Fri, 12 Feb 2010 15:27:36 -0800
Merge
1 /*
2 * Copyright 1997-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 // Portions of code courtesy of Clifford Click
27 // Optimization - Graph Style
29 #include "incls/_precompiled.incl"
30 #include "incls/_callnode.cpp.incl"
32 //=============================================================================
33 uint StartNode::size_of() const { return sizeof(*this); }
34 uint StartNode::cmp( const Node &n ) const
35 { return _domain == ((StartNode&)n)._domain; }
36 const Type *StartNode::bottom_type() const { return _domain; }
37 const Type *StartNode::Value(PhaseTransform *phase) const { return _domain; }
38 #ifndef PRODUCT
39 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);}
40 #endif
42 //------------------------------Ideal------------------------------------------
43 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){
44 return remove_dead_region(phase, can_reshape) ? this : NULL;
45 }
47 //------------------------------calling_convention-----------------------------
48 void StartNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
49 Matcher::calling_convention( sig_bt, parm_regs, argcnt, false );
50 }
52 //------------------------------Registers--------------------------------------
53 const RegMask &StartNode::in_RegMask(uint) const {
54 return RegMask::Empty;
55 }
57 //------------------------------match------------------------------------------
58 // Construct projections for incoming parameters, and their RegMask info
59 Node *StartNode::match( const ProjNode *proj, const Matcher *match ) {
60 switch (proj->_con) {
61 case TypeFunc::Control:
62 case TypeFunc::I_O:
63 case TypeFunc::Memory:
64 return new (match->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
65 case TypeFunc::FramePtr:
66 return new (match->C, 1) MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP);
67 case TypeFunc::ReturnAdr:
68 return new (match->C, 1) MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP);
69 case TypeFunc::Parms:
70 default: {
71 uint parm_num = proj->_con - TypeFunc::Parms;
72 const Type *t = _domain->field_at(proj->_con);
73 if (t->base() == Type::Half) // 2nd half of Longs and Doubles
74 return new (match->C, 1) ConNode(Type::TOP);
75 uint ideal_reg = Matcher::base2reg[t->base()];
76 RegMask &rm = match->_calling_convention_mask[parm_num];
77 return new (match->C, 1) MachProjNode(this,proj->_con,rm,ideal_reg);
78 }
79 }
80 return NULL;
81 }
83 //------------------------------StartOSRNode----------------------------------
84 // The method start node for an on stack replacement adapter
86 //------------------------------osr_domain-----------------------------
87 const TypeTuple *StartOSRNode::osr_domain() {
88 const Type **fields = TypeTuple::fields(2);
89 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer
91 return TypeTuple::make(TypeFunc::Parms+1, fields);
92 }
94 //=============================================================================
95 const char * const ParmNode::names[TypeFunc::Parms+1] = {
96 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms"
97 };
99 #ifndef PRODUCT
100 void ParmNode::dump_spec(outputStream *st) const {
101 if( _con < TypeFunc::Parms ) {
102 st->print(names[_con]);
103 } else {
104 st->print("Parm%d: ",_con-TypeFunc::Parms);
105 // Verbose and WizardMode dump bottom_type for all nodes
106 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st);
107 }
108 }
109 #endif
111 uint ParmNode::ideal_reg() const {
112 switch( _con ) {
113 case TypeFunc::Control : // fall through
114 case TypeFunc::I_O : // fall through
115 case TypeFunc::Memory : return 0;
116 case TypeFunc::FramePtr : // fall through
117 case TypeFunc::ReturnAdr: return Op_RegP;
118 default : assert( _con > TypeFunc::Parms, "" );
119 // fall through
120 case TypeFunc::Parms : {
121 // Type of argument being passed
122 const Type *t = in(0)->as_Start()->_domain->field_at(_con);
123 return Matcher::base2reg[t->base()];
124 }
125 }
126 ShouldNotReachHere();
127 return 0;
128 }
130 //=============================================================================
131 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) {
132 init_req(TypeFunc::Control,cntrl);
133 init_req(TypeFunc::I_O,i_o);
134 init_req(TypeFunc::Memory,memory);
135 init_req(TypeFunc::FramePtr,frameptr);
136 init_req(TypeFunc::ReturnAdr,retadr);
137 }
139 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){
140 return remove_dead_region(phase, can_reshape) ? this : NULL;
141 }
143 const Type *ReturnNode::Value( PhaseTransform *phase ) const {
144 return ( phase->type(in(TypeFunc::Control)) == Type::TOP)
145 ? Type::TOP
146 : Type::BOTTOM;
147 }
149 // Do we Match on this edge index or not? No edges on return nodes
150 uint ReturnNode::match_edge(uint idx) const {
151 return 0;
152 }
155 #ifndef PRODUCT
156 void ReturnNode::dump_req() const {
157 // Dump the required inputs, enclosed in '(' and ')'
158 uint i; // Exit value of loop
159 for( i=0; i<req(); i++ ) { // For all required inputs
160 if( i == TypeFunc::Parms ) tty->print("returns");
161 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
162 else tty->print("_ ");
163 }
164 }
165 #endif
167 //=============================================================================
168 RethrowNode::RethrowNode(
169 Node* cntrl,
170 Node* i_o,
171 Node* memory,
172 Node* frameptr,
173 Node* ret_adr,
174 Node* exception
175 ) : Node(TypeFunc::Parms + 1) {
176 init_req(TypeFunc::Control , cntrl );
177 init_req(TypeFunc::I_O , i_o );
178 init_req(TypeFunc::Memory , memory );
179 init_req(TypeFunc::FramePtr , frameptr );
180 init_req(TypeFunc::ReturnAdr, ret_adr);
181 init_req(TypeFunc::Parms , exception);
182 }
184 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){
185 return remove_dead_region(phase, can_reshape) ? this : NULL;
186 }
188 const Type *RethrowNode::Value( PhaseTransform *phase ) const {
189 return (phase->type(in(TypeFunc::Control)) == Type::TOP)
190 ? Type::TOP
191 : Type::BOTTOM;
192 }
194 uint RethrowNode::match_edge(uint idx) const {
195 return 0;
196 }
198 #ifndef PRODUCT
199 void RethrowNode::dump_req() const {
200 // Dump the required inputs, enclosed in '(' and ')'
201 uint i; // Exit value of loop
202 for( i=0; i<req(); i++ ) { // For all required inputs
203 if( i == TypeFunc::Parms ) tty->print("exception");
204 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
205 else tty->print("_ ");
206 }
207 }
208 #endif
210 //=============================================================================
211 // Do we Match on this edge index or not? Match only target address & method
212 uint TailCallNode::match_edge(uint idx) const {
213 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
214 }
216 //=============================================================================
