Thu, 16 Feb 2012 17:12:49 -0800
7145346: VerifyStackAtCalls is broken
Summary: Replace call_epilog() encoding with macroassembler use. Moved duplicated code to x86.ad. Fixed return_addr() definition.
Reviewed-by: never
1 /*
2 * Copyright (c) 1997, 2010, 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 */
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/oopMap.hpp"
28 #include "opto/callnode.hpp"
29 #include "opto/escape.hpp"
30 #include "opto/locknode.hpp"
31 #include "opto/machnode.hpp"
32 #include "opto/matcher.hpp"
33 #include "opto/parse.hpp"
34 #include "opto/regalloc.hpp"
35 #include "opto/regmask.hpp"
36 #include "opto/rootnode.hpp"
37 #include "opto/runtime.hpp"
39 // Portions of code courtesy of Clifford Click
41 // Optimization - Graph Style
43 //=============================================================================
44 uint StartNode::size_of() const { return sizeof(*this); }
45 uint StartNode::cmp( const Node &n ) const
46 { return _domain == ((StartNode&)n)._domain; }
47 const Type *StartNode::bottom_type() const { return _domain; }
48 const Type *StartNode::Value(PhaseTransform *phase) const { return _domain; }
49 #ifndef PRODUCT
50 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);}
51 #endif
53 //------------------------------Ideal------------------------------------------
54 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){
55 return remove_dead_region(phase, can_reshape) ? this : NULL;
56 }
58 //------------------------------calling_convention-----------------------------
59 void StartNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
60 Matcher::calling_convention( sig_bt, parm_regs, argcnt, false );
61 }
63 //------------------------------Registers--------------------------------------
64 const RegMask &StartNode::in_RegMask(uint) const {
65 return RegMask::Empty;
66 }
68 //------------------------------match------------------------------------------
69 // Construct projections for incoming parameters, and their RegMask info
70 Node *StartNode::match( const ProjNode *proj, const Matcher *match ) {
71 switch (proj->_con) {
72 case TypeFunc::Control:
73 case TypeFunc::I_O:
74 case TypeFunc::Memory:
75 return new (match->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
76 case TypeFunc::FramePtr:
77 return new (match->C, 1) MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP);
78 case TypeFunc::ReturnAdr:
79 return new (match->C, 1) MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP);
80 case TypeFunc::Parms:
81 default: {
82 uint parm_num = proj->_con - TypeFunc::Parms;
83 const Type *t = _domain->field_at(proj->_con);
84 if (t->base() == Type::Half) // 2nd half of Longs and Doubles
85 return new (match->C, 1) ConNode(Type::TOP);
86 uint ideal_reg = Matcher::base2reg[t->base()];
87 RegMask &rm = match->_calling_convention_mask[parm_num];
88 return new (match->C, 1) MachProjNode(this,proj->_con,rm,ideal_reg);
89 }
90 }
91 return NULL;
92 }
94 //------------------------------StartOSRNode----------------------------------
95 // The method start node for an on stack replacement adapter
97 //------------------------------osr_domain-----------------------------
98 const TypeTuple *StartOSRNode::osr_domain() {
99 const Type **fields = TypeTuple::fields(2);
100 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer
102 return TypeTuple::make(TypeFunc::Parms+1, fields);
103 }
105 //=============================================================================
106 const char * const ParmNode::names[TypeFunc::Parms+1] = {
107 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms"
108 };
110 #ifndef PRODUCT
111 void ParmNode::dump_spec(outputStream *st) const {
112 if( _con < TypeFunc::Parms ) {
113 st->print(names[_con]);
114 } else {
115 st->print("Parm%d: ",_con-TypeFunc::Parms);
116 // Verbose and WizardMode dump bottom_type for all nodes
117 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st);
118 }
119 }
120 #endif
122 uint ParmNode::ideal_reg() const {
123 switch( _con ) {
124 case TypeFunc::Control : // fall through
125 case TypeFunc::I_O : // fall through
126 case TypeFunc::Memory : return 0;
127 case TypeFunc::FramePtr : // fall through
128 case TypeFunc::ReturnAdr: return Op_RegP;
129 default : assert( _con > TypeFunc::Parms, "" );
130 // fall through
131 case TypeFunc::Parms : {
132 // Type of argument being passed
133 const Type *t = in(0)->as_Start()->_domain->field_at(_con);
134 return Matcher::base2reg[t->base()];
135 }
136 }
137 ShouldNotReachHere();
138 return 0;
139 }
141 //=============================================================================
142 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) {
143 init_req(TypeFunc::Control,cntrl);
144 init_req(TypeFunc::I_O,i_o);
145 init_req(TypeFunc::Memory,memory);
146 init_req(TypeFunc::FramePtr,frameptr);
147 init_req(TypeFunc::ReturnAdr,retadr);
148 }
150 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){
151 return remove_dead_region(phase, can_reshape) ? this : NULL;
152 }
154 const Type *ReturnNode::Value( PhaseTransform *phase ) const {
155 return ( phase->type(in(TypeFunc::Control)) == Type::TOP)
156 ? Type::TOP
157 : Type::BOTTOM;
158 }
160 // Do we Match on this edge index or not? No edges on return nodes
161 uint ReturnNode::match_edge(uint idx) const {
162 return 0;
163 }
166 #ifndef PRODUCT
167 void ReturnNode::dump_req() const {
168 // Dump the required inputs, enclosed in '(' and ')'
