expanded have
787 // different signatures. Look for the second pointer argument,
788 // that is the destination of the copy.
789 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
790 if (args->field_at(i)->isa_ptr()) {
791 j++;
792 if (j == 2) {
793 dest = in(i);
794 break;
795 }
796 }
797 }
798 guarantee(dest != NULL, "Call had only one ptr in, broken IR!");
799 if (!dest->is_top() && may_modify_arraycopy_helper(phase->type(dest)->is_oopptr(), t_oop, phase)) {
800 return true;
801 }
802 return false;
803 }
804 if (t_oop->is_known_instance()) {
805 // The instance_id is set only for scalar-replaceable allocations which
806 // are not passed as arguments according to Escape Analysis.
807 return false;
808 }
809 if (t_oop->is_ptr_to_boxed_value()) {
810 ciKlass* boxing_klass = t_oop->klass();
811 if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) {
812 // Skip unrelated boxing methods.
813 Node* proj = proj_out_or_null(TypeFunc::Parms);
814 if ((proj == NULL) || (phase->type(proj)->is_instptr()->klass() != boxing_klass)) {
815 return false;
816 }
817 }
818 if (is_CallJava() && as_CallJava()->method() != NULL) {
819 ciMethod* meth = as_CallJava()->method();
820 if (meth->is_getter()) {
821 return false;
822 }
823 // May modify (by reflection) if an boxing object is passed
824 // as argument or returned.
825 Node* proj = returns_pointer() ? proj_out_or_null(TypeFunc::Parms) : NULL;
826 if (proj != NULL) {
827 const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr();
828 if ((inst_t != NULL) && (!inst_t->klass_is_exact() ||
829 (inst_t->klass() == boxing_klass))) {
830 return true;
831 }
832 }
833 const TypeTuple* d = tf()->domain();
834 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
835 const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr();
836 if ((inst_t != NULL) && (!inst_t->klass_is_exact() ||
837 (inst_t->klass() == boxing_klass))) {
838 return true;
839 }
840 }
841 return false;
842 }
843 }
844 return true;
845 }
846
847 // Does this call have a direct reference to n other than debug information?
848 bool CallNode::has_non_debug_use(Node *n) {
849 const TypeTuple * d = tf()->domain();
850 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
851 Node *arg = in(i);
852 if (arg == n) {
853 return true;
854 }
855 }
856 return false;
857 }
858
859 // Returns the unique CheckCastPP of a call
860 // or 'this' if there are several CheckCastPP or unexpected uses
861 // or returns NULL if there is no one.
862 Node *CallNode::result_cast() {
863 Node *cast = NULL;
864
865 Node *p = proj_out_or_null(TypeFunc::Parms);
866 if (p == NULL)
867 return NULL;
868
869 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
870 Node *use = p->fast_out(i);
871 if (use->is_CheckCastPP()) {
872 if (cast != NULL) {
873 return this; // more than 1 CheckCastPP
874 }
875 cast = use;
876 } else if (!use->is_Initialize() &&
877 !use->is_AddP() &&
878 use->Opcode() != Op_MemBarStoreStore) {
879 // Expected uses are restricted to a CheckCastPP, an Initialize
880 // node, a MemBarStoreStore (clone) and AddP nodes. If we
881 // encounter any other use (a Phi node can be seen in rare
882 // cases) return this to prevent incorrect optimizations.
883 return this;
884 }
885 }
886 return cast;
887 }
888
889
890 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts) {
891 projs->fallthrough_proj = NULL;
892 projs->fallthrough_catchproj = NULL;
893 projs->fallthrough_ioproj = NULL;
894 projs->catchall_ioproj = NULL;
895 projs->catchall_catchproj = NULL;
896 projs->fallthrough_memproj = NULL;
897 projs->catchall_memproj = NULL;
898 projs->resproj = NULL;
899 projs->exobj = NULL;
900
901 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
902 ProjNode *pn = fast_out(i)->as_Proj();
903 if (pn->outcnt() == 0) continue;
904 switch (pn->_con) {
905 case TypeFunc::Control:
906 {
907 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
908 projs->fallthrough_proj = pn;
909 const Node* cn = pn->unique_ctrl_out_or_null();
910 if (cn != NULL && cn->is_Catch()) {
911 ProjNode *cpn = NULL;
912 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
913 cpn = cn->fast_out(k)->as_Proj();
914 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
915 if (cpn->_con == CatchProjNode::fall_through_index)
916 projs->fallthrough_catchproj = cpn;
917 else {
918 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
919 projs->catchall_catchproj = cpn;
920 }
921 }
922 }
923 break;
924 }
925 case TypeFunc::I_O:
926 if (pn->_is_io_use)
927 projs->catchall_ioproj = pn;
928 else
929 projs->fallthrough_ioproj = pn;
930 for (DUIterator j = pn->outs(); pn->has_out(j); j++) {
931 Node* e = pn->out(j);
932 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) {
933 assert(projs->exobj == NULL, "only one");
934 projs->exobj = e;
935 }
936 }
937 break;
938 case TypeFunc::Memory:
939 if (pn->_is_io_use)
940 projs->catchall_memproj = pn;
941 else
942 projs->fallthrough_memproj = pn;
943 break;
944 case TypeFunc::Parms:
945 projs->resproj = pn;
946 break;
947 default:
948 assert(false, "unexpected projection from allocation node.");
949 }
950 }
951
952 // The resproj may not exist because the result could be ignored
953 // and the exception object may not exist if an exception handler
954 // swallows the exception but all the other must exist and be found.
955 assert(projs->fallthrough_proj != NULL, "must be found");
956 do_asserts = do_asserts && !Compile::current()->inlining_incrementally();
957 assert(!do_asserts || projs->fallthrough_catchproj != NULL, "must be found");
958 assert(!do_asserts || projs->fallthrough_memproj != NULL, "must be found");
959 assert(!do_asserts || projs->fallthrough_ioproj != NULL, "must be found");
960 assert(!do_asserts || projs->catchall_catchproj != NULL, "must be found");
961 if (separate_io_proj) {
962 assert(!do_asserts || projs->catchall_memproj != NULL, "must be found");
963 assert(!do_asserts || projs->catchall_ioproj != NULL, "must be found");
964 }
965 }
966
967 Node* CallNode::Ideal(PhaseGVN* phase, bool can_reshape) {
968 #ifdef ASSERT
969 // Validate attached generator
970 CallGenerator* cg = generator();
971 if (cg != NULL) {
972 assert(is_CallStaticJava() && cg->is_mh_late_inline() ||
973 is_CallDynamicJava() && cg->is_virtual_late_inline(), "mismatch");
974 }
975 #endif // ASSERT
976 return SafePointNode::Ideal(phase, can_reshape);
977 }
978
979 bool CallNode::is_call_to_arraycopystub() const {
980 if (_name != NULL && strstr(_name, "arraycopy") != 0) {
981 return true;
982 }
983 return false;
984 }
985
986 //=============================================================================
987 uint CallJavaNode::size_of() const { return sizeof(*this); }
988 bool CallJavaNode::cmp( const Node &n ) const {
989 CallJavaNode &call = (CallJavaNode&)n;
990 return CallNode::cmp(call) && _method == call._method &&
991 _override_symbolic_info == call._override_symbolic_info;
992 }
993
994 void CallJavaNode::copy_call_debug_info(PhaseIterGVN* phase, SafePointNode* sfpt) {
995 // Copy debug information and adjust JVMState information
996 uint old_dbg_start = sfpt->is_Call() ? sfpt->as_Call()->tf()->domain()->cnt() : (uint)TypeFunc::Parms+1;
997 uint new_dbg_start = tf()->domain()->cnt();
998 int jvms_adj = new_dbg_start - old_dbg_start;
999 assert (new_dbg_start == req(), "argument count mismatch");
1000 Compile* C = phase->C;
1001
1002 // SafePointScalarObject node could be referenced several times in debug info.