217 // Do we Match on this edge index or not? Match only target address & oop
218 uint TailJumpNode::match_edge(uint idx) const {
219 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
220 }
222 //=============================================================================
223 JVMState::JVMState(ciMethod* method, JVMState* caller) {
224 assert(method != NULL, "must be valid call site");
225 _method = method;
226 _reexecute = Reexecute_Undefined;
227 debug_only(_bci = -99); // random garbage value
228 debug_only(_map = (SafePointNode*)-1);
229 _caller = caller;
230 _depth = 1 + (caller == NULL ? 0 : caller->depth());
231 _locoff = TypeFunc::Parms;
232 _stkoff = _locoff + _method->max_locals();
233 _monoff = _stkoff + _method->max_stack();
234 _scloff = _monoff;
235 _endoff = _monoff;
236 _sp = 0;
237 }
238 JVMState::JVMState(int stack_size) {
239 _method = NULL;
240 _bci = InvocationEntryBci;
241 _reexecute = Reexecute_Undefined;
242 debug_only(_map = (SafePointNode*)-1);
243 _caller = NULL;
244 _depth = 1;
245 _locoff = TypeFunc::Parms;
246 _stkoff = _locoff;
247 _monoff = _stkoff + stack_size;
248 _scloff = _monoff;
249 _endoff = _monoff;
250 _sp = 0;
251 }
253 //--------------------------------of_depth-------------------------------------
254 JVMState* JVMState::of_depth(int d) const {
255 const JVMState* jvmp = this;
256 assert(0 < d && (uint)d <= depth(), "oob");
257 for (int skip = depth() - d; skip > 0; skip--) {
258 jvmp = jvmp->caller();
259 }
260 assert(jvmp->depth() == (uint)d, "found the right one");
261 return (JVMState*)jvmp;
262 }
264 //-----------------------------same_calls_as-----------------------------------
265 bool JVMState::same_calls_as(const JVMState* that) const {
266 if (this == that) return true;
267 if (this->depth() != that->depth()) return false;
268 const JVMState* p = this;
269 const JVMState* q = that;
270 for (;;) {
271 if (p->_method != q->_method) return false;
272 if (p->_method == NULL) return true; // bci is irrelevant
273 if (p->_bci != q->_bci) return false;
274 if (p->_reexecute != q->_reexecute) return false;
275 p = p->caller();
276 q = q->caller();
277 if (p == q) return true;
278 assert(p != NULL && q != NULL, "depth check ensures we don't run off end");
279 }
280 }
282 //------------------------------debug_start------------------------------------
283 uint JVMState::debug_start() const {
284 debug_only(JVMState* jvmroot = of_depth(1));
285 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
286 return of_depth(1)->locoff();
287 }
289 //-------------------------------debug_end-------------------------------------
290 uint JVMState::debug_end() const {
291 debug_only(JVMState* jvmroot = of_depth(1));
292 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
293 return endoff();
294 }
296 //------------------------------debug_depth------------------------------------
297 uint JVMState::debug_depth() const {
298 uint total = 0;
299 for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) {
300 total += jvmp->debug_size();
301 }
302 return total;
303 }
305 #ifndef PRODUCT
307 //------------------------------format_helper----------------------------------
308 // Given an allocation (a Chaitin object) and a Node decide if the Node carries
309 // any defined value or not. If it does, print out the register or constant.
310 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
311 if (n == NULL) { st->print(" NULL"); return; }
312 if (n->is_SafePointScalarObject()) {
313 // Scalar replacement.
314 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
315 scobjs->append_if_missing(spobj);
316 int sco_n = scobjs->find(spobj);
317 assert(sco_n >= 0, "");
318 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
319 return;
320 }
321 if( OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
322 char buf[50];
323 regalloc->dump_register(n,buf);
324 st->print(" %s%d]=%s",msg,i,buf);
325 } else { // No register, but might be constant
326 const Type *t = n->bottom_type();
327 switch (t->base()) {
328 case Type::Int:
329 st->print(" %s%d]=#"INT32_FORMAT,msg,i,t->is_int()->get_con());
330 break;
331 case Type::AnyPtr:
332 assert( t == TypePtr::NULL_PTR, "" );
333 st->print(" %s%d]=#NULL",msg,i);
334 break;
335 case Type::AryPtr:
336 case Type::KlassPtr:
337 case Type::InstPtr:
338 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->isa_oopptr()->const_oop());
339 break;
340 case Type::NarrowOop:
341 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->make_ptr()->isa_oopptr()->const_oop());
342 break;
343 case Type::RawPtr:
344 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,t->is_rawptr());
345 break;
346 case Type::DoubleCon:
347 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
348 break;
349 case Type::FloatCon:
350 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
351 break;
352 case Type::Long:
353 st->print(" %s%d]=#"INT64_FORMAT,msg,i,t->is_long()->get_con());
354 break;
355 case Type::Half:
356 case Type::Top:
357 st->print(" %s%d]=_",msg,i);
358 break;
359 default: ShouldNotReachHere();
360 }
361 }
362 }
364 //------------------------------format-----------------------------------------
365 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
366 st->print(" #");
367 if( _method ) {
368 _method->print_short_name(st);
369 st->print(" @ bci:%d ",_bci);
370 } else {
371 st->print_cr(" runtime stub ");
372 return;
373 }
374 if (n->is_MachSafePoint()) {
375 GrowableArray<SafePointScalarObjectNode*> scobjs;
376 MachSafePointNode *mcall = n->as_MachSafePoint();
377 uint i;
378 // Print locals
379 for( i = 0; i < (uint)loc_size(); i++ )
380 format_helper( regalloc, st, mcall->local(this, i), "L[", i, &scobjs );
381 // Print stack
382 for (i = 0; i < (uint)stk_size(); i++) {
383 if ((uint)(_stkoff + i) >= mcall->len())
384 st->print(" oob ");
385 else
386 format_helper( regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs );
387 }
388 for (i = 0; (int)i < nof_monitors(); i++) {
389 Node *box = mcall->monitor_box(this, i);
390 Node *obj = mcall->monitor_obj(this, i);
391 if ( OptoReg::is_valid(regalloc->get_reg_first(box)) ) {
392 while( !box->is_BoxLock() ) box = box->in(1);
393 format_helper( regalloc, st, box, "MON-BOX[", i, &scobjs );
394 } else {
395 OptoReg::Name box_reg = BoxLockNode::stack_slot(box);
396 st->print(" MON-BOX%d=%s+%d",
397 i,
398 OptoReg::regname(OptoReg::c_frame_pointer),
399 regalloc->reg2offset(box_reg));
400 }
401 const char* obj_msg = "MON-OBJ[";
402 if (EliminateLocks) {
403 while( !box->is_BoxLock() ) box = box->in(1);
404 if (box->as_BoxLock()->is_eliminated())
405 obj_msg = "MON-OBJ(LOCK ELIMINATED)[";
406 }
407 format_helper( regalloc, st, obj, obj_msg, i, &scobjs );
408 }
410 for (i = 0; i < (uint)scobjs.length(); i++) {
411 // Scalar replaced objects.