169 uint i; // Exit value of loop
170 for( i=0; i<req(); i++ ) { // For all required inputs
171 if( i == TypeFunc::Parms ) tty->print("returns");
172 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
173 else tty->print("_ ");
174 }
175 }
176 #endif
178 //=============================================================================
179 RethrowNode::RethrowNode(
180 Node* cntrl,
181 Node* i_o,
182 Node* memory,
183 Node* frameptr,
184 Node* ret_adr,
185 Node* exception
186 ) : Node(TypeFunc::Parms + 1) {
187 init_req(TypeFunc::Control , cntrl );
188 init_req(TypeFunc::I_O , i_o );
189 init_req(TypeFunc::Memory , memory );
190 init_req(TypeFunc::FramePtr , frameptr );
191 init_req(TypeFunc::ReturnAdr, ret_adr);
192 init_req(TypeFunc::Parms , exception);
193 }
195 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){
196 return remove_dead_region(phase, can_reshape) ? this : NULL;
197 }
199 const Type *RethrowNode::Value( PhaseTransform *phase ) const {
200 return (phase->type(in(TypeFunc::Control)) == Type::TOP)
201 ? Type::TOP
202 : Type::BOTTOM;
203 }
205 uint RethrowNode::match_edge(uint idx) const {
206 return 0;
207 }
209 #ifndef PRODUCT
210 void RethrowNode::dump_req() const {
211 // Dump the required inputs, enclosed in '(' and ')'
212 uint i; // Exit value of loop
213 for( i=0; i<req(); i++ ) { // For all required inputs
214 if( i == TypeFunc::Parms ) tty->print("exception");
215 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
216 else tty->print("_ ");
217 }
218 }
219 #endif
221 //=============================================================================
222 // Do we Match on this edge index or not? Match only target address & method
223 uint TailCallNode::match_edge(uint idx) const {
224 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
225 }
227 //=============================================================================
228 // Do we Match on this edge index or not? Match only target address & oop
229 uint TailJumpNode::match_edge(uint idx) const {
230 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
231 }
233 //=============================================================================
234 JVMState::JVMState(ciMethod* method, JVMState* caller) {
235 assert(method != NULL, "must be valid call site");
236 _method = method;
237 _reexecute = Reexecute_Undefined;
238 debug_only(_bci = -99); // random garbage value
239 debug_only(_map = (SafePointNode*)-1);
240 _caller = caller;
241 _depth = 1 + (caller == NULL ? 0 : caller->depth());
242 _locoff = TypeFunc::Parms;
243 _stkoff = _locoff + _method->max_locals();
244 _monoff = _stkoff + _method->max_stack();
245 _scloff = _monoff;
246 _endoff = _monoff;
247 _sp = 0;
248 }
249 JVMState::JVMState(int stack_size) {
250 _method = NULL;
251 _bci = InvocationEntryBci;
252 _reexecute = Reexecute_Undefined;
253 debug_only(_map = (SafePointNode*)-1);
254 _caller = NULL;
255 _depth = 1;
256 _locoff = TypeFunc::Parms;
257 _stkoff = _locoff;
258 _monoff = _stkoff + stack_size;
259 _scloff = _monoff;
260 _endoff = _monoff;
261 _sp = 0;
262 }
264 //--------------------------------of_depth-------------------------------------
265 JVMState* JVMState::of_depth(int d) const {
266 const JVMState* jvmp = this;
267 assert(0 < d && (uint)d <= depth(), "oob");
268 for (int skip = depth() - d; skip > 0; skip--) {
269 jvmp = jvmp->caller();
270 }
271 assert(jvmp->depth() == (uint)d, "found the right one");
272 return (JVMState*)jvmp;
273 }
275 //-----------------------------same_calls_as-----------------------------------
276 bool JVMState::same_calls_as(const JVMState* that) const {
277 if (this == that) return true;
278 if (this->depth() != that->depth()) return false;
279 const JVMState* p = this;
280 const JVMState* q = that;
281 for (;;) {
282 if (p->_method != q->_method) return false;
283 if (p->_method == NULL) return true; // bci is irrelevant
284 if (p->_bci != q->_bci) return false;
285 if (p->_reexecute != q->_reexecute) return false;
286 p = p->caller();
287 q = q->caller();
288 if (p == q) return true;
289 assert(p != NULL && q != NULL, "depth check ensures we don't run off end");
290 }
291 }
293 //------------------------------debug_start------------------------------------
294 uint JVMState::debug_start() const {
295 debug_only(JVMState* jvmroot = of_depth(1));
296 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
297 return of_depth(1)->locoff();
298 }
300 //-------------------------------debug_end-------------------------------------
301 uint JVMState::debug_end() const {
302 debug_only(JVMState* jvmroot = of_depth(1));
303 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
304 return endoff();
305 }
307 //------------------------------debug_depth------------------------------------
308 uint JVMState::debug_depth() const {
309 uint total = 0;
310 for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) {
311 total += jvmp->debug_size();
312 }
313 return total;
314 }
316 #ifndef PRODUCT
318 //------------------------------format_helper----------------------------------
319 // Given an allocation (a Chaitin object) and a Node decide if the Node carries
320 // any defined value or not. If it does, print out the register or constant.
321 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
322 if (n == NULL) { st->print(" NULL"); return; }
323 if (n->is_SafePointScalarObject()) {
324 // Scalar replacement.