1003 // Use Dict to record cloned nodes.
1004 Dict* sosn_map = new Dict(cmpkey,hashkey);
1005 for (uint i = old_dbg_start; i < sfpt->req(); i++) {
1006 Node* old_in = sfpt->in(i);
1007 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
1008 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
1009 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
1010 bool new_node;
1011 Node* new_in = old_sosn->clone(sosn_map, new_node);
1012 if (new_node) { // New node?
1013 new_in->set_req(0, C->root()); // reset control edge
1014 new_in = phase->transform(new_in); // Register new node.
1015 }
1016 old_in = new_in;
1017 }
1018 add_req(old_in);
1019 }
1020
1021 // JVMS may be shared so clone it before we modify it
1022 set_jvms(sfpt->jvms() != NULL ? sfpt->jvms()->clone_deep(C) : NULL);
1023 for (JVMState *jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
1024 jvms->set_map(this);
1025 jvms->set_locoff(jvms->locoff()+jvms_adj);
1026 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
1027 jvms->set_monoff(jvms->monoff()+jvms_adj);
1028 jvms->set_scloff(jvms->scloff()+jvms_adj);
1029 jvms->set_endoff(jvms->endoff()+jvms_adj);
1030 }
1031 }
1032
1033 #ifdef ASSERT
1034 bool CallJavaNode::validate_symbolic_info() const {
1035 if (method() == NULL) {
1036 return true; // call into runtime or uncommon trap
1037 }
1038 ciMethod* symbolic_info = jvms()->method()->get_method_at_bci(jvms()->bci());
1039 ciMethod* callee = method();
1040 if (symbolic_info->is_method_handle_intrinsic() && !callee->is_method_handle_intrinsic()) {
1041 assert(override_symbolic_info(), "should be set");
1042 }
1043 assert(ciMethod::is_consistent_info(symbolic_info, callee), "inconsistent info");
1044 return true;
1045 }
1046 #endif
1047
1048 #ifndef PRODUCT
1049 void CallJavaNode::dump_spec(outputStream* st) const {
1050 if( _method ) _method->print_short_name(st);
1051 CallNode::dump_spec(st);
1052 }
1053
1054 void CallJavaNode::dump_compact_spec(outputStream* st) const {
1055 if (_method) {
1056 _method->print_short_name(st);
1057 } else {
1058 st->print("<?>");
1059 }
1060 }
1061 #endif
1062
1063 //=============================================================================
1064 uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
1065 bool CallStaticJavaNode::cmp( const Node &n ) const {
1066 CallStaticJavaNode &call = (CallStaticJavaNode&)n;
1067 return CallJavaNode::cmp(call);
1068 }
1069
1070 Node* CallStaticJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1071 CallGenerator* cg = generator();
1072 if (can_reshape && cg != NULL) {
1073 assert(IncrementalInlineMH, "required");
1074 assert(cg->call_node() == this, "mismatch");
1075 assert(cg->is_mh_late_inline(), "not virtual");
1076
1077 // Check whether this MH handle call becomes a candidate for inlining.
1078 ciMethod* callee = cg->method();
1079 vmIntrinsics::ID iid = callee->intrinsic_id();
1080 if (iid == vmIntrinsics::_invokeBasic) {
1081 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) {
1082 phase->C->prepend_late_inline(cg);
1083 set_generator(NULL);
1084 }
1085 } else if (iid == vmIntrinsics::_linkToNative) {
1086 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP /* NEP */
1087 && in(TypeFunc::Parms + 1)->Opcode() == Op_ConL /* address */) {
1088 phase->C->prepend_late_inline(cg);
1089 set_generator(NULL);
1090 }
1091 } else {
1092 assert(callee->has_member_arg(), "wrong type of call?");
1093 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) {
1094 phase->C->prepend_late_inline(cg);
1095 set_generator(NULL);
1096 }
1097 }
1098 }
1099 return CallNode::Ideal(phase, can_reshape);
1100 }
1101
1102 //----------------------------uncommon_trap_request----------------------------
1103 // If this is an uncommon trap, return the request code, else zero.
1104 int CallStaticJavaNode::uncommon_trap_request() const {
1105 if (_name != NULL && !strcmp(_name, "uncommon_trap")) {
1106 return extract_uncommon_trap_request(this);
1107 }
1108 return 0;
1109 }
1110 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
1111 #ifndef PRODUCT
1112 if (!(call->req() > TypeFunc::Parms &&
1113 call->in(TypeFunc::Parms) != NULL &&
1114 call->in(TypeFunc::Parms)->is_Con() &&
1115 call->in(TypeFunc::Parms)->bottom_type()->isa_int())) {
1116 assert(in_dump() != 0, "OK if dumping");
1117 tty->print("[bad uncommon trap]");
1118 return 0;
1119 }
1120 #endif
1121 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
1122 }
1123
1124 #ifndef PRODUCT
1125 void CallStaticJavaNode::dump_spec(outputStream *st) const {
1126 st->print("# Static ");
1127 if (_name != NULL) {
1128 st->print("%s", _name);
1129 int trap_req = uncommon_trap_request();
1130 if (trap_req != 0) {
1131 char buf[100];
1132 st->print("(%s)",
1133 Deoptimization::format_trap_request(buf, sizeof(buf),
1134 trap_req));
1135 }
1136 st->print(" ");
1137 }
1138 CallJavaNode::dump_spec(st);
1139 }
1140
1141 void CallStaticJavaNode::dump_compact_spec(outputStream* st) const {
1142 if (_method) {
1143 _method->print_short_name(st);
1144 } else if (_name) {
1145 st->print("%s", _name);
1146 } else {
1147 st->print("<?>");
1148 }
1149 }
1150 #endif
1151
1152 //=============================================================================
1153 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
1154 bool CallDynamicJavaNode::cmp( const Node &n ) const {
1155 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
1156 return CallJavaNode::cmp(call);
1157 }
1158
1159 Node* CallDynamicJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1160 CallGenerator* cg = generator();
1161 if (can_reshape && cg != NULL) {
1162 assert(IncrementalInlineVirtual, "required");
1163 assert(cg->call_node() == this, "mismatch");
1164 assert(cg->is_virtual_late_inline(), "not virtual");
1165
1166 // Recover symbolic info for method resolution.