412 st->print_cr("");
413 st->print(" # ScObj" INT32_FORMAT " ", i);
414 SafePointScalarObjectNode* spobj = scobjs.at(i);
415 ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass();
416 assert(cik->is_instance_klass() ||
417 cik->is_array_klass(), "Not supported allocation.");
418 ciInstanceKlass *iklass = NULL;
419 if (cik->is_instance_klass()) {
420 cik->print_name_on(st);
421 iklass = cik->as_instance_klass();
422 } else if (cik->is_type_array_klass()) {
423 cik->as_array_klass()->base_element_type()->print_name_on(st);
424 st->print("[%d]", spobj->n_fields());
425 } else if (cik->is_obj_array_klass()) {
426 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass();
427 if (cie->is_instance_klass()) {
428 cie->print_name_on(st);
429 } else if (cie->is_type_array_klass()) {
430 cie->as_array_klass()->base_element_type()->print_name_on(st);
431 } else {
432 ShouldNotReachHere();
433 }
434 st->print("[%d]", spobj->n_fields());
435 int ndim = cik->as_array_klass()->dimension() - 1;
436 while (ndim-- > 0) {
437 st->print("[]");
438 }
439 }
440 st->print("={");
441 uint nf = spobj->n_fields();
442 if (nf > 0) {
443 uint first_ind = spobj->first_index();
444 Node* fld_node = mcall->in(first_ind);
445 ciField* cifield;
446 if (iklass != NULL) {
447 st->print(" [");
448 cifield = iklass->nonstatic_field_at(0);
449 cifield->print_name_on(st);
450 format_helper( regalloc, st, fld_node, ":", 0, &scobjs );
451 } else {
452 format_helper( regalloc, st, fld_node, "[", 0, &scobjs );
453 }
454 for (uint j = 1; j < nf; j++) {
455 fld_node = mcall->in(first_ind+j);
456 if (iklass != NULL) {
457 st->print(", [");
458 cifield = iklass->nonstatic_field_at(j);
459 cifield->print_name_on(st);
460 format_helper( regalloc, st, fld_node, ":", j, &scobjs );
461 } else {
462 format_helper( regalloc, st, fld_node, ", [", j, &scobjs );
463 }
464 }
465 }
466 st->print(" }");
467 }
468 }
469 st->print_cr("");
470 if (caller() != NULL) caller()->format(regalloc, n, st);
471 }
474 void JVMState::dump_spec(outputStream *st) const {
475 if (_method != NULL) {
476 bool printed = false;
477 if (!Verbose) {
478 // The JVMS dumps make really, really long lines.
479 // Take out the most boring parts, which are the package prefixes.
480 char buf[500];
481 stringStream namest(buf, sizeof(buf));
482 _method->print_short_name(&namest);
483 if (namest.count() < sizeof(buf)) {
484 const char* name = namest.base();
485 if (name[0] == ' ') ++name;
486 const char* endcn = strchr(name, ':'); // end of class name
487 if (endcn == NULL) endcn = strchr(name, '(');
488 if (endcn == NULL) endcn = name + strlen(name);
489 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
490 --endcn;
491 st->print(" %s", endcn);
492 printed = true;
493 }
494 }
495 if (!printed)
496 _method->print_short_name(st);
497 st->print(" @ bci:%d",_bci);
498 if(_reexecute == Reexecute_True)
499 st->print(" reexecute");
500 } else {
501 st->print(" runtime stub");
502 }
503 if (caller() != NULL) caller()->dump_spec(st);
504 }
507 void JVMState::dump_on(outputStream* st) const {
508 if (_map && !((uintptr_t)_map & 1)) {
509 if (_map->len() > _map->req()) { // _map->has_exceptions()
510 Node* ex = _map->in(_map->req()); // _map->next_exception()
511 // skip the first one; it's already being printed
512 while (ex != NULL && ex->len() > ex->req()) {
513 ex = ex->in(ex->req()); // ex->next_exception()
514 ex->dump(1);
515 }
516 }
517 _map->dump(2);
518 }
519 st->print("JVMS depth=%d loc=%d stk=%d mon=%d scalar=%d end=%d mondepth=%d sp=%d bci=%d reexecute=%s method=",
520 depth(), locoff(), stkoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false");
521 if (_method == NULL) {
522 st->print_cr("(none)");
523 } else {
524 _method->print_name(st);
525 st->cr();
526 if (bci() >= 0 && bci() < _method->code_size()) {
527 st->print(" bc: ");
528 _method->print_codes_on(bci(), bci()+1, st);
529 }
530 }
531 if (caller() != NULL) {
532 caller()->dump_on(st);
533 }
534 }
536 // Extra way to dump a jvms from the debugger,
537 // to avoid a bug with C++ member function calls.
538 void dump_jvms(JVMState* jvms) {
539 jvms->dump();
540 }
541 #endif
543 //--------------------------clone_shallow--------------------------------------
544 JVMState* JVMState::clone_shallow(Compile* C) const {
545 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
546 n->set_bci(_bci);
547 n->_reexecute = _reexecute;
548 n->set_locoff(_locoff);
549 n->set_stkoff(_stkoff);
550 n->set_monoff(_monoff);
551 n->set_scloff(_scloff);
552 n->set_endoff(_endoff);
553 n->set_sp(_sp);
554 n->set_map(_map);
555 return n;
556 }
558 //---------------------------clone_deep----------------------------------------
559 JVMState* JVMState::clone_deep(Compile* C) const {
560 JVMState* n = clone_shallow(C);
561 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) {
562 p->_caller = p->_caller->clone_shallow(C);
563 }
564 assert(n->depth() == depth(), "sanity");
565 assert(n->debug_depth() == debug_depth(), "sanity");
566 return n;
567 }
569 //=============================================================================
570 uint CallNode::cmp( const Node &n ) const
571 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
572 #ifndef PRODUCT
573 void CallNode::dump_req() const {
574 // Dump the required inputs, enclosed in '(' and ')'
575 uint i; // Exit value of loop
576 for( i=0; i<req(); i++ ) { // For all required inputs
577 if( i == TypeFunc::Parms ) tty->print("(");
578 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
579 else tty->print("_ ");
580 }
581 tty->print(")");
582 }
584 void CallNode::dump_spec(outputStream *st) const {
585 st->print(" ");
586 tf()->dump_on(st);
587 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt);
588 if (jvms() != NULL) jvms()->dump_spec(st);
589 }
590 #endif
592 const Type *CallNode::bottom_type() const { return tf()->range(); }
593 const Type *CallNode::Value(PhaseTransform *phase) const {
594 if (phase->type(in(0)) == Type::TOP) return Type::TOP;
595 return tf()->range();
596 }
598 //------------------------------calling_convention-----------------------------
599 void CallNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
600 // Use the standard compiler calling convention
601 Matcher::calling_convention( sig_bt, parm_regs, argcnt, true );
602 }
605 //------------------------------match------------------------------------------
606 // Construct projections for control, I/O, memory-fields, ..., and
607 // return result(s) along with their RegMask info
608 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
609 switch (proj->_con) {
610 case TypeFunc::Control:
611 case TypeFunc::I_O:
612 case TypeFunc::Memory:
613 return new (match->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
615 case TypeFunc::Parms+1: // For LONG & DOUBLE returns
616 assert(tf()->_range->field_at(TypeFunc::Parms+1) == Type::HALF, "");
617 // 2nd half of doubles and longs
618 return new (match->C, 1) MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad);
620 case TypeFunc::Parms: { // Normal returns
621 uint ideal_reg = Matcher::base2reg[tf()->range()->field_at(TypeFunc::Parms)->base()];
622 OptoRegPair regs = is_CallRuntime()
623 ? match->c_return_value(ideal_reg,true) // Calls into C runtime
624 : match-> return_value(ideal_reg,true); // Calls into compiled Java code
625 RegMask rm = RegMask(regs.first());
626 if( OptoReg::is_valid(regs.second()) )
627 rm.Insert( regs.second() );
628 return new (match->C, 1) MachProjNode(this,proj->_con,rm,ideal_reg);
629 }
631 case TypeFunc::ReturnAdr:
632 case TypeFunc::FramePtr:
633 default:
634 ShouldNotReachHere();
635 }
636 return NULL;
637 }
639 // Do we Match on this edge index or not? Match no edges
640 uint CallNode::match_edge(uint idx) const {
641 return 0;
642 }
644 //
645 // Determine whether the call could modify the field of the specified
646 // instance at the specified offset.