325 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
326 scobjs->append_if_missing(spobj);
327 int sco_n = scobjs->find(spobj);
328 assert(sco_n >= 0, "");
329 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
330 return;
331 }
332 if( OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
333 char buf[50];
334 regalloc->dump_register(n,buf);
335 st->print(" %s%d]=%s",msg,i,buf);
336 } else { // No register, but might be constant
337 const Type *t = n->bottom_type();
338 switch (t->base()) {
339 case Type::Int:
340 st->print(" %s%d]=#"INT32_FORMAT,msg,i,t->is_int()->get_con());
341 break;
342 case Type::AnyPtr:
343 assert( t == TypePtr::NULL_PTR, "" );
344 st->print(" %s%d]=#NULL",msg,i);
345 break;
346 case Type::AryPtr:
347 case Type::KlassPtr:
348 case Type::InstPtr:
349 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->isa_oopptr()->const_oop());
350 break;
351 case Type::NarrowOop:
352 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->make_ptr()->isa_oopptr()->const_oop());
353 break;
354 case Type::RawPtr:
355 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,t->is_rawptr());
356 break;
357 case Type::DoubleCon:
358 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
359 break;
360 case Type::FloatCon:
361 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
362 break;
363 case Type::Long:
364 st->print(" %s%d]=#"INT64_FORMAT,msg,i,t->is_long()->get_con());
365 break;
366 case Type::Half:
367 case Type::Top:
368 st->print(" %s%d]=_",msg,i);
369 break;
370 default: ShouldNotReachHere();
371 }
372 }
373 }
375 //------------------------------format-----------------------------------------
376 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
377 st->print(" #");
378 if( _method ) {
379 _method->print_short_name(st);
380 st->print(" @ bci:%d ",_bci);
381 } else {
382 st->print_cr(" runtime stub ");
383 return;
384 }
385 if (n->is_MachSafePoint()) {
386 GrowableArray<SafePointScalarObjectNode*> scobjs;
387 MachSafePointNode *mcall = n->as_MachSafePoint();
388 uint i;
389 // Print locals
390 for( i = 0; i < (uint)loc_size(); i++ )
391 format_helper( regalloc, st, mcall->local(this, i), "L[", i, &scobjs );
392 // Print stack
393 for (i = 0; i < (uint)stk_size(); i++) {
394 if ((uint)(_stkoff + i) >= mcall->len())
395 st->print(" oob ");
396 else
397 format_helper( regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs );
398 }
399 for (i = 0; (int)i < nof_monitors(); i++) {
400 Node *box = mcall->monitor_box(this, i);
401 Node *obj = mcall->monitor_obj(this, i);
402 if ( OptoReg::is_valid(regalloc->get_reg_first(box)) ) {
403 box = BoxLockNode::box_node(box);
404 format_helper( regalloc, st, box, "MON-BOX[", i, &scobjs );
405 } else {
406 OptoReg::Name box_reg = BoxLockNode::reg(box);
407 st->print(" MON-BOX%d=%s+%d",
408 i,
409 OptoReg::regname(OptoReg::c_frame_pointer),
410 regalloc->reg2offset(box_reg));
411 }
412 const char* obj_msg = "MON-OBJ[";
413 if (EliminateLocks) {
414 if (BoxLockNode::box_node(box)->is_eliminated())
415 obj_msg = "MON-OBJ(LOCK ELIMINATED)[";
416 }
417 format_helper( regalloc, st, obj, obj_msg, i, &scobjs );
418 }
420 for (i = 0; i < (uint)scobjs.length(); i++) {
421 // Scalar replaced objects.
422 st->print_cr("");
423 st->print(" # ScObj" INT32_FORMAT " ", i);
424 SafePointScalarObjectNode* spobj = scobjs.at(i);
425 ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass();
426 assert(cik->is_instance_klass() ||
427 cik->is_array_klass(), "Not supported allocation.");
428 ciInstanceKlass *iklass = NULL;
429 if (cik->is_instance_klass()) {
430 cik->print_name_on(st);
431 iklass = cik->as_instance_klass();
432 } else if (cik->is_type_array_klass()) {
433 cik->as_array_klass()->base_element_type()->print_name_on(st);
434 st->print("[%d]", spobj->n_fields());
435 } else if (cik->is_obj_array_klass()) {
436 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass();
437 if (cie->is_instance_klass()) {
438 cie->print_name_on(st);
439 } else if (cie->is_type_array_klass()) {
440 cie->as_array_klass()->base_element_type()->print_name_on(st);
441 } else {
442 ShouldNotReachHere();
443 }
444 st->print("[%d]", spobj->n_fields());
445 int ndim = cik->as_array_klass()->dimension() - 1;
446 while (ndim-- > 0) {
447 st->print("[]");
448 }
449 }
450 st->print("={");
451 uint nf = spobj->n_fields();
452 if (nf > 0) {
453 uint first_ind = spobj->first_index();
454 Node* fld_node = mcall->in(first_ind);
455 ciField* cifield;
456 if (iklass != NULL) {
457 st->print(" [");
458 cifield = iklass->nonstatic_field_at(0);
459 cifield->print_name_on(st);
460 format_helper( regalloc, st, fld_node, ":", 0, &scobjs );
461 } else {
462 format_helper( regalloc, st, fld_node, "[", 0, &scobjs );
463 }
464 for (uint j = 1; j < nf; j++) {
465 fld_node = mcall->in(first_ind+j);
466 if (iklass != NULL) {
467 st->print(", [");
468 cifield = iklass->nonstatic_field_at(j);
469 cifield->print_name_on(st);
470 format_helper( regalloc, st, fld_node, ":", j, &scobjs );
471 } else {
472 format_helper( regalloc, st, fld_node, ", [", j, &scobjs );
473 }
474 }
475 }
476 st->print(" }");
477 }
478 }
479 st->print_cr("");
480 if (caller() != NULL) caller()->format(regalloc, n, st);
481 }
484 void JVMState::dump_spec(outputStream *st) const {
485 if (_method != NULL) {
486 bool printed = false;
487 if (!Verbose) {
488 // The JVMS dumps make really, really long lines.
489 // Take out the most boring parts, which are the package prefixes.
490 char buf[500];
491 stringStream namest(buf, sizeof(buf));
492 _method->print_short_name(&namest);
493 if (namest.count() < sizeof(buf)) {
494 const char* name = namest.base();
495 if (name[0] == ' ') ++name;
496 const char* endcn = strchr(name, ':'); // end of class name
497 if (endcn == NULL) endcn = strchr(name, '(');
498 if (endcn == NULL) endcn = name + strlen(name);
499 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
500 --endcn;
501 st->print(" %s", endcn);
502 printed = true;
503 }
504 }
505 if (!printed)
506 _method->print_short_name(st);
507 st->print(" @ bci:%d",_bci);
508 if(_reexecute == Reexecute_True)
509 st->print(" reexecute");
510 } else {
511 st->print(" runtime stub");
512 }
513 if (caller() != NULL) caller()->dump_spec(st);
514 }
517 void JVMState::dump_on(outputStream* st) const {
518 if (_map && !((uintptr_t)_map & 1)) {
519 if (_map->len() > _map->req()) { // _map->has_exceptions()
520 Node* ex = _map->in(_map->req()); // _map->next_exception()
521 // skip the first one; it's already being printed
522 while (ex != NULL && ex->len() > ex->req()) {
523 ex = ex->in(ex->req()); // ex->next_exception()
524 ex->dump(1);
525 }
526 }
527 _map->dump(2);
528 }
529 st->print("JVMS depth=%d loc=%d stk=%d mon=%d scalar=%d end=%d mondepth=%d sp=%d bci=%d reexecute=%s method=",
530 depth(), locoff(), stkoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false");
531 if (_method == NULL) {
532 st->print_cr("(none)");
533 } else {
534 _method->print_name(st);
535 st->cr();
536 if (bci() >= 0 && bci() < _method->code_size()) {
537 st->print(" bc: ");
538 _method->print_codes_on(bci(), bci()+1, st);
539 }
540 }
541 if (caller() != NULL) {
542 caller()->dump_on(st);
543 }
544 }
546 // Extra way to dump a jvms from the debugger,
547 // to avoid a bug with C++ member function calls.