1167 ciMethod* caller = jvms()->method();
1168 ciBytecodeStream iter(caller);
1169 iter.force_bci(jvms()->bci());
1170
1171 bool not_used1;
1172 ciSignature* not_used2;
1173 ciMethod* orig_callee = iter.get_method(not_used1, ¬_used2); // callee in the bytecode
1174 ciKlass* holder = iter.get_declared_method_holder();
1175 if (orig_callee->is_method_handle_intrinsic()) {
1176 assert(_override_symbolic_info, "required");
1177 orig_callee = method();
1178 holder = method()->holder();
1179 }
1180
1181 ciInstanceKlass* klass = ciEnv::get_instance_klass_for_declared_method_holder(holder);
1182
1183 Node* receiver_node = in(TypeFunc::Parms);
1184 const TypeOopPtr* receiver_type = phase->type(receiver_node)->isa_oopptr();
1185
1186 int not_used3;
1187 bool call_does_dispatch;
1188 ciMethod* callee = phase->C->optimize_virtual_call(caller, klass, holder, orig_callee, receiver_type, true /*is_virtual*/,
1189 call_does_dispatch, not_used3); // out-parameters
1190 if (!call_does_dispatch) {
1191 // Register for late inlining.
1192 cg->set_callee_method(callee);
1193 phase->C->prepend_late_inline(cg); // MH late inlining prepends to the list, so do the same
1194 set_generator(NULL);
1195 }
1196 }
1197 return CallNode::Ideal(phase, can_reshape);
1198 }
1199
1200 #ifndef PRODUCT
1201 void CallDynamicJavaNode::dump_spec(outputStream *st) const {
1202 st->print("# Dynamic ");
1203 CallJavaNode::dump_spec(st);
1204 }
1205 #endif
1206
1207 //=============================================================================
1208 uint CallRuntimeNode::size_of() const { return sizeof(*this); }
1209 bool CallRuntimeNode::cmp( const Node &n ) const {
1210 CallRuntimeNode &call = (CallRuntimeNode&)n;
1211 return CallNode::cmp(call) && !strcmp(_name,call._name);
1212 }
1213 #ifndef PRODUCT
1214 void CallRuntimeNode::dump_spec(outputStream *st) const {
1215 st->print("# ");
1216 st->print("%s", _name);
1217 CallNode::dump_spec(st);
1218 }
1219 #endif
1220 uint CallLeafVectorNode::size_of() const { return sizeof(*this); }
1221 bool CallLeafVectorNode::cmp( const Node &n ) const {
1222 CallLeafVectorNode &call = (CallLeafVectorNode&)n;
1223 return CallLeafNode::cmp(call) && _num_bits == call._num_bits;
1224 }
1225
1226 //=============================================================================
1227 uint CallNativeNode::size_of() const { return sizeof(*this); }
1228 bool CallNativeNode::cmp( const Node &n ) const {
1229 CallNativeNode &call = (CallNativeNode&)n;
1230 return CallNode::cmp(call) && !strcmp(_name,call._name)
1231 && _arg_regs == call._arg_regs && _ret_regs == call._ret_regs;
1232 }
1233 Node* CallNativeNode::match(const ProjNode *proj, const Matcher *matcher) {
1234 switch (proj->_con) {
1235 case TypeFunc::Control:
1236 case TypeFunc::I_O:
1237 case TypeFunc::Memory:
1238 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
1239 case TypeFunc::ReturnAdr:
1240 case TypeFunc::FramePtr:
1241 ShouldNotReachHere();
1242 case TypeFunc::Parms: {
1243 const Type* field_at_con = tf()->range()->field_at(proj->_con);
1244 const BasicType bt = field_at_con->basic_type();
1245 OptoReg::Name optoreg = OptoReg::as_OptoReg(_ret_regs.at(proj->_con - TypeFunc::Parms));
1246 OptoRegPair regs;
1247 if (bt == T_DOUBLE || bt == T_LONG) {
1248 regs.set2(optoreg);
1249 } else {
1250 regs.set1(optoreg);
1251 }
1252 RegMask rm = RegMask(regs.first());
1253 if(OptoReg::is_valid(regs.second()))
1254 rm.Insert(regs.second());
1255 return new MachProjNode(this, proj->_con, rm, field_at_con->ideal_reg());
1256 }
1257 case TypeFunc::Parms + 1: {
1258 assert(tf()->range()->field_at(proj->_con) == Type::HALF, "Expected HALF");
1259 assert(_ret_regs.at(proj->_con - TypeFunc::Parms) == VMRegImpl::Bad(), "Unexpected register for Type::HALF");
1260 // 2nd half of doubles and longs
1261 return new MachProjNode(this, proj->_con, RegMask::Empty, (uint) OptoReg::Bad);
1262 }
1263 default:
1264 ShouldNotReachHere();
1265 }
1266 return NULL;
1267 }
1268 #ifndef PRODUCT
1269 void CallNativeNode::print_regs(const GrowableArray<VMReg>& regs, outputStream* st) {
1270 st->print("{ ");
1271 for (int i = 0; i < regs.length(); i++) {
1272 regs.at(i)->print_on(st);
1273 if (i < regs.length() - 1) {
1274 st->print(", ");
1275 }
1276 }
1277 st->print(" } ");
1278 }
1279
1280 void CallNativeNode::dump_spec(outputStream *st) const {
1281 st->print("# ");
1282 st->print("%s ", _name);
1283 st->print("_arg_regs: ");
1284 print_regs(_arg_regs, st);
1285 st->print("_ret_regs: ");
1286 print_regs(_ret_regs, st);
1287 CallNode::dump_spec(st);
1288 }
1289 #endif
1290
1291 //------------------------------calling_convention-----------------------------
1292 void CallRuntimeNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const {
1293 SharedRuntime::c_calling_convention(sig_bt, parm_regs, /*regs2=*/nullptr, argcnt);
1294 }
1295
1296 void CallLeafVectorNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
1297 #ifdef ASSERT
1298 assert(tf()->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1299 "return vector size must match");
1300 const TypeTuple* d = tf()->domain();
1301 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1302 Node* arg = in(i);
1303 assert(arg->bottom_type()->is_vect()->length_in_bytes() * BitsPerByte == _num_bits,
1304 "vector argument size must match");
1305 }
1306 #endif
1307
1308 SharedRuntime::vector_calling_convention(parm_regs, _num_bits, argcnt);
1309 }
1310
1311 void CallNativeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
1312 assert((tf()->domain()->cnt() - TypeFunc::Parms) == argcnt, "arg counts must match!");
1313 #ifdef ASSERT
1314 for (uint i = 0; i < argcnt; i++) {
1315 assert(tf()->domain()->field_at(TypeFunc::Parms + i)->basic_type() == sig_bt[i], "types must match!");
1316 }
1317 #endif
1318 for (uint i = 0; i < argcnt; i++) {
1319 switch (sig_bt[i]) {
1320 case T_BOOLEAN:
1321 case T_CHAR:
1322 case T_BYTE:
1323 case T_SHORT:
1324 case T_INT:
1325 case T_FLOAT:
1326 parm_regs[i].set1(_arg_regs.at(i));
1327 break;
1328 case T_LONG:
1329 case T_DOUBLE:
1330 assert((i + 1) < argcnt && sig_bt[i + 1] == T_VOID, "expecting half");
1331 parm_regs[i].set2(_arg_regs.at(i));
1332 break;
1333 case T_VOID: // Halves of longs and doubles
1334 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
1335 assert(_arg_regs.at(i) == VMRegImpl::Bad(), "expecting bad reg");
1336 parm_regs[i].set_bad();
1337 break;
1338 default:
1339 ShouldNotReachHere();
1340 break;
1341 }
1342 }
1343 }
1344
1345 //=============================================================================
1346 //------------------------------calling_convention-----------------------------
1347
1348
1349 //=============================================================================
1350 #ifndef PRODUCT
1351 void CallLeafNode::dump_spec(outputStream *st) const {
1352 st->print("# ");
1353 st->print("%s", _name);
1354 CallNode::dump_spec(st);
1355 }
1356 #endif
1357
1358 //=============================================================================
1359
1360 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) {
1361 assert(verify_jvms(jvms), "jvms must match");
1362 int loc = jvms->locoff() + idx;
1363 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
1364 // If current local idx is top then local idx - 1 could
1365 // be a long/double that needs to be killed since top could
1366 // represent the 2nd half of the long/double.
