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