548 void dump_jvms(JVMState* jvms) {
549 jvms->dump();
550 }
551 #endif
553 //--------------------------clone_shallow--------------------------------------
554 JVMState* JVMState::clone_shallow(Compile* C) const {
555 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
556 n->set_bci(_bci);
557 n->_reexecute = _reexecute;
558 n->set_locoff(_locoff);
559 n->set_stkoff(_stkoff);
560 n->set_monoff(_monoff);
561 n->set_scloff(_scloff);
562 n->set_endoff(_endoff);
563 n->set_sp(_sp);
564 n->set_map(_map);
565 return n;
566 }
568 //---------------------------clone_deep----------------------------------------
569 JVMState* JVMState::clone_deep(Compile* C) const {
570 JVMState* n = clone_shallow(C);
571 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) {
572 p->_caller = p->_caller->clone_shallow(C);
573 }
574 assert(n->depth() == depth(), "sanity");
575 assert(n->debug_depth() == debug_depth(), "sanity");
576 return n;
577 }
579 //=============================================================================
580 uint CallNode::cmp( const Node &n ) const
581 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
582 #ifndef PRODUCT
583 void CallNode::dump_req() const {
584 // Dump the required inputs, enclosed in '(' and ')'
585 uint i; // Exit value of loop
586 for( i=0; i<req(); i++ ) { // For all required inputs
587 if( i == TypeFunc::Parms ) tty->print("(");
588 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
589 else tty->print("_ ");
590 }
591 tty->print(")");
592 }
594 void CallNode::dump_spec(outputStream *st) const {
595 st->print(" ");
596 tf()->dump_on(st);
597 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt);
598 if (jvms() != NULL) jvms()->dump_spec(st);
599 }
600 #endif
602 const Type *CallNode::bottom_type() const { return tf()->range(); }
603 const Type *CallNode::Value(PhaseTransform *phase) const {
604 if (phase->type(in(0)) == Type::TOP) return Type::TOP;
605 return tf()->range();
606 }
608 //------------------------------calling_convention-----------------------------
609 void CallNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
610 // Use the standard compiler calling convention
611 Matcher::calling_convention( sig_bt, parm_regs, argcnt, true );
612 }
615 //------------------------------match------------------------------------------
616 // Construct projections for control, I/O, memory-fields, ..., and
617 // return result(s) along with their RegMask info
618 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
619 switch (proj->_con) {
620 case TypeFunc::Control:
621 case TypeFunc::I_O:
622 case TypeFunc::Memory:
623 return new (match->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
625 case TypeFunc::Parms+1: // For LONG & DOUBLE returns
626 assert(tf()->_range->field_at(TypeFunc::Parms+1) == Type::HALF, "");
627 // 2nd half of doubles and longs
628 return new (match->C, 1) MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad);
630 case TypeFunc::Parms: { // Normal returns
631 uint ideal_reg = Matcher::base2reg[tf()->range()->field_at(TypeFunc::Parms)->base()];
632 OptoRegPair regs = is_CallRuntime()
633 ? match->c_return_value(ideal_reg,true) // Calls into C runtime
634 : match-> return_value(ideal_reg,true); // Calls into compiled Java code
635 RegMask rm = RegMask(regs.first());
636 if( OptoReg::is_valid(regs.second()) )
637 rm.Insert( regs.second() );
638 return new (match->C, 1) MachProjNode(this,proj->_con,rm,ideal_reg);
639 }
641 case TypeFunc::ReturnAdr:
642 case TypeFunc::FramePtr:
643 default:
644 ShouldNotReachHere();
645 }
646 return NULL;
647 }
649 // Do we Match on this edge index or not? Match no edges
650 uint CallNode::match_edge(uint idx) const {
651 return 0;
652 }
654 //
655 // Determine whether the call could modify the field of the specified
656 // instance at the specified offset.
657 //
658 bool CallNode::may_modify(const TypePtr *addr_t, PhaseTransform *phase) {
659 const TypeOopPtr *adrInst_t = addr_t->isa_oopptr();
661 // If not an OopPtr or not an instance type, assume the worst.
662 // Note: currently this method is called only for instance types.
663 if (adrInst_t == NULL || !adrInst_t->is_known_instance()) {
664 return true;
665 }
666 // The instance_id is set only for scalar-replaceable allocations which
667 // are not passed as arguments according to Escape Analysis.
668 return false;
669 }
671 // Does this call have a direct reference to n other than debug information?
672 bool CallNode::has_non_debug_use(Node *n) {
673 const TypeTuple * d = tf()->domain();
674 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
675 Node *arg = in(i);
676 if (arg == n) {
677 return true;
678 }
679 }
680 return false;
681 }
683 // Returns the unique CheckCastPP of a call
684 // or 'this' if there are several CheckCastPP
685 // or returns NULL if there is no one.