1367 uint ideal = in(loc -1)->ideal_reg();
1368 if (ideal == Op_RegD || ideal == Op_RegL) {
1369 // set other (low index) half to top
1370 set_req(loc - 1, in(loc));
1371 }
1372 }
1373 set_req(loc, c);
1374 }
1375
1376 uint SafePointNode::size_of() const { return sizeof(*this); }
1377 bool SafePointNode::cmp( const Node &n ) const {
1378 return (&n == this); // Always fail except on self
1379 }
1380
1381 //-------------------------set_next_exception----------------------------------
1382 void SafePointNode::set_next_exception(SafePointNode* n) {
1383 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception");
1384 if (len() == req()) {
1385 if (n != NULL) add_prec(n);
1386 } else {
1387 set_prec(req(), n);
1388 }
1389 }
1390
1391
1392 //----------------------------next_exception-----------------------------------
1393 SafePointNode* SafePointNode::next_exception() const {
1394 if (len() == req()) {
1395 return NULL;
1396 } else {
1397 Node* n = in(req());
1398 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
1399 return (SafePointNode*) n;
1400 }
1401 }
1402
1403
1404 //------------------------------Ideal------------------------------------------
1405 // Skip over any collapsed Regions
1406 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1407 assert(_jvms == NULL || ((uintptr_t)_jvms->map() & 1) || _jvms->map() == this, "inconsistent JVMState");
1408 return remove_dead_region(phase, can_reshape) ? this : NULL;
1409 }
1410
1411 //------------------------------Identity---------------------------------------
1412 // Remove obviously duplicate safepoints
1413 Node* SafePointNode::Identity(PhaseGVN* phase) {
1414
1415 // If you have back to back safepoints, remove one
1416 if (in(TypeFunc::Control)->is_SafePoint()) {
1417 Node* out_c = unique_ctrl_out_or_null();
1418 // This can be the safepoint of an outer strip mined loop if the inner loop's backedge was removed. Replacing the
1419 // outer loop's safepoint could confuse removal of the outer loop.
1420 if (out_c != NULL && !out_c->is_OuterStripMinedLoopEnd()) {
1421 return in(TypeFunc::Control);
1422 }
1423 }
1424
1425 // Transforming long counted loops requires a safepoint node. Do not
1426 // eliminate a safepoint until loop opts are over.
1427 if (in(0)->is_Proj() && !phase->C->major_progress()) {
1428 Node *n0 = in(0)->in(0);
1429 // Check if he is a call projection (except Leaf Call)
1430 if( n0->is_Catch() ) {
1431 n0 = n0->in(0)->in(0);
1432 assert( n0->is_Call(), "expect a call here" );
1433 }
1434 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
1435 // Don't remove a safepoint belonging to an OuterStripMinedLoopEndNode.
1436 // If the loop dies, they will be removed together.
1437 if (has_out_with(Op_OuterStripMinedLoopEnd)) {
1438 return this;
1439 }
1440 // Useless Safepoint, so remove it
1441 return in(TypeFunc::Control);
1442 }
1443 }
1444
1445 return this;
1446 }
1447
1448 //------------------------------Value------------------------------------------
1449 const Type* SafePointNode::Value(PhaseGVN* phase) const {
1450 if (phase->type(in(0)) == Type::TOP) {
1451 return Type::TOP;
1452 }
1453 if (in(0) == this) {
1454 return Type::TOP; // Dead infinite loop
1455 }
1456 return Type::CONTROL;
1457 }
1458
1459 #ifndef PRODUCT
1460 void SafePointNode::dump_spec(outputStream *st) const {
1461 st->print(" SafePoint ");
1462 _replaced_nodes.dump(st);
1463 }
1464
1465 // The related nodes of a SafepointNode are all data inputs, excluding the
1466 // control boundary, as well as all outputs till level 2 (to include projection
1467 // nodes and targets). In compact mode, just include inputs till level 1 and
1468 // outputs as before.
1469 void SafePointNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
1470 if (compact) {
1471 this->collect_nodes(in_rel, 1, false, false);
1472 } else {
1473 this->collect_nodes_in_all_data(in_rel, false);
1474 }
1475 this->collect_nodes(out_rel, -2, false, false);
1476 }
1477 #endif
1478
1479 const RegMask &SafePointNode::in_RegMask(uint idx) const {
1480 if( idx < TypeFunc::Parms ) return RegMask::Empty;
1481 // Values outside the domain represent debug info
1482 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1483 }
1484 const RegMask &SafePointNode::out_RegMask() const {
1485 return RegMask::Empty;
1486 }
1487
1488
1489 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
1490 assert((int)grow_by > 0, "sanity");
1491 int monoff = jvms->monoff();
1492 int scloff = jvms->scloff();
1493 int endoff = jvms->endoff();
1494 assert(endoff == (int)req(), "no other states or debug info after me");
1495 Node* top = Compile::current()->top();
1496 for (uint i = 0; i < grow_by; i++) {
1497 ins_req(monoff, top);
1498 }
1499 jvms->set_monoff(monoff + grow_by);
1500 jvms->set_scloff(scloff + grow_by);
1501 jvms->set_endoff(endoff + grow_by);
1502 }
1503
1504 void SafePointNode::push_monitor(const FastLockNode *lock) {
1505 // Add a LockNode, which points to both the original BoxLockNode (the
1506 // stack space for the monitor) and the Object being locked.