647 //
648 bool CallNode::may_modify(const TypePtr *addr_t, PhaseTransform *phase) {
649 const TypeOopPtr *adrInst_t = addr_t->isa_oopptr();
651 // If not an OopPtr or not an instance type, assume the worst.
652 // Note: currently this method is called only for instance types.
653 if (adrInst_t == NULL || !adrInst_t->is_known_instance()) {
654 return true;
655 }
656 // The instance_id is set only for scalar-replaceable allocations which
657 // are not passed as arguments according to Escape Analysis.
658 return false;
659 }
661 // Does this call have a direct reference to n other than debug information?
662 bool CallNode::has_non_debug_use(Node *n) {
663 const TypeTuple * d = tf()->domain();
664 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
665 Node *arg = in(i);
666 if (arg == n) {
667 return true;
668 }
669 }
670 return false;
671 }
673 // Returns the unique CheckCastPP of a call
674 // or 'this' if there are several CheckCastPP
675 // or returns NULL if there is no one.
676 Node *CallNode::result_cast() {
677 Node *cast = NULL;
679 Node *p = proj_out(TypeFunc::Parms);
680 if (p == NULL)
681 return NULL;
683 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
684 Node *use = p->fast_out(i);
685 if (use->is_CheckCastPP()) {
686 if (cast != NULL) {
687 return this; // more than 1 CheckCastPP
688 }
689 cast = use;
690 }
691 }
692 return cast;
693 }
696 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj) {
697 projs->fallthrough_proj = NULL;
698 projs->fallthrough_catchproj = NULL;
699 projs->fallthrough_ioproj = NULL;
700 projs->catchall_ioproj = NULL;
701 projs->catchall_catchproj = NULL;
702 projs->fallthrough_memproj = NULL;
703 projs->catchall_memproj = NULL;
704 projs->resproj = NULL;
705 projs->exobj = NULL;
707 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
708 ProjNode *pn = fast_out(i)->as_Proj();
709 if (pn->outcnt() == 0) continue;
710 switch (pn->_con) {
711 case TypeFunc::Control:
712 {
713 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
714 projs->fallthrough_proj = pn;
715 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
716 const Node *cn = pn->fast_out(j);
717 if (cn->is_Catch()) {
718 ProjNode *cpn = NULL;
719 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
720 cpn = cn->fast_out(k)->as_Proj();
721 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
722 if (cpn->_con == CatchProjNode::fall_through_index)
723 projs->fallthrough_catchproj = cpn;
724 else {
725 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
726 projs->catchall_catchproj = cpn;
727 }
728 }
729 }
730 break;
731 }
732 case TypeFunc::I_O:
733 if (pn->_is_io_use)
734 projs->catchall_ioproj = pn;
735 else
736 projs->fallthrough_ioproj = pn;
737 for (DUIterator j = pn->outs(); pn->has_out(j); j++) {
738 Node* e = pn->out(j);
739 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj()) {
740 assert(projs->exobj == NULL, "only one");
741 projs->exobj = e;
742 }
743 }
744 break;
745 case TypeFunc::Memory:
746 if (pn->_is_io_use)
747 projs->catchall_memproj = pn;
748 else
749 projs->fallthrough_memproj = pn;
750 break;
751 case TypeFunc::Parms:
752 projs->resproj = pn;
753 break;
754 default:
755 assert(false, "unexpected projection from allocation node.");
756 }
757 }
759 // The resproj may not exist because the result couuld be ignored
760 // and the exception object may not exist if an exception handler
761 // swallows the exception but all the other must exist and be found.
762 assert(projs->fallthrough_proj != NULL, "must be found");
763 assert(projs->fallthrough_catchproj != NULL, "must be found");
764 assert(projs->fallthrough_memproj != NULL, "must be found");
765 assert(projs->fallthrough_ioproj != NULL, "must be found");
766 assert(projs->catchall_catchproj != NULL, "must be found");
767 if (separate_io_proj) {
768 assert(projs->catchall_memproj != NULL, "must be found");
769 assert(projs->catchall_ioproj != NULL, "must be found");
770 }
771 }
774 //=============================================================================
775 uint CallJavaNode::size_of() const { return sizeof(*this); }
776 uint CallJavaNode::cmp( const Node &n ) const {
777 CallJavaNode &call = (CallJavaNode&)n;
778 return CallNode::cmp(call) && _method == call._method;
779 }
780 #ifndef PRODUCT
781 void CallJavaNode::dump_spec(outputStream *st) const {
782 if( _method ) _method->print_short_name(st);
783 CallNode::dump_spec(st);
784 }
785 #endif
787 //=============================================================================
788 uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
789 uint CallStaticJavaNode::cmp( const Node &n ) const {
790 CallStaticJavaNode &call = (CallStaticJavaNode&)n;
791 return CallJavaNode::cmp(call);
792 }
794 //----------------------------uncommon_trap_request----------------------------
795 // If this is an uncommon trap, return the request code, else zero.