686 Node *CallNode::result_cast() {
687 Node *cast = NULL;
689 Node *p = proj_out(TypeFunc::Parms);
690 if (p == NULL)
691 return NULL;
693 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
694 Node *use = p->fast_out(i);
695 if (use->is_CheckCastPP()) {
696 if (cast != NULL) {
697 return this; // more than 1 CheckCastPP
698 }
699 cast = use;
700 }
701 }
702 return cast;
703 }
706 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj) {
707 projs->fallthrough_proj = NULL;
708 projs->fallthrough_catchproj = NULL;
709 projs->fallthrough_ioproj = NULL;
710 projs->catchall_ioproj = NULL;
711 projs->catchall_catchproj = NULL;
712 projs->fallthrough_memproj = NULL;
713 projs->catchall_memproj = NULL;
714 projs->resproj = NULL;
715 projs->exobj = NULL;
717 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
718 ProjNode *pn = fast_out(i)->as_Proj();
719 if (pn->outcnt() == 0) continue;
720 switch (pn->_con) {
721 case TypeFunc::Control:
722 {
723 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
724 projs->fallthrough_proj = pn;
725 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
726 const Node *cn = pn->fast_out(j);
727 if (cn->is_Catch()) {
728 ProjNode *cpn = NULL;
729 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
730 cpn = cn->fast_out(k)->as_Proj();
731 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
732 if (cpn->_con == CatchProjNode::fall_through_index)
733 projs->fallthrough_catchproj = cpn;
734 else {
735 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
736 projs->catchall_catchproj = cpn;
737 }
738 }
739 }
740 break;
741 }
742 case TypeFunc::I_O:
743 if (pn->_is_io_use)
744 projs->catchall_ioproj = pn;
745 else
746 projs->fallthrough_ioproj = pn;
747 for (DUIterator j = pn->outs(); pn->has_out(j); j++) {
748 Node* e = pn->out(j);
749 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj()) {
750 assert(projs->exobj == NULL, "only one");
751 projs->exobj = e;
752 }
753 }
754 break;
755 case TypeFunc::Memory:
756 if (pn->_is_io_use)
757 projs->catchall_memproj = pn;
758 else
759 projs->fallthrough_memproj = pn;
760 break;
761 case TypeFunc::Parms:
762 projs->resproj = pn;
763 break;
764 default:
765 assert(false, "unexpected projection from allocation node.");
766 }
767 }
769 // The resproj may not exist because the result couuld be ignored
770 // and the exception object may not exist if an exception handler
771 // swallows the exception but all the other must exist and be found.
772 assert(projs->fallthrough_proj != NULL, "must be found");
773 assert(projs->fallthrough_catchproj != NULL, "must be found");
774 assert(projs->fallthrough_memproj != NULL, "must be found");
775 assert(projs->fallthrough_ioproj != NULL, "must be found");
776 assert(projs->catchall_catchproj != NULL, "must be found");
777 if (separate_io_proj) {
778 assert(projs->catchall_memproj != NULL, "must be found");
779 assert(projs->catchall_ioproj != NULL, "must be found");
780 }
781 }
784 //=============================================================================
785 uint CallJavaNode::size_of() const { return sizeof(*this); }
786 uint CallJavaNode::cmp( const Node &n ) const {
787 CallJavaNode &call = (CallJavaNode&)n;
788 return CallNode::cmp(call) && _method == call._method;
789 }
790 #ifndef PRODUCT
791 void CallJavaNode::dump_spec(outputStream *st) const {
792 if( _method ) _method->print_short_name(st);
793 CallNode::dump_spec(st);
794 }
795 #endif
797 //=============================================================================
798 uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
799 uint CallStaticJavaNode::cmp( const Node &n ) const {
800 CallStaticJavaNode &call = (CallStaticJavaNode&)n;
801 return CallJavaNode::cmp(call);
802 }
804 //----------------------------uncommon_trap_request----------------------------
805 // If this is an uncommon trap, return the request code, else zero.
806 int CallStaticJavaNode::uncommon_trap_request() const {
807 if (_name != NULL && !strcmp(_name, "uncommon_trap")) {
808 return extract_uncommon_trap_request(this);
809 }
810 return 0;
811 }
812 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
813 #ifndef PRODUCT
814 if (!(call->req() > TypeFunc::Parms &&
815 call->in(TypeFunc::Parms) != NULL &&
816 call->in(TypeFunc::Parms)->is_Con())) {
817 assert(_in_dump_cnt != 0, "OK if dumping");
818 tty->print("[bad uncommon trap]");
819 return 0;
820 }
821 #endif
822 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
823 }
825 #ifndef PRODUCT
826 void CallStaticJavaNode::dump_spec(outputStream *st) const {
827 st->print("# Static ");
828 if (_name != NULL) {
829 st->print("%s", _name);
830 int trap_req = uncommon_trap_request();
831 if (trap_req != 0) {
832 char buf[100];
833 st->print("(%s)",
834 Deoptimization::format_trap_request(buf, sizeof(buf),
835 trap_req));
836 }
837 st->print(" ");
838 }
839 CallJavaNode::dump_spec(st);
840 }
841 #endif
843 //=============================================================================
844 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
845 uint CallDynamicJavaNode::cmp( const Node &n ) const {
846 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
847 return CallJavaNode::cmp(call);
848 }
849 #ifndef PRODUCT
850 void CallDynamicJavaNode::dump_spec(outputStream *st) const {
851 st->print("# Dynamic ");
852 CallJavaNode::dump_spec(st);
853 }
854 #endif
856 //=============================================================================
857 uint CallRuntimeNode::size_of() const { return sizeof(*this); }
858 uint CallRuntimeNode::cmp( const Node &n ) const {
859 CallRuntimeNode &call = (CallRuntimeNode&)n;
860 return CallNode::cmp(call) && !strcmp(_name,call._name);
861 }
862 #ifndef PRODUCT
863 void CallRuntimeNode::dump_spec(outputStream *st) const {
864 st->print("# ");
865 st->print(_name);
866 CallNode::dump_spec(st);
867 }
868 #endif
870 //------------------------------calling_convention-----------------------------
871 void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
872 Matcher::c_calling_convention( sig_bt, parm_regs, argcnt );
873 }
875 //=============================================================================
876 //------------------------------calling_convention-----------------------------
879 //=============================================================================
880 #ifndef PRODUCT
881 void CallLeafNode::dump_spec(outputStream *st) const {
882 st->print("# ");
883 st->print(_name);
884 CallNode::dump_spec(st);
885 }
886 #endif
888 //=============================================================================
890 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) {
891 assert(verify_jvms(jvms), "jvms must match");
892 int loc = jvms->locoff() + idx;
893 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
894 // If current local idx is top then local idx - 1 could
895 // be a long/double that needs to be killed since top could
896 // represent the 2nd half ofthe long/double.