1507 const int MonitorEdges = 2;
1508 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1509 assert(req() == jvms()->endoff(), "correct sizing");
1510 int nextmon = jvms()->scloff();
1511 if (GenerateSynchronizationCode) {
1512 ins_req(nextmon, lock->box_node());
1513 ins_req(nextmon+1, lock->obj_node());
1514 } else {
1515 Node* top = Compile::current()->top();
1516 ins_req(nextmon, top);
1517 ins_req(nextmon, top);
1518 }
1519 jvms()->set_scloff(nextmon + MonitorEdges);
1520 jvms()->set_endoff(req());
1521 }
1522
1523 void SafePointNode::pop_monitor() {
1524 // Delete last monitor from debug info
1525 debug_only(int num_before_pop = jvms()->nof_monitors());
1526 const int MonitorEdges = 2;
1527 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1528 int scloff = jvms()->scloff();
1529 int endoff = jvms()->endoff();
1530 int new_scloff = scloff - MonitorEdges;
1531 int new_endoff = endoff - MonitorEdges;
1532 jvms()->set_scloff(new_scloff);
1533 jvms()->set_endoff(new_endoff);
1534 while (scloff > new_scloff) del_req_ordered(--scloff);
1535 assert(jvms()->nof_monitors() == num_before_pop-1, "");
1536 }
1537
1538 Node *SafePointNode::peek_monitor_box() const {
1539 int mon = jvms()->nof_monitors() - 1;
1540 assert(mon >= 0, "must have a monitor");
1541 return monitor_box(jvms(), mon);
1542 }
1543
1544 Node *SafePointNode::peek_monitor_obj() const {
1545 int mon = jvms()->nof_monitors() - 1;
1546 assert(mon >= 0, "must have a monitor");
1547 return monitor_obj(jvms(), mon);
1548 }
1549
1550 // Do we Match on this edge index or not? Match no edges
1551 uint SafePointNode::match_edge(uint idx) const {
1552 return (TypeFunc::Parms == idx);
1553 }
1554
1555 void SafePointNode::disconnect_from_root(PhaseIterGVN *igvn) {
1556 assert(Opcode() == Op_SafePoint, "only value for safepoint in loops");
1557 int nb = igvn->C->root()->find_prec_edge(this);
1558 if (nb != -1) {
1559 igvn->C->root()->rm_prec(nb);
1560 }
1561 }
1562
1563 //============== SafePointScalarObjectNode ==============
1564
1565 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp,
1566 #ifdef ASSERT
1567 Node* alloc,
1568 #endif
1569 uint first_index,
1570 uint n_fields) :
1571 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1572 _first_index(first_index),
1573 _n_fields(n_fields)
1574 #ifdef ASSERT
1575 , _alloc(alloc)
1576 #endif
1577 {
1578 #ifdef ASSERT
1579 if (!alloc->is_Allocate()
1580 && !(alloc->Opcode() == Op_VectorBox)) {
1581 alloc->dump();
1582 assert(false, "unexpected call node");
1583 }
1584 #endif
1585 init_class_id(Class_SafePointScalarObject);
1586 }
1587
1588 // Do not allow value-numbering for SafePointScalarObject node.
1589 uint SafePointScalarObjectNode::hash() const { return NO_HASH; }
1590 bool SafePointScalarObjectNode::cmp( const Node &n ) const {
1591 return (&n == this); // Always fail except on self
1592 }
1593
1594 uint SafePointScalarObjectNode::ideal_reg() const {
1595 return 0; // No matching to machine instruction
1596 }
1597
1598 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1599 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1600 }
1601
1602 const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1603 return RegMask::Empty;
1604 }
1605
1606 uint SafePointScalarObjectNode::match_edge(uint idx) const {
1607 return 0;
1608 }
1609
1610 SafePointScalarObjectNode*
1611 SafePointScalarObjectNode::clone(Dict* sosn_map, bool& new_node) const {
1612 void* cached = (*sosn_map)[(void*)this];
1613 if (cached != NULL) {
1614 new_node = false;
1615 return (SafePointScalarObjectNode*)cached;
1616 }
1617 new_node = true;
1618 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1619 sosn_map->Insert((void*)this, (void*)res);
1620 return res;
1621 }
1622
1623
1624 #ifndef PRODUCT
1625 void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1626 st->print(" # fields@[%d..%d]", first_index(),
1627 first_index() + n_fields() - 1);
1628 }
1629
1630 #endif
1631
1632 //=============================================================================
1633 uint AllocateNode::size_of() const { return sizeof(*this); }
1634
1635 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1636 Node *ctrl, Node *mem, Node *abio,
1637 Node *size, Node *klass_node, Node *initial_test)
1638 : CallNode(atype, NULL, TypeRawPtr::BOTTOM)
1639 {
1640 init_class_id(Class_Allocate);
1641 init_flags(Flag_is_macro);
1642 _is_scalar_replaceable = false;
1643 _is_non_escaping = false;
1644 _is_allocation_MemBar_redundant = false;
1645 Node *topnode = C->top();
1646
1647 init_req( TypeFunc::Control , ctrl );
1648 init_req( TypeFunc::I_O , abio );
1649 init_req( TypeFunc::Memory , mem );
1650 init_req( TypeFunc::ReturnAdr, topnode );
1651 init_req( TypeFunc::FramePtr , topnode );
1652 init_req( AllocSize , size);
1653 init_req( KlassNode , klass_node);
1654 init_req( InitialTest , initial_test);
1655 init_req( ALength , topnode);
1656 init_req( ValidLengthTest , topnode);
1657 C->add_macro_node(this);
1658 }
1659
1660 void AllocateNode::compute_MemBar_redundancy(ciMethod* initializer)
1661 {
1662 assert(initializer != NULL &&
1663 initializer->is_initializer() &&
1664 !initializer->is_static(),
1665 "unexpected initializer method");
1666 BCEscapeAnalyzer* analyzer = initializer->get_bcea();
1667 if (analyzer == NULL) {
1668 return;
1669 }
1670
1671 // Allocation node is first parameter in its initializer
1672 if (analyzer->is_arg_stack(0) || analyzer->is_arg_local(0)) {
1673 _is_allocation_MemBar_redundant = true;
1674 }
1675 }
1676 Node *AllocateNode::make_ideal_mark(PhaseGVN *phase, Node* obj, Node* control, Node* mem) {
1677 Node* mark_node = NULL;
1678 // For now only enable fast locking for non-array types
1679 mark_node = phase->MakeConX(markWord::prototype().value());
1680 return mark_node;
1681 }
1682
1683 // Retrieve the length from the AllocateArrayNode. Narrow the type with a
1684 // CastII, if appropriate. If we are not allowed to create new nodes, and
1685 // a CastII is appropriate, return NULL.
1686 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) {
1687 Node *length = in(AllocateNode::ALength);
1688 assert(length != NULL, "length is not null");
1689
1690 const TypeInt* length_type = phase->find_int_type(length);
1691 const TypeAryPtr* ary_type = oop_type->isa_aryptr();
1692
1693 if (ary_type != NULL && length_type != NULL) {
1694 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type);
1695 if (narrow_length_type != length_type) {
1696 // Assert one of:
1697 // - the narrow_length is 0
1698 // - the narrow_length is not wider than length
1699 assert(narrow_length_type == TypeInt::ZERO ||
1700 length_type->is_con() && narrow_length_type->is_con() &&
1701 (narrow_length_type->_hi <= length_type->_lo) ||
1702 (narrow_length_type->_hi <= length_type->_hi &&
1703 narrow_length_type->_lo >= length_type->_lo),
1704 "narrow type must be narrower than length type");
1705
1706 // Return NULL if new nodes are not allowed
1707 if (!allow_new_nodes) return NULL;
1708 // Create a cast which is control dependent on the initialization to
1709 // propagate the fact that the array length must be positive.
1710 InitializeNode* init = initialization();
1711 assert(init != NULL, "initialization not found");
1712 length = new CastIINode(length, narrow_length_type);
1713 length->set_req(TypeFunc::Control, init->proj_out_or_null(TypeFunc::Control));
1714 }
1715 }
1716
1717 return length;
1718 }
1719
1720 //=============================================================================
1721 uint LockNode::size_of() const { return sizeof(*this); }
1722
1723 // Redundant lock elimination
1724 //
1725 // There are various patterns of locking where we release and
1726 // immediately reacquire a lock in a piece of code where no operations
1727 // occur in between that would be observable. In those cases we can
1728 // skip releasing and reacquiring the lock without violating any
1729 // fairness requirements. Doing this around a loop could cause a lock
1730 // to be held for a very long time so we concentrate on non-looping
1731 // control flow. We also require that the operations are fully
1732 // redundant meaning that we don't introduce new lock operations on
1733 // some paths so to be able to eliminate it on others ala PRE. This
1734 // would probably require some more extensive graph manipulation to
1735 // guarantee that the memory edges were all handled correctly.