796 int CallStaticJavaNode::uncommon_trap_request() const {
797 if (_name != NULL && !strcmp(_name, "uncommon_trap")) {
798 return extract_uncommon_trap_request(this);
799 }
800 return 0;
801 }
802 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
803 #ifndef PRODUCT
804 if (!(call->req() > TypeFunc::Parms &&
805 call->in(TypeFunc::Parms) != NULL &&
806 call->in(TypeFunc::Parms)->is_Con())) {
807 assert(_in_dump_cnt != 0, "OK if dumping");
808 tty->print("[bad uncommon trap]");
809 return 0;
810 }
811 #endif
812 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
813 }
815 #ifndef PRODUCT
816 void CallStaticJavaNode::dump_spec(outputStream *st) const {
817 st->print("# Static ");
818 if (_name != NULL) {
819 st->print("%s", _name);
820 int trap_req = uncommon_trap_request();
821 if (trap_req != 0) {
822 char buf[100];
823 st->print("(%s)",
824 Deoptimization::format_trap_request(buf, sizeof(buf),
825 trap_req));
826 }
827 st->print(" ");
828 }
829 CallJavaNode::dump_spec(st);
830 }
831 #endif
833 //=============================================================================
834 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
835 uint CallDynamicJavaNode::cmp( const Node &n ) const {
836 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
837 return CallJavaNode::cmp(call);
838 }
839 #ifndef PRODUCT
840 void CallDynamicJavaNode::dump_spec(outputStream *st) const {
841 st->print("# Dynamic ");
842 CallJavaNode::dump_spec(st);
843 }
844 #endif
846 //=============================================================================
847 uint CallRuntimeNode::size_of() const { return sizeof(*this); }
848 uint CallRuntimeNode::cmp( const Node &n ) const {
849 CallRuntimeNode &call = (CallRuntimeNode&)n;
850 return CallNode::cmp(call) && !strcmp(_name,call._name);
851 }
852 #ifndef PRODUCT
853 void CallRuntimeNode::dump_spec(outputStream *st) const {
854 st->print("# ");
855 st->print(_name);
856 CallNode::dump_spec(st);
857 }
858 #endif
860 //------------------------------calling_convention-----------------------------
861 void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
862 Matcher::c_calling_convention( sig_bt, parm_regs, argcnt );
863 }
865 //=============================================================================
866 //------------------------------calling_convention-----------------------------
869 //=============================================================================
870 #ifndef PRODUCT
871 void CallLeafNode::dump_spec(outputStream *st) const {
872 st->print("# ");
873 st->print(_name);
874 CallNode::dump_spec(st);
875 }
876 #endif
878 //=============================================================================
880 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) {
881 assert(verify_jvms(jvms), "jvms must match");
882 int loc = jvms->locoff() + idx;
883 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
884 // If current local idx is top then local idx - 1 could
885 // be a long/double that needs to be killed since top could
886 // represent the 2nd half ofthe long/double.
887 uint ideal = in(loc -1)->ideal_reg();
888 if (ideal == Op_RegD || ideal == Op_RegL) {
889 // set other (low index) half to top
890 set_req(loc - 1, in(loc));
891 }
892 }
893 set_req(loc, c);
894 }
896 uint SafePointNode::size_of() const { return sizeof(*this); }
897 uint SafePointNode::cmp( const Node &n ) const {
898 return (&n == this); // Always fail except on self
899 }
901 //-------------------------set_next_exception----------------------------------
902 void SafePointNode::set_next_exception(SafePointNode* n) {
903 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception");
904 if (len() == req()) {
905 if (n != NULL) add_prec(n);
906 } else {
907 set_prec(req(), n);
908 }
909 }
912 //----------------------------next_exception-----------------------------------
913 SafePointNode* SafePointNode::next_exception() const {
914 if (len() == req()) {
915 return NULL;
916 } else {
917 Node* n = in(req());
918 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
919 return (SafePointNode*) n;
920 }
921 }
924 //------------------------------Ideal------------------------------------------
925 // Skip over any collapsed Regions
926 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
927 return remove_dead_region(phase, can_reshape) ? this : NULL;
928 }
930 //------------------------------Identity---------------------------------------
931 // Remove obviously duplicate safepoints
932 Node *SafePointNode::Identity( PhaseTransform *phase ) {
934 // If you have back to back safepoints, remove one
935 if( in(TypeFunc::Control)->is_SafePoint() )
936 return in(TypeFunc::Control);
938 if( in(0)->is_Proj() ) {
939 Node *n0 = in(0)->in(0);
940 // Check if he is a call projection (except Leaf Call)
941 if( n0->is_Catch() ) {
942 n0 = n0->in(0)->in(0);
943 assert( n0->is_Call(), "expect a call here" );
944 }
945 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
946 // Useless Safepoint, so remove it
947 return in(TypeFunc::Control);
948 }
949 }
951 return this;
952 }
954 //------------------------------Value------------------------------------------
955 const Type *SafePointNode::Value( PhaseTransform *phase ) const {
956 if( phase->type(in(0)) == Type::TOP ) return Type::TOP;
957 if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop
958 return Type::CONTROL;
959 }
961 #ifndef PRODUCT
962 void SafePointNode::dump_spec(outputStream *st) const {
963 st->print(" SafePoint ");
964 }
965 #endif
967 const RegMask &SafePointNode::in_RegMask(uint idx) const {
968 if( idx < TypeFunc::Parms ) return RegMask::Empty;
969 // Values outside the domain represent debug info
970 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
971 }
972 const RegMask &SafePointNode::out_RegMask() const {
973 return RegMask::Empty;
974 }
977 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
978 assert((int)grow_by > 0, "sanity");
979 int monoff = jvms->monoff();
980 int scloff = jvms->scloff();
981 int endoff = jvms->endoff();
982 assert(endoff == (int)req(), "no other states or debug info after me");
983 Node* top = Compile::current()->top();
984 for (uint i = 0; i < grow_by; i++) {
985 ins_req(monoff, top);
986 }
987 jvms->set_monoff(monoff + grow_by);
988 jvms->set_scloff(scloff + grow_by);
989 jvms->set_endoff(endoff + grow_by);
990 }
992 void SafePointNode::push_monitor(const FastLockNode *lock) {
993 // Add a LockNode, which points to both the original BoxLockNode (the
994 // stack space for the monitor) and the Object being locked.