897 uint ideal = in(loc -1)->ideal_reg();
898 if (ideal == Op_RegD || ideal == Op_RegL) {
899 // set other (low index) half to top
900 set_req(loc - 1, in(loc));
901 }
902 }
903 set_req(loc, c);
904 }
906 uint SafePointNode::size_of() const { return sizeof(*this); }
907 uint SafePointNode::cmp( const Node &n ) const {
908 return (&n == this); // Always fail except on self
909 }
911 //-------------------------set_next_exception----------------------------------
912 void SafePointNode::set_next_exception(SafePointNode* n) {
913 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception");
914 if (len() == req()) {
915 if (n != NULL) add_prec(n);
916 } else {
917 set_prec(req(), n);
918 }
919 }
922 //----------------------------next_exception-----------------------------------
923 SafePointNode* SafePointNode::next_exception() const {
924 if (len() == req()) {
925 return NULL;
926 } else {
927 Node* n = in(req());
928 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
929 return (SafePointNode*) n;
930 }
931 }
934 //------------------------------Ideal------------------------------------------
935 // Skip over any collapsed Regions
936 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
937 return remove_dead_region(phase, can_reshape) ? this : NULL;
938 }
940 //------------------------------Identity---------------------------------------
941 // Remove obviously duplicate safepoints
942 Node *SafePointNode::Identity( PhaseTransform *phase ) {
944 // If you have back to back safepoints, remove one
945 if( in(TypeFunc::Control)->is_SafePoint() )
946 return in(TypeFunc::Control);
948 if( in(0)->is_Proj() ) {
949 Node *n0 = in(0)->in(0);
950 // Check if he is a call projection (except Leaf Call)
951 if( n0->is_Catch() ) {
952 n0 = n0->in(0)->in(0);
953 assert( n0->is_Call(), "expect a call here" );
954 }
955 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
956 // Useless Safepoint, so remove it
957 return in(TypeFunc::Control);
958 }
959 }
961 return this;
962 }
964 //------------------------------Value------------------------------------------
965 const Type *SafePointNode::Value( PhaseTransform *phase ) const {
966 if( phase->type(in(0)) == Type::TOP ) return Type::TOP;
967 if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop
968 return Type::CONTROL;
969 }
971 #ifndef PRODUCT
972 void SafePointNode::dump_spec(outputStream *st) const {
973 st->print(" SafePoint ");
974 }
975 #endif
977 const RegMask &SafePointNode::in_RegMask(uint idx) const {
978 if( idx < TypeFunc::Parms ) return RegMask::Empty;
979 // Values outside the domain represent debug info
980 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
981 }
982 const RegMask &SafePointNode::out_RegMask() const {
983 return RegMask::Empty;
984 }
987 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
988 assert((int)grow_by > 0, "sanity");
989 int monoff = jvms->monoff();
990 int scloff = jvms->scloff();
991 int endoff = jvms->endoff();
992 assert(endoff == (int)req(), "no other states or debug info after me");
993 Node* top = Compile::current()->top();
994 for (uint i = 0; i < grow_by; i++) {
995 ins_req(monoff, top);
996 }
997 jvms->set_monoff(monoff + grow_by);
998 jvms->set_scloff(scloff + grow_by);
999 jvms->set_endoff(endoff + grow_by);
1000 }
1002 void SafePointNode::push_monitor(const FastLockNode *lock) {
1003 // Add a LockNode, which points to both the original BoxLockNode (the
1004 // stack space for the monitor) and the Object being locked.
1005 const int MonitorEdges = 2;
1006 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1007 assert(req() == jvms()->endoff(), "correct sizing");
1008 int nextmon = jvms()->scloff();
1009 if (GenerateSynchronizationCode) {
1010 add_req(lock->box_node());
1011 add_req(lock->obj_node());
1012 } else {
1013 Node* top = Compile::current()->top();
1014 add_req(top);
1015 add_req(top);
1016 }
1017 jvms()->set_scloff(nextmon+MonitorEdges);
1018 jvms()->set_endoff(req());
1019 }
1021 void SafePointNode::pop_monitor() {
1022 // Delete last monitor from debug info
1023 debug_only(int num_before_pop = jvms()->nof_monitors());
1024 const int MonitorEdges = (1<<JVMState::logMonitorEdges);
1025 int scloff = jvms()->scloff();
1026 int endoff = jvms()->endoff();
1027 int new_scloff = scloff - MonitorEdges;
1028 int new_endoff = endoff - MonitorEdges;
1029 jvms()->set_scloff(new_scloff);
1030 jvms()->set_endoff(new_endoff);
1031 while (scloff > new_scloff) del_req(--scloff);
1032 assert(jvms()->nof_monitors() == num_before_pop-1, "");
1033 }
1035 Node *SafePointNode::peek_monitor_box() const {
1036 int mon = jvms()->nof_monitors() - 1;
1037 assert(mon >= 0, "most have a monitor");
1038 return monitor_box(jvms(), mon);
1039 }
1041 Node *SafePointNode::peek_monitor_obj() const {
1042 int mon = jvms()->nof_monitors() - 1;
1043 assert(mon >= 0, "most have a monitor");
1044 return monitor_obj(jvms(), mon);
1045 }
1047 // Do we Match on this edge index or not? Match no edges
1048 uint SafePointNode::match_edge(uint idx) const {
1049 if( !needs_polling_address_input() )
1050 return 0;
1052 return (TypeFunc::Parms == idx);
1053 }
1055 //============== SafePointScalarObjectNode ==============
1057 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp,
1058 #ifdef ASSERT
1059 AllocateNode* alloc,
1060 #endif
1061 uint first_index,
1062 uint n_fields) :
1063 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1064 #ifdef ASSERT
1065 _alloc(alloc),
1066 #endif
1067 _first_index(first_index),
1068 _n_fields(n_fields)
1069 {
1070 init_class_id(Class_SafePointScalarObject);
1071 }
1073 // Do not allow value-numbering for SafePointScalarObject node.