1736 //
1737 // Assuming p is a simple predicate which can't trap in any way and s
1738 // is a synchronized method consider this code:
1739 //
1740 // s();
1741 // if (p)
1742 // s();
1743 // else
1744 // s();
1745 // s();
1746 //
1747 // 1. The unlocks of the first call to s can be eliminated if the
1748 // locks inside the then and else branches are eliminated.
1749 //
1750 // 2. The unlocks of the then and else branches can be eliminated if
1751 // the lock of the final call to s is eliminated.
1752 //
1753 // Either of these cases subsumes the simple case of sequential control flow
1754 //
1755 // Additionally we can eliminate versions without the else case:
1756 //
1757 // s();
1758 // if (p)
1759 // s();
1760 // s();
1761 //
1762 // 3. In this case we eliminate the unlock of the first s, the lock
1763 // and unlock in the then case and the lock in the final s.
1764 //
1765 // Note also that in all these cases the then/else pieces don't have
1766 // to be trivial as long as they begin and end with synchronization
1767 // operations.
1768 //
1769 // s();
1770 // if (p)
1771 // s();
1772 // f();
1773 // s();
1774 // s();
1775 //
1776 // The code will work properly for this case, leaving in the unlock
1777 // before the call to f and the relock after it.
1778 //
1779 // A potentially interesting case which isn't handled here is when the
1780 // locking is partially redundant.
1781 //
1782 // s();
1783 // if (p)
1784 // s();
1785 //
1786 // This could be eliminated putting unlocking on the else case and
1787 // eliminating the first unlock and the lock in the then side.
1788 // Alternatively the unlock could be moved out of the then side so it
1789 // was after the merge and the first unlock and second lock
1790 // eliminated. This might require less manipulation of the memory
1791 // state to get correct.
1792 //
1793 // Additionally we might allow work between a unlock and lock before
1794 // giving up eliminating the locks. The current code disallows any
1795 // conditional control flow between these operations. A formulation
1796 // similar to partial redundancy elimination computing the
1797 // availability of unlocking and the anticipatability of locking at a
1798 // program point would allow detection of fully redundant locking with
1799 // some amount of work in between. I'm not sure how often I really
1800 // think that would occur though. Most of the cases I've seen
1801 // indicate it's likely non-trivial work would occur in between.
1802 // There may be other more complicated constructs where we could
1803 // eliminate locking but I haven't seen any others appear as hot or
1804 // interesting.
1805 //
1806 // Locking and unlocking have a canonical form in ideal that looks
1807 // roughly like this:
1808 //
1809 // <obj>
1810 // | \\------+
1811 // | \ \
1812 // | BoxLock \
1813 // | | | \
1814 // | | \ \
1815 // | | FastLock
1816 // | | /
1817 // | | /
1818 // | | |
1819 //
1820 // Lock
1821 // |
1822 // Proj #0
1823 // |
1824 // MembarAcquire
1825 // |
1826 // Proj #0
1827 //
1828 // MembarRelease
1829 // |
1830 // Proj #0
1831 // |
1832 // Unlock
1833 // |
1834 // Proj #0
1835 //
1836 //
1837 // This code proceeds by processing Lock nodes during PhaseIterGVN
1838 // and searching back through its control for the proper code
1839 // patterns. Once it finds a set of lock and unlock operations to
1840 // eliminate they are marked as eliminatable which causes the
1841 // expansion of the Lock and Unlock macro nodes to make the operation a NOP
1842 //
1843 //=============================================================================
1844
1845 //
1846 // Utility function to skip over uninteresting control nodes. Nodes skipped are:
1847 // - copy regions. (These may not have been optimized away yet.)
1848 // - eliminated locking nodes
1849 //
1850 static Node *next_control(Node *ctrl) {
1851 if (ctrl == NULL)
1852 return NULL;
1853 while (1) {
1854 if (ctrl->is_Region()) {
1855 RegionNode *r = ctrl->as_Region();
1856 Node *n = r->is_copy();
1857 if (n == NULL)
1858 break; // hit a region, return it
1859 else
1860 ctrl = n;
1861 } else if (ctrl->is_Proj()) {
1862 Node *in0 = ctrl->in(0);
1863 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
1864 ctrl = in0->in(0);
1865 } else {
1866 break;
1867 }
1868 } else {
1869 break; // found an interesting control
1870 }
1871 }
1872 return ctrl;
1873 }
1874 //
1875 // Given a control, see if it's the control projection of an Unlock which
1876 // operating on the same object as lock.
1877 //
1878 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
1879 GrowableArray<AbstractLockNode*> &lock_ops) {
1880 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL;
1881 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) {
1882 Node *n = ctrl_proj->in(0);
1883 if (n != NULL && n->is_Unlock()) {
1884 UnlockNode *unlock = n->as_Unlock();
1885 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1886 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1887 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
1888 if (lock_obj->eqv_uncast(unlock_obj) &&
1889 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) &&
1890 !unlock->is_eliminated()) {
1891 lock_ops.append(unlock);
1892 return true;
1893 }
1894 }
1895 }
1896 return false;
1897 }
1898
1899 //
1900 // Find the lock matching an unlock. Returns null if a safepoint
1901 // or complicated control is encountered first.
1902 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
1903 LockNode *lock_result = NULL;
1904 // find the matching lock, or an intervening safepoint
1905 Node *ctrl = next_control(unlock->in(0));
1906 while (1) {
1907 assert(ctrl != NULL, "invalid control graph");
1908 assert(!ctrl->is_Start(), "missing lock for unlock");
1909 if (ctrl->is_top()) break; // dead control path
1910 if (ctrl->is_Proj()) ctrl = ctrl->in(0);
1911 if (ctrl->is_SafePoint()) {
1912 break; // found a safepoint (may be the lock we are searching for)
1913 } else if (ctrl->is_Region()) {
1914 // Check for a simple diamond pattern. Punt on anything more complicated
1915 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) {
1916 Node *in1 = next_control(ctrl->in(1));
1917 Node *in2 = next_control(ctrl->in(2));
1918 if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
1919 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
1920 ctrl = next_control(in1->in(0)->in(0));
1921 } else {
1922 break;
1923 }
1924 } else {
1925 break;
1926 }
1927 } else {
1928 ctrl = next_control(ctrl->in(0)); // keep searching
1929 }
1930 }
1931 if (ctrl->is_Lock()) {
1932 LockNode *lock = ctrl->as_Lock();
1933 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1934 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1935 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node());
1936 if (lock_obj->eqv_uncast(unlock_obj) &&
1937 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) {
1938 lock_result = lock;
1939 }
1940 }
1941 return lock_result;
1942 }
1943
1944 // This code corresponds to case 3 above.