995 const int MonitorEdges = 2;
996 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
997 assert(req() == jvms()->endoff(), "correct sizing");
998 int nextmon = jvms()->scloff();
999 if (GenerateSynchronizationCode) {
1000 add_req(lock->box_node());
1001 add_req(lock->obj_node());
1002 } else {
1003 Node* top = Compile::current()->top();
1004 add_req(top);
1005 add_req(top);
1006 }
1007 jvms()->set_scloff(nextmon+MonitorEdges);
1008 jvms()->set_endoff(req());
1009 }
1011 void SafePointNode::pop_monitor() {
1012 // Delete last monitor from debug info
1013 debug_only(int num_before_pop = jvms()->nof_monitors());
1014 const int MonitorEdges = (1<<JVMState::logMonitorEdges);
1015 int scloff = jvms()->scloff();
1016 int endoff = jvms()->endoff();
1017 int new_scloff = scloff - MonitorEdges;
1018 int new_endoff = endoff - MonitorEdges;
1019 jvms()->set_scloff(new_scloff);
1020 jvms()->set_endoff(new_endoff);
1021 while (scloff > new_scloff) del_req(--scloff);
1022 assert(jvms()->nof_monitors() == num_before_pop-1, "");
1023 }
1025 Node *SafePointNode::peek_monitor_box() const {
1026 int mon = jvms()->nof_monitors() - 1;
1027 assert(mon >= 0, "most have a monitor");
1028 return monitor_box(jvms(), mon);
1029 }
1031 Node *SafePointNode::peek_monitor_obj() const {
1032 int mon = jvms()->nof_monitors() - 1;
1033 assert(mon >= 0, "most have a monitor");
1034 return monitor_obj(jvms(), mon);
1035 }
1037 // Do we Match on this edge index or not? Match no edges
1038 uint SafePointNode::match_edge(uint idx) const {
1039 if( !needs_polling_address_input() )
1040 return 0;
1042 return (TypeFunc::Parms == idx);
1043 }
1045 //============== SafePointScalarObjectNode ==============
1047 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp,
1048 #ifdef ASSERT
1049 AllocateNode* alloc,
1050 #endif
1051 uint first_index,
1052 uint n_fields) :
1053 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1054 #ifdef ASSERT
1055 _alloc(alloc),
1056 #endif
1057 _first_index(first_index),
1058 _n_fields(n_fields)
1059 {
1060 init_class_id(Class_SafePointScalarObject);
1061 }
1063 bool SafePointScalarObjectNode::pinned() const { return true; }
1064 bool SafePointScalarObjectNode::depends_only_on_test() const { return false; }
1066 uint SafePointScalarObjectNode::ideal_reg() const {
1067 return 0; // No matching to machine instruction
1068 }
1070 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1071 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1072 }
1074 const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1075 return RegMask::Empty;
1076 }
1078 uint SafePointScalarObjectNode::match_edge(uint idx) const {
1079 return 0;
1080 }
1082 SafePointScalarObjectNode*
1083 SafePointScalarObjectNode::clone(int jvms_adj, Dict* sosn_map) const {
1084 void* cached = (*sosn_map)[(void*)this];
1085 if (cached != NULL) {
1086 return (SafePointScalarObjectNode*)cached;
1087 }
1088 Compile* C = Compile::current();
1089 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1090 res->_first_index += jvms_adj;
1091 sosn_map->Insert((void*)this, (void*)res);
1092 return res;
1093 }
1096 #ifndef PRODUCT
1097 void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1098 st->print(" # fields@[%d..%d]", first_index(),
1099 first_index() + n_fields() - 1);
1100 }
1102 #endif
1104 //=============================================================================
1105 uint AllocateNode::size_of() const { return sizeof(*this); }
1107 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1108 Node *ctrl, Node *mem, Node *abio,
1109 Node *size, Node *klass_node, Node *initial_test)
1110 : CallNode(atype, NULL, TypeRawPtr::BOTTOM)
1111 {
1112 init_class_id(Class_Allocate);
1113 init_flags(Flag_is_macro);
1114 _is_scalar_replaceable = false;
1115 Node *topnode = C->top();
1117 init_req( TypeFunc::Control , ctrl );
1118 init_req( TypeFunc::I_O , abio );
1119 init_req( TypeFunc::Memory , mem );
1120 init_req( TypeFunc::ReturnAdr, topnode );
1121 init_req( TypeFunc::FramePtr , topnode );
1122 init_req( AllocSize , size);
1123 init_req( KlassNode , klass_node);
1124 init_req( InitialTest , initial_test);
1125 init_req( ALength , topnode);
1126 C->add_macro_node(this);
1127 }
1129 //=============================================================================
1130 uint AllocateArrayNode::size_of() const { return sizeof(*this); }
1132 Node* AllocateArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1133 if (remove_dead_region(phase, can_reshape)) return this;
1135 const Type* type = phase->type(Ideal_length());
1136 if (type->isa_int() && type->is_int()->_hi < 0) {
1137 if (can_reshape) {
1138 PhaseIterGVN *igvn = phase->is_IterGVN();
1139 // Unreachable fall through path (negative array length),
1140 // the allocation can only throw so disconnect it.
1141 Node* proj = proj_out(TypeFunc::Control);
1142 Node* catchproj = NULL;
1143 if (proj != NULL) {
1144 for (DUIterator_Fast imax, i = proj->fast_outs(imax); i < imax; i++) {
1145 Node *cn = proj->fast_out(i);
1146 if (cn->is_Catch()) {
1147 catchproj = cn->as_Multi()->proj_out(CatchProjNode::fall_through_index);
1148 break;
1149 }
1150 }
1151 }
1152 if (catchproj != NULL && catchproj->outcnt() > 0 &&
1153 (catchproj->outcnt() > 1 ||
1154 catchproj->unique_out()->Opcode() != Op_Halt)) {
1155 assert(catchproj->is_CatchProj(), "must be a CatchProjNode");
1156 Node* nproj = catchproj->clone();
1157 igvn->register_new_node_with_optimizer(nproj);
1159 Node *frame = new (phase->C, 1) ParmNode( phase->C->start(), TypeFunc::FramePtr );
1160 frame = phase->transform(frame);
1161 // Halt & Catch Fire
1162 Node *halt = new (phase->C, TypeFunc::Parms) HaltNode( nproj, frame );
1163 phase->C->root()->add_req(halt);
1164 phase->transform(halt);
1166 igvn->replace_node(catchproj, phase->C->top());
1167 return this;
1168 }
1169 } else {
1170 // Can't correct it during regular GVN so register for IGVN
1171 phase->C->record_for_igvn(this);
1172 }
1173 }
1174 return NULL;
1175 }
1177 // Retrieve the length from the AllocateArrayNode. Narrow the type with a
1178 // CastII, if appropriate. If we are not allowed to create new nodes, and
1179 // a CastII is appropriate, return NULL.
1180 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) {
1181 Node *length = in(AllocateNode::ALength);
1182 assert(length != NULL, "length is not null");
1184 const TypeInt* length_type = phase->find_int_type(length);
1185 const TypeAryPtr* ary_type = oop_type->isa_aryptr();
1187 if (ary_type != NULL && length_type != NULL) {
1188 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type);
1189 if (narrow_length_type != length_type) {
1190 // Assert one of:
1191 // - the narrow_length is 0
1192 // - the narrow_length is not wider than length
1193 assert(narrow_length_type == TypeInt::ZERO ||
1194 (narrow_length_type->_hi <= length_type->_hi &&
1195 narrow_length_type->_lo >= length_type->_lo),
1196 "narrow type must be narrower than length type");
1198 // Return NULL if new nodes are not allowed
1199 if (!allow_new_nodes) return NULL;
1200 // Create a cast which is control dependent on the initialization to
1201 // propagate the fact that the array length must be positive.
1202 length = new (phase->C, 2) CastIINode(length, narrow_length_type);
1203 length->set_req(0, initialization()->proj_out(0));
1204 }
1205 }
1207 return length;
1208 }
1210 //=============================================================================
1211 uint LockNode::size_of() const { return sizeof(*this); }
1213 // Redundant lock elimination
1214 //
1215 // There are various patterns of locking where we release and
1216 // immediately reacquire a lock in a piece of code where no operations
1217 // occur in between that would be observable. In those cases we can
1218 // skip releasing and reacquiring the lock without violating any
1219 // fairness requirements. Doing this around a loop could cause a lock
1220 // to be held for a very long time so we concentrate on non-looping
1221 // control flow. We also require that the operations are fully
1222 // redundant meaning that we don't introduce new lock operations on
1223 // some paths so to be able to eliminate it on others ala PRE. This
1224 // would probably require some more extensive graph manipulation to
1225 // guarantee that the memory edges were all handled correctly.