1074 uint SafePointScalarObjectNode::hash() const { return NO_HASH; }
1075 uint SafePointScalarObjectNode::cmp( const Node &n ) const {
1076 return (&n == this); // Always fail except on self
1077 }
1079 uint SafePointScalarObjectNode::ideal_reg() const {
1080 return 0; // No matching to machine instruction
1081 }
1083 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1084 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1085 }
1087 const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1088 return RegMask::Empty;
1089 }
1091 uint SafePointScalarObjectNode::match_edge(uint idx) const {
1092 return 0;
1093 }
1095 SafePointScalarObjectNode*
1096 SafePointScalarObjectNode::clone(int jvms_adj, Dict* sosn_map) const {
1097 void* cached = (*sosn_map)[(void*)this];
1098 if (cached != NULL) {
1099 return (SafePointScalarObjectNode*)cached;
1100 }
1101 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1102 res->_first_index += jvms_adj;
1103 sosn_map->Insert((void*)this, (void*)res);
1104 return res;
1105 }
1108 #ifndef PRODUCT
1109 void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1110 st->print(" # fields@[%d..%d]", first_index(),
1111 first_index() + n_fields() - 1);
1112 }
1114 #endif
1116 //=============================================================================
1117 uint AllocateNode::size_of() const { return sizeof(*this); }
1119 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1120 Node *ctrl, Node *mem, Node *abio,
1121 Node *size, Node *klass_node, Node *initial_test)
1122 : CallNode(atype, NULL, TypeRawPtr::BOTTOM)
1123 {
1124 init_class_id(Class_Allocate);
1125 init_flags(Flag_is_macro);
1126 _is_scalar_replaceable = false;
1127 Node *topnode = C->top();
1129 init_req( TypeFunc::Control , ctrl );
1130 init_req( TypeFunc::I_O , abio );
1131 init_req( TypeFunc::Memory , mem );
1132 init_req( TypeFunc::ReturnAdr, topnode );
1133 init_req( TypeFunc::FramePtr , topnode );
1134 init_req( AllocSize , size);
1135 init_req( KlassNode , klass_node);
1136 init_req( InitialTest , initial_test);
1137 init_req( ALength , topnode);
1138 C->add_macro_node(this);
1139 }
1141 //=============================================================================
1142 uint AllocateArrayNode::size_of() const { return sizeof(*this); }
1144 Node* AllocateArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1145 if (remove_dead_region(phase, can_reshape)) return this;
1146 // Don't bother trying to transform a dead node
1147 if (in(0) && in(0)->is_top()) return NULL;
1149 const Type* type = phase->type(Ideal_length());
1150 if (type->isa_int() && type->is_int()->_hi < 0) {
1151 if (can_reshape) {
1152 PhaseIterGVN *igvn = phase->is_IterGVN();
1153 // Unreachable fall through path (negative array length),
1154 // the allocation can only throw so disconnect it.
1155 Node* proj = proj_out(TypeFunc::Control);
1156 Node* catchproj = NULL;
1157 if (proj != NULL) {
1158 for (DUIterator_Fast imax, i = proj->fast_outs(imax); i < imax; i++) {
1159 Node *cn = proj->fast_out(i);
1160 if (cn->is_Catch()) {
1161 catchproj = cn->as_Multi()->proj_out(CatchProjNode::fall_through_index);
1162 break;
1163 }
1164 }
1165 }
1166 if (catchproj != NULL && catchproj->outcnt() > 0 &&
1167 (catchproj->outcnt() > 1 ||
1168 catchproj->unique_out()->Opcode() != Op_Halt)) {
1169 assert(catchproj->is_CatchProj(), "must be a CatchProjNode");
1170 Node* nproj = catchproj->clone();
1171 igvn->register_new_node_with_optimizer(nproj);
1173 Node *frame = new (phase->C, 1) ParmNode( phase->C->start(), TypeFunc::FramePtr );
1174 frame = phase->transform(frame);
1175 // Halt & Catch Fire
1176 Node *halt = new (phase->C, TypeFunc::Parms) HaltNode( nproj, frame );
1177 phase->C->root()->add_req(halt);
1178 phase->transform(halt);
1180 igvn->replace_node(catchproj, phase->C->top());
1181 return this;
1182 }
1183 } else {
1184 // Can't correct it during regular GVN so register for IGVN
1185 phase->C->record_for_igvn(this);
1186 }
1187 }
1188 return NULL;
1189 }
1191 // Retrieve the length from the AllocateArrayNode. Narrow the type with a
1192 // CastII, if appropriate. If we are not allowed to create new nodes, and
1193 // a CastII is appropriate, return NULL.
1194 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) {
1195 Node *length = in(AllocateNode::ALength);
1196 assert(length != NULL, "length is not null");
1198 const TypeInt* length_type = phase->find_int_type(length);
1199 const TypeAryPtr* ary_type = oop_type->isa_aryptr();
1201 if (ary_type != NULL && length_type != NULL) {
1202 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type);
1203 if (narrow_length_type != length_type) {
1204 // Assert one of:
1205 // - the narrow_length is 0
1206 // - the narrow_length is not wider than length
1207 assert(narrow_length_type == TypeInt::ZERO ||
1208 (narrow_length_type->_hi <= length_type->_hi &&
1209 narrow_length_type->_lo >= length_type->_lo),
1210 "narrow type must be narrower than length type");
1212 // Return NULL if new nodes are not allowed
1213 if (!allow_new_nodes) return NULL;
1214 // Create a cast which is control dependent on the initialization to
1215 // propagate the fact that the array length must be positive.