1945
1946 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
1947 GrowableArray<AbstractLockNode*> &lock_ops) {
1948 Node* if_node = node->in(0);
1949 bool if_true = node->is_IfTrue();
1950
1951 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
1952 Node *lock_ctrl = next_control(if_node->in(0));
1953 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
1954 Node* lock1_node = NULL;
1955 ProjNode* proj = if_node->as_If()->proj_out(!if_true);
1956 if (if_true) {
1957 if (proj->is_IfFalse() && proj->outcnt() == 1) {
1958 lock1_node = proj->unique_out();
1959 }
1960 } else {
1961 if (proj->is_IfTrue() && proj->outcnt() == 1) {
1962 lock1_node = proj->unique_out();
1963 }
1964 }
1965 if (lock1_node != NULL && lock1_node->is_Lock()) {
1966 LockNode *lock1 = lock1_node->as_Lock();
1967 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1968 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node());
1969 Node* lock1_obj = bs->step_over_gc_barrier(lock1->obj_node());
1970 if (lock_obj->eqv_uncast(lock1_obj) &&
1971 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) &&
1972 !lock1->is_eliminated()) {
1973 lock_ops.append(lock1);
1974 return true;
1975 }
1976 }
1977 }
1978 }
1979
1980 lock_ops.trunc_to(0);
1981 return false;
1982 }
1983
1984 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
1985 GrowableArray<AbstractLockNode*> &lock_ops) {
1986 // check each control merging at this point for a matching unlock.
1987 // in(0) should be self edge so skip it.
1988 for (int i = 1; i < (int)region->req(); i++) {
1989 Node *in_node = next_control(region->in(i));
1990 if (in_node != NULL) {
1991 if (find_matching_unlock(in_node, lock, lock_ops)) {
1992 // found a match so keep on checking.
1993 continue;
1994 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
1995 continue;
1996 }
1997
1998 // If we fall through to here then it was some kind of node we
1999 // don't understand or there wasn't a matching unlock, so give
2000 // up trying to merge locks.
2001 lock_ops.trunc_to(0);
2002 return false;
2003 }
2004 }
2005 return true;
2006
2007 }
2008
2009 const char* AbstractLockNode::_kind_names[] = {"Regular", "NonEscObj", "Coarsened", "Nested"};
2010
2011 const char * AbstractLockNode::kind_as_string() const {
2012 return _kind_names[_kind];
2013 }
2014
2015 #ifndef PRODUCT
2016 //
2017 // Create a counter which counts the number of times this lock is acquired
2018 //
2019 void AbstractLockNode::create_lock_counter(JVMState* state) {
2020 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
2021 }
2022
2023 void AbstractLockNode::set_eliminated_lock_counter() {
2024 if (_counter) {
2025 // Update the counter to indicate that this lock was eliminated.
2026 // The counter update code will stay around even though the
2027 // optimizer will eliminate the lock operation itself.
2028 _counter->set_tag(NamedCounter::EliminatedLockCounter);
2029 }
2030 }
2031
2032 void AbstractLockNode::dump_spec(outputStream* st) const {
2033 st->print("%s ", _kind_names[_kind]);
2034 CallNode::dump_spec(st);
2035 }
2036
2037 void AbstractLockNode::dump_compact_spec(outputStream* st) const {
2038 st->print("%s", _kind_names[_kind]);
2039 }
2040
2041 // The related set of lock nodes includes the control boundary.
2042 void AbstractLockNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const {
2043 if (compact) {
2044 this->collect_nodes(in_rel, 1, false, false);
2045 } else {
2046 this->collect_nodes_in_all_data(in_rel, true);
2047 }
2048 this->collect_nodes(out_rel, -2, false, false);
2049 }
2050 #endif
2051
2052 //=============================================================================
2053 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2054
2055 // perform any generic optimizations first (returns 'this' or NULL)
2056 Node *result = SafePointNode::Ideal(phase, can_reshape);
2057 if (result != NULL) return result;
2058 // Don't bother trying to transform a dead node
2059 if (in(0) && in(0)->is_top()) return NULL;
2060
2061 // Now see if we can optimize away this lock. We don't actually
2062 // remove the locking here, we simply set the _eliminate flag which
2063 // prevents macro expansion from expanding the lock. Since we don't
2064 // modify the graph, the value returned from this function is the
2065 // one computed above.
2066 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
2067 //
2068 // If we are locking an non-escaped object, the lock/unlock is unnecessary
2069 //
2070 ConnectionGraph *cgr = phase->C->congraph();
2071 if (cgr != NULL && cgr->not_global_escape(obj_node())) {
2072 assert(!is_eliminated() || is_coarsened(), "sanity");
2073 // The lock could be marked eliminated by lock coarsening
2074 // code during first IGVN before EA. Replace coarsened flag
2075 // to eliminate all associated locks/unlocks.
2076 #ifdef ASSERT
2077 this->log_lock_optimization(phase->C,"eliminate_lock_set_non_esc1");
2078 #endif
2079 this->set_non_esc_obj();
2080 return result;
2081 }
2082
2083 if (!phase->C->do_locks_coarsening()) {
2084 return result; // Compiling without locks coarsening
2085 }
2086 //
2087 // Try lock coarsening
2088 //
2089 PhaseIterGVN* iter = phase->is_IterGVN();
2090 if (iter != NULL && !is_eliminated()) {
2091
2092 GrowableArray<AbstractLockNode*> lock_ops;
2093
2094 Node *ctrl = next_control(in(0));
2095
2096 // now search back for a matching Unlock
2097 if (find_matching_unlock(ctrl, this, lock_ops)) {
2098 // found an unlock directly preceding this lock. This is the
2099 // case of single unlock directly control dependent on a
2100 // single lock which is the trivial version of case 1 or 2.
2101 } else if (ctrl->is_Region() ) {
2102 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
2103 // found lock preceded by multiple unlocks along all paths
2104 // joining at this point which is case 3 in description above.
2105 }
2106 } else {
2107 // see if this lock comes from either half of an if and the
2108 // predecessors merges unlocks and the other half of the if
2109 // performs a lock.
2110 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
2111 // found unlock splitting to an if with locks on both branches.
2112 }
2113 }
2114
2115 if (lock_ops.length() > 0) {
2116 // add ourselves to the list of locks to be eliminated.
2117 lock_ops.append(this);
2118
2119 #ifndef PRODUCT
2120 if (PrintEliminateLocks) {
2121 int locks = 0;
2122 int unlocks = 0;
2123 if (Verbose) {
2124 tty->print_cr("=== Locks coarsening ===");
2125 }
2126 for (int i = 0; i < lock_ops.length(); i++) {
2127 AbstractLockNode* lock = lock_ops.at(i);
2128 if (lock->Opcode() == Op_Lock)
2129 locks++;
2130 else
2131 unlocks++;
2132 if (Verbose) {
2133 tty->print(" %d: ", i);
2134 lock->dump();
2135 }
2136 }
2137 tty->print_cr("=== Coarsened %d unlocks and %d locks", unlocks, locks);
2138 }
2139 #endif
2140
2141 // for each of the identified locks, mark them
2142 // as eliminatable
2143 for (int i = 0; i < lock_ops.length(); i++) {
2144 AbstractLockNode* lock = lock_ops.at(i);
2145
2146 // Mark it eliminated by coarsening and update any counters
2147 #ifdef ASSERT
2148 lock->log_lock_optimization(phase->C, "eliminate_lock_set_coarsened");
2149 #endif
2150 lock->set_coarsened();
2151 }
2152 // Record this coarsened group.
2153 phase->C->add_coarsened_locks(lock_ops);
2154 } else if (ctrl->is_Region() &&
2155 iter->_worklist.member(ctrl)) {
2156 // We weren't able to find any opportunities but the region this
2157 // lock is control dependent on hasn't been processed yet so put
2158 // this lock back on the worklist so we can check again once any
2159 // region simplification has occurred.