1226 //
1227 // Assuming p is a simple predicate which can't trap in any way and s
1228 // is a synchronized method consider this code:
1229 //
1230 // s();
1231 // if (p)
1232 // s();
1233 // else
1234 // s();
1235 // s();
1236 //
1237 // 1. The unlocks of the first call to s can be eliminated if the
1238 // locks inside the then and else branches are eliminated.
1239 //
1240 // 2. The unlocks of the then and else branches can be eliminated if
1241 // the lock of the final call to s is eliminated.
1242 //
1243 // Either of these cases subsumes the simple case of sequential control flow
1244 //
1245 // Addtionally we can eliminate versions without the else case:
1246 //
1247 // s();
1248 // if (p)
1249 // s();
1250 // s();
1251 //
1252 // 3. In this case we eliminate the unlock of the first s, the lock
1253 // and unlock in the then case and the lock in the final s.
1254 //
1255 // Note also that in all these cases the then/else pieces don't have
1256 // to be trivial as long as they begin and end with synchronization
1257 // operations.
1258 //
1259 // s();
1260 // if (p)
1261 // s();
1262 // f();
1263 // s();
1264 // s();
1265 //
1266 // The code will work properly for this case, leaving in the unlock
1267 // before the call to f and the relock after it.
1268 //
1269 // A potentially interesting case which isn't handled here is when the
1270 // locking is partially redundant.
1271 //
1272 // s();
1273 // if (p)
1274 // s();
1275 //
1276 // This could be eliminated putting unlocking on the else case and
1277 // eliminating the first unlock and the lock in the then side.
1278 // Alternatively the unlock could be moved out of the then side so it
1279 // was after the merge and the first unlock and second lock
1280 // eliminated. This might require less manipulation of the memory
1281 // state to get correct.
1282 //
1283 // Additionally we might allow work between a unlock and lock before
1284 // giving up eliminating the locks. The current code disallows any
1285 // conditional control flow between these operations. A formulation
1286 // similar to partial redundancy elimination computing the
1287 // availability of unlocking and the anticipatability of locking at a
1288 // program point would allow detection of fully redundant locking with
1289 // some amount of work in between. I'm not sure how often I really
1290 // think that would occur though. Most of the cases I've seen
1291 // indicate it's likely non-trivial work would occur in between.
1292 // There may be other more complicated constructs where we could
1293 // eliminate locking but I haven't seen any others appear as hot or
1294 // interesting.
1295 //
1296 // Locking and unlocking have a canonical form in ideal that looks
1297 // roughly like this:
1298 //
1299 // <obj>
1300 // | \\------+
1301 // | \ \
1302 // | BoxLock \
1303 // | | | \
1304 // | | \ \
1305 // | | FastLock
1306 // | | /
1307 // | | /
1308 // | | |
1309 //
1310 // Lock
1311 // |
1312 // Proj #0
1313 // |
1314 // MembarAcquire
1315 // |
1316 // Proj #0
1317 //
1318 // MembarRelease
1319 // |
1320 // Proj #0
1321 // |
1322 // Unlock
1323 // |
1324 // Proj #0
1325 //
1326 //
1327 // This code proceeds by processing Lock nodes during PhaseIterGVN
1328 // and searching back through its control for the proper code
1329 // patterns. Once it finds a set of lock and unlock operations to
1330 // eliminate they are marked as eliminatable which causes the
1331 // expansion of the Lock and Unlock macro nodes to make the operation a NOP
1332 //
1333 //=============================================================================
1335 //
1336 // Utility function to skip over uninteresting control nodes. Nodes skipped are:
1337 // - copy regions. (These may not have been optimized away yet.)
1338 // - eliminated locking nodes
1339 //
1340 static Node *next_control(Node *ctrl) {
1341 if (ctrl == NULL)
1342 return NULL;
1343 while (1) {
1344 if (ctrl->is_Region()) {
1345 RegionNode *r = ctrl->as_Region();
1346 Node *n = r->is_copy();
1347 if (n == NULL)
1348 break; // hit a region, return it
1349 else
1350 ctrl = n;
1351 } else if (ctrl->is_Proj()) {
1352 Node *in0 = ctrl->in(0);
1353 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
1354 ctrl = in0->in(0);
1355 } else {
1356 break;
1357 }
1358 } else {
1359 break; // found an interesting control
1360 }
1361 }
1362 return ctrl;
1363 }
1364 //
1365 // Given a control, see if it's the control projection of an Unlock which
1366 // operating on the same object as lock.
1367 //
1368 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
1369 GrowableArray<AbstractLockNode*> &lock_ops) {
1370 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL;
1371 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) {
1372 Node *n = ctrl_proj->in(0);
1373 if (n != NULL && n->is_Unlock()) {
1374 UnlockNode *unlock = n->as_Unlock();
1375 if ((lock->obj_node() == unlock->obj_node()) &&
1376 (lock->box_node() == unlock->box_node()) && !unlock->is_eliminated()) {
1377 lock_ops.append(unlock);
1378 return true;
1379 }
1380 }
1381 }
1382 return false;
1383 }
1385 //
1386 // Find the lock matching an unlock. Returns null if a safepoint
1387 // or complicated control is encountered first.
1388 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
1389 LockNode *lock_result = NULL;
1390 // find the matching lock, or an intervening safepoint
1391 Node *ctrl = next_control(unlock->in(0));
1392 while (1) {
1393 assert(ctrl != NULL, "invalid control graph");
1394 assert(!ctrl->is_Start(), "missing lock for unlock");
1395 if (ctrl->is_top()) break; // dead control path
1396 if (ctrl->is_Proj()) ctrl = ctrl->in(0);
1397 if (ctrl->is_SafePoint()) {
1398 break; // found a safepoint (may be the lock we are searching for)
1399 } else if (ctrl->is_Region()) {
1400 // Check for a simple diamond pattern. Punt on anything more complicated
1401 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) {
1402 Node *in1 = next_control(ctrl->in(1));
1403 Node *in2 = next_control(ctrl->in(2));
1404 if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
1405 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
1406 ctrl = next_control(in1->in(0)->in(0));
1407 } else {
1408 break;
1409 }
1410 } else {
1411 break;
1412 }
1413 } else {
1414 ctrl = next_control(ctrl->in(0)); // keep searching
1415 }
1416 }
1417 if (ctrl->is_Lock()) {
1418 LockNode *lock = ctrl->as_Lock();
1419 if ((lock->obj_node() == unlock->obj_node()) &&
1420 (lock->box_node() == unlock->box_node())) {
1421 lock_result = lock;
1422 }
1423 }
1424 return lock_result;
1425 }
1427 // This code corresponds to case 3 above.