2160 iter->_worklist.push(this);
2161 }
2162 }
2163 }
2164
2165 return result;
2166 }
2167
2168 //=============================================================================
2169 bool LockNode::is_nested_lock_region() {
2170 return is_nested_lock_region(NULL);
2171 }
2172
2173 // p is used for access to compilation log; no logging if NULL
2174 bool LockNode::is_nested_lock_region(Compile * c) {
2175 BoxLockNode* box = box_node()->as_BoxLock();
2176 int stk_slot = box->stack_slot();
2177 if (stk_slot <= 0) {
2178 #ifdef ASSERT
2179 this->log_lock_optimization(c, "eliminate_lock_INLR_1");
2180 #endif
2181 return false; // External lock or it is not Box (Phi node).
2182 }
2183
2184 // Ignore complex cases: merged locks or multiple locks.
2185 Node* obj = obj_node();
2186 LockNode* unique_lock = NULL;
2187 Node* bad_lock = NULL;
2188 if (!box->is_simple_lock_region(&unique_lock, obj, &bad_lock)) {
2189 #ifdef ASSERT
2190 this->log_lock_optimization(c, "eliminate_lock_INLR_2a", bad_lock);
2191 #endif
2192 return false;
2193 }
2194 if (unique_lock != this) {
2195 #ifdef ASSERT
2196 this->log_lock_optimization(c, "eliminate_lock_INLR_2b", (unique_lock != NULL ? unique_lock : bad_lock));
2197 if (PrintEliminateLocks && Verbose) {
2198 tty->print_cr("=============== unique_lock != this ============");
2199 tty->print(" this: ");
2200 this->dump();
2201 tty->print(" box: ");
2202 box->dump();
2203 tty->print(" obj: ");
2204 obj->dump();
2205 if (unique_lock != NULL) {
2206 tty->print(" unique_lock: ");
2207 unique_lock->dump();
2208 }
2209 if (bad_lock != NULL) {
2210 tty->print(" bad_lock: ");
2211 bad_lock->dump();
2212 }
2213 tty->print_cr("===============");
2214 }
2215 #endif
2216 return false;
2217 }
2218
2219 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2220 obj = bs->step_over_gc_barrier(obj);
2221 // Look for external lock for the same object.
2222 SafePointNode* sfn = this->as_SafePoint();
2223 JVMState* youngest_jvms = sfn->jvms();
2224 int max_depth = youngest_jvms->depth();
2225 for (int depth = 1; depth <= max_depth; depth++) {
2226 JVMState* jvms = youngest_jvms->of_depth(depth);
2227 int num_mon = jvms->nof_monitors();
2228 // Loop over monitors
2229 for (int idx = 0; idx < num_mon; idx++) {
2230 Node* obj_node = sfn->monitor_obj(jvms, idx);
2231 obj_node = bs->step_over_gc_barrier(obj_node);
2232 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock();
2233 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) {
2234 return true;
2235 }
2236 }
2237 }
2238 #ifdef ASSERT
2239 this->log_lock_optimization(c, "eliminate_lock_INLR_3");
2240 #endif
2241 return false;
2242 }
2243
2244 //=============================================================================
2245 uint UnlockNode::size_of() const { return sizeof(*this); }
2246
2247 //=============================================================================
2248 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2249
2250 // perform any generic optimizations first (returns 'this' or NULL)
2251 Node *result = SafePointNode::Ideal(phase, can_reshape);
2252 if (result != NULL) return result;
2253 // Don't bother trying to transform a dead node
2254 if (in(0) && in(0)->is_top()) return NULL;
2255
2256 // Now see if we can optimize away this unlock. We don't actually
2257 // remove the unlocking here, we simply set the _eliminate flag which
2258 // prevents macro expansion from expanding the unlock. Since we don't
2259 // modify the graph, the value returned from this function is the
2260 // one computed above.
2261 // Escape state is defined after Parse phase.
2262 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
2263 //
2264 // If we are unlocking an non-escaped object, the lock/unlock is unnecessary.
2265 //
2266 ConnectionGraph *cgr = phase->C->congraph();
2267 if (cgr != NULL && cgr->not_global_escape(obj_node())) {
2268 assert(!is_eliminated() || is_coarsened(), "sanity");
2269 // The lock could be marked eliminated by lock coarsening
2270 // code during first IGVN before EA. Replace coarsened flag
2271 // to eliminate all associated locks/unlocks.
2272 #ifdef ASSERT
2273 this->log_lock_optimization(phase->C, "eliminate_lock_set_non_esc2");
2274 #endif
2275 this->set_non_esc_obj();
2276 }
2277 }
2278 return result;
2279 }
2280
2281 void AbstractLockNode::log_lock_optimization(Compile *C, const char * tag, Node* bad_lock) const {
2282 if (C == NULL) {
2283 return;
2284 }
2285 CompileLog* log = C->log();
2286 if (log != NULL) {
2287 Node* box = box_node();
2288 Node* obj = obj_node();
2289 int box_id = box != NULL ? box->_idx : -1;
2290 int obj_id = obj != NULL ? obj->_idx : -1;
2291
2292 log->begin_head("%s compile_id='%d' lock_id='%d' class='%s' kind='%s' box_id='%d' obj_id='%d' bad_id='%d'",
2293 tag, C->compile_id(), this->_idx,
2294 is_Unlock() ? "unlock" : is_Lock() ? "lock" : "?",
2295 kind_as_string(), box_id, obj_id, (bad_lock != NULL ? bad_lock->_idx : -1));
2296 log->stamp();
2297 log->end_head();
2298 JVMState* p = is_Unlock() ? (as_Unlock()->dbg_jvms()) : jvms();
2299 while (p != NULL) {
2300 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
2301 p = p->caller();
2302 }
2303 log->tail(tag);
2304 }
2305 }
2306
2307 bool CallNode::may_modify_arraycopy_helper(const TypeOopPtr* dest_t, const TypeOopPtr *t_oop, PhaseTransform *phase) {
2308 if (dest_t->is_known_instance() && t_oop->is_known_instance()) {
2309 return dest_t->instance_id() == t_oop->instance_id();
2310 }
2311
2312 if (dest_t->isa_instptr() && !dest_t->klass()->equals(phase->C->env()->Object_klass())) {
2313 // clone
2314 if (t_oop->isa_aryptr()) {
2315 return false;
2316 }
2317 if (!t_oop->isa_instptr()) {
2318 return true;
2319 }
2320 if (dest_t->klass()->is_subtype_of(t_oop->klass()) || t_oop->klass()->is_subtype_of(dest_t->klass())) {
2321 return true;
2322 }
2323 // unrelated
2324 return false;
2325 }
2326
2327 if (dest_t->isa_aryptr()) {
2328 // arraycopy or array clone
2329 if (t_oop->isa_instptr()) {
2330 return false;
2331 }
2332 if (!t_oop->isa_aryptr()) {
2333 return true;
2334 }
2335
2336 const Type* elem = dest_t->is_aryptr()->elem();
2337 if (elem == Type::BOTTOM) {
2338 // An array but we don't know what elements are
2339 return true;
2340 }
2341
2342 dest_t = dest_t->add_offset(Type::OffsetBot)->is_oopptr();
2343 uint dest_alias = phase->C->get_alias_index(dest_t);
2344 uint t_oop_alias = phase->C->get_alias_index(t_oop);
2345
2346 return dest_alias == t_oop_alias;
2347 }
2348
2349 return true;
2350 }