1429 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
1430 GrowableArray<AbstractLockNode*> &lock_ops) {
1431 Node* if_node = node->in(0);
1432 bool if_true = node->is_IfTrue();
1434 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
1435 Node *lock_ctrl = next_control(if_node->in(0));
1436 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
1437 Node* lock1_node = NULL;
1438 ProjNode* proj = if_node->as_If()->proj_out(!if_true);
1439 if (if_true) {
1440 if (proj->is_IfFalse() && proj->outcnt() == 1) {
1441 lock1_node = proj->unique_out();
1442 }
1443 } else {
1444 if (proj->is_IfTrue() && proj->outcnt() == 1) {
1445 lock1_node = proj->unique_out();
1446 }
1447 }
1448 if (lock1_node != NULL && lock1_node->is_Lock()) {
1449 LockNode *lock1 = lock1_node->as_Lock();
1450 if ((lock->obj_node() == lock1->obj_node()) &&
1451 (lock->box_node() == lock1->box_node()) && !lock1->is_eliminated()) {
1452 lock_ops.append(lock1);
1453 return true;
1454 }
1455 }
1456 }
1457 }
1459 lock_ops.trunc_to(0);
1460 return false;
1461 }
1463 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
1464 GrowableArray<AbstractLockNode*> &lock_ops) {
1465 // check each control merging at this point for a matching unlock.
1466 // in(0) should be self edge so skip it.
1467 for (int i = 1; i < (int)region->req(); i++) {
1468 Node *in_node = next_control(region->in(i));
1469 if (in_node != NULL) {
1470 if (find_matching_unlock(in_node, lock, lock_ops)) {
1471 // found a match so keep on checking.
1472 continue;
1473 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
1474 continue;
1475 }
1477 // If we fall through to here then it was some kind of node we
1478 // don't understand or there wasn't a matching unlock, so give
1479 // up trying to merge locks.
1480 lock_ops.trunc_to(0);
1481 return false;
1482 }
1483 }
1484 return true;
1486 }
1488 #ifndef PRODUCT
1489 //
1490 // Create a counter which counts the number of times this lock is acquired
1491 //
1492 void AbstractLockNode::create_lock_counter(JVMState* state) {
1493 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
1494 }
1495 #endif
1497 void AbstractLockNode::set_eliminated() {
1498 _eliminate = true;
1499 #ifndef PRODUCT
1500 if (_counter) {
1501 // Update the counter to indicate that this lock was eliminated.
1502 // The counter update code will stay around even though the
1503 // optimizer will eliminate the lock operation itself.
1504 _counter->set_tag(NamedCounter::EliminatedLockCounter);
1505 }
1506 #endif
1507 }
1509 //=============================================================================
1510 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1512 // perform any generic optimizations first (returns 'this' or NULL)
1513 Node *result = SafePointNode::Ideal(phase, can_reshape);
1515 // Now see if we can optimize away this lock. We don't actually
1516 // remove the locking here, we simply set the _eliminate flag which
1517 // prevents macro expansion from expanding the lock. Since we don't
1518 // modify the graph, the value returned from this function is the
1519 // one computed above.
1520 if (result == NULL && can_reshape && EliminateLocks && !is_eliminated()) {
1521 //
1522 // If we are locking an unescaped object, the lock/unlock is unnecessary
1523 //
1524 ConnectionGraph *cgr = phase->C->congraph();
1525 PointsToNode::EscapeState es = PointsToNode::GlobalEscape;
1526 if (cgr != NULL)
1527 es = cgr->escape_state(obj_node(), phase);
1528 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1529 // Mark it eliminated to update any counters
1530 this->set_eliminated();
1531 return result;
1532 }
1534 //
1535 // Try lock coarsening
1536 //
1537 PhaseIterGVN* iter = phase->is_IterGVN();
1538 if (iter != NULL) {
1540 GrowableArray<AbstractLockNode*> lock_ops;
1542 Node *ctrl = next_control(in(0));
1544 // now search back for a matching Unlock
1545 if (find_matching_unlock(ctrl, this, lock_ops)) {
1546 // found an unlock directly preceding this lock. This is the
1547 // case of single unlock directly control dependent on a
1548 // single lock which is the trivial version of case 1 or 2.
1549 } else if (ctrl->is_Region() ) {
1550 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
1551 // found lock preceded by multiple unlocks along all paths
1552 // joining at this point which is case 3 in description above.
1553 }
1554 } else {
1555 // see if this lock comes from either half of an if and the
1556 // predecessors merges unlocks and the other half of the if
1557 // performs a lock.
1558 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
1559 // found unlock splitting to an if with locks on both branches.
1560 }
1561 }
1563 if (lock_ops.length() > 0) {
1564 // add ourselves to the list of locks to be eliminated.
1565 lock_ops.append(this);
1567 #ifndef PRODUCT
1568 if (PrintEliminateLocks) {
1569 int locks = 0;
1570 int unlocks = 0;
1571 for (int i = 0; i < lock_ops.length(); i++) {
1572 AbstractLockNode* lock = lock_ops.at(i);
1573 if (lock->Opcode() == Op_Lock)
1574 locks++;
1575 else
1576 unlocks++;
1577 if (Verbose) {
1578 lock->dump(1);
1579 }
1580 }
1581 tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks);
1582 }
1583 #endif
1585 // for each of the identified locks, mark them
1586 // as eliminatable
1587 for (int i = 0; i < lock_ops.length(); i++) {
1588 AbstractLockNode* lock = lock_ops.at(i);
1590 // Mark it eliminated to update any counters
1591 lock->set_eliminated();
1592 lock->set_coarsened();
1593 }
1594 } else if (result != NULL && ctrl->is_Region() &&
1595 iter->_worklist.member(ctrl)) {
1596 // We weren't able to find any opportunities but the region this
1597 // lock is control dependent on hasn't been processed yet so put
1598 // this lock back on the worklist so we can check again once any
1599 // region simplification has occurred.
1600 iter->_worklist.push(this);
1601 }
1602 }
1603 }
1605 return result;
1606 }
1608 //=============================================================================
1609 uint UnlockNode::size_of() const { return sizeof(*this); }
1611 //=============================================================================
1612 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1614 // perform any generic optimizations first (returns 'this' or NULL)
1615 Node * result = SafePointNode::Ideal(phase, can_reshape);
1617 // Now see if we can optimize away this unlock. We don't actually
1618 // remove the unlocking here, we simply set the _eliminate flag which
1619 // prevents macro expansion from expanding the unlock. Since we don't
1620 // modify the graph, the value returned from this function is the
1621 // one computed above.
1622 // Escape state is defined after Parse phase.
1623 if (result == NULL && can_reshape && EliminateLocks && !is_eliminated()) {
1624 //
1625 // If we are unlocking an unescaped object, the lock/unlock is unnecessary.
1626 //
1627 ConnectionGraph *cgr = phase->C->congraph();
1628 PointsToNode::EscapeState es = PointsToNode::GlobalEscape;
1629 if (cgr != NULL)
1630 es = cgr->escape_state(obj_node(), phase);
1631 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1632 // Mark it eliminated to update any counters
1633 this->set_eliminated();
1634 }
1635 }
1636 return result;
1637 }