1216 length = new (phase->C, 2) CastIINode(length, narrow_length_type);
1217 length->set_req(0, initialization()->proj_out(0));
1218 }
1219 }
1221 return length;
1222 }
1224 //=============================================================================
1225 uint LockNode::size_of() const { return sizeof(*this); }
1227 // Redundant lock elimination
1228 //
1229 // There are various patterns of locking where we release and
1230 // immediately reacquire a lock in a piece of code where no operations
1231 // occur in between that would be observable. In those cases we can
1232 // skip releasing and reacquiring the lock without violating any
1233 // fairness requirements. Doing this around a loop could cause a lock
1234 // to be held for a very long time so we concentrate on non-looping
1235 // control flow. We also require that the operations are fully
1236 // redundant meaning that we don't introduce new lock operations on
1237 // some paths so to be able to eliminate it on others ala PRE. This
1238 // would probably require some more extensive graph manipulation to
1239 // guarantee that the memory edges were all handled correctly.
1240 //
1241 // Assuming p is a simple predicate which can't trap in any way and s
1242 // is a synchronized method consider this code:
1243 //
1244 // s();
1245 // if (p)
1246 // s();
1247 // else
1248 // s();
1249 // s();
1250 //
1251 // 1. The unlocks of the first call to s can be eliminated if the
1252 // locks inside the then and else branches are eliminated.
1253 //
1254 // 2. The unlocks of the then and else branches can be eliminated if
1255 // the lock of the final call to s is eliminated.
1256 //
1257 // Either of these cases subsumes the simple case of sequential control flow
1258 //
1259 // Addtionally we can eliminate versions without the else case:
1260 //
1261 // s();
1262 // if (p)
1263 // s();
1264 // s();
1265 //
1266 // 3. In this case we eliminate the unlock of the first s, the lock
1267 // and unlock in the then case and the lock in the final s.
1268 //
1269 // Note also that in all these cases the then/else pieces don't have
1270 // to be trivial as long as they begin and end with synchronization
1271 // operations.
1272 //
1273 // s();
1274 // if (p)
1275 // s();
1276 // f();
1277 // s();
1278 // s();
1279 //
1280 // The code will work properly for this case, leaving in the unlock
1281 // before the call to f and the relock after it.
1282 //
1283 // A potentially interesting case which isn't handled here is when the
1284 // locking is partially redundant.
1285 //
1286 // s();
1287 // if (p)
1288 // s();
1289 //
1290 // This could be eliminated putting unlocking on the else case and
1291 // eliminating the first unlock and the lock in the then side.
1292 // Alternatively the unlock could be moved out of the then side so it
1293 // was after the merge and the first unlock and second lock
1294 // eliminated. This might require less manipulation of the memory
1295 // state to get correct.
1296 //
1297 // Additionally we might allow work between a unlock and lock before
1298 // giving up eliminating the locks. The current code disallows any
1299 // conditional control flow between these operations. A formulation
1300 // similar to partial redundancy elimination computing the
1301 // availability of unlocking and the anticipatability of locking at a
1302 // program point would allow detection of fully redundant locking with
1303 // some amount of work in between. I'm not sure how often I really
1304 // think that would occur though. Most of the cases I've seen
1305 // indicate it's likely non-trivial work would occur in between.
1306 // There may be other more complicated constructs where we could
1307 // eliminate locking but I haven't seen any others appear as hot or
1308 // interesting.
1309 //
1310 // Locking and unlocking have a canonical form in ideal that looks
1311 // roughly like this:
1312 //
1313 // <obj>
1314 // | \\------+
1315 // | \ \
1316 // | BoxLock \
1317 // | | | \
1318 // | | \ \
1319 // | | FastLock
1320 // | | /
1321 // | | /
1322 // | | |
1323 //
1324 // Lock
1325 // |
1326 // Proj #0
1327 // |
1328 // MembarAcquire
1329 // |
1330 // Proj #0
1331 //
1332 // MembarRelease
1333 // |
1334 // Proj #0
1335 // |
1336 // Unlock
1337 // |
1338 // Proj #0
1339 //
1340 //
1341 // This code proceeds by processing Lock nodes during PhaseIterGVN
1342 // and searching back through its control for the proper code
1343 // patterns. Once it finds a set of lock and unlock operations to
1344 // eliminate they are marked as eliminatable which causes the
1345 // expansion of the Lock and Unlock macro nodes to make the operation a NOP
1346 //
1347 //=============================================================================
1349 //
1350 // Utility function to skip over uninteresting control nodes. Nodes skipped are:
1351 // - copy regions. (These may not have been optimized away yet.)
1352 // - eliminated locking nodes
1353 //
1354 static Node *next_control(Node *ctrl) {
1355 if (ctrl == NULL)
1356 return NULL;
1357 while (1) {
1358 if (ctrl->is_Region()) {
1359 RegionNode *r = ctrl->as_Region();
1360 Node *n = r->is_copy();
1361 if (n == NULL)
1362 break; // hit a region, return it
1363 else
1364 ctrl = n;
1365 } else if (ctrl->is_Proj()) {
1366 Node *in0 = ctrl->in(0);
1367 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
1368 ctrl = in0->in(0);
1369 } else {
1370 break;
1371 }
1372 } else {
1373 break; // found an interesting control
1374 }
1375 }
1376 return ctrl;
1377 }
1378 //
1379 // Given a control, see if it's the control projection of an Unlock which
1380 // operating on the same object as lock.
1381 //
1382 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
1383 GrowableArray<AbstractLockNode*> &lock_ops) {
1384 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL;
1385 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) {
1386 Node *n = ctrl_proj->in(0);
1387 if (n != NULL && n->is_Unlock()) {
1388 UnlockNode *unlock = n->as_Unlock();
1389 if (lock->obj_node()->eqv_uncast(unlock->obj_node()) &&
1390 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) &&
1391 !unlock->is_eliminated()) {
1392 lock_ops.append(unlock);
1393 return true;
1394 }
1395 }
1396 }
1397 return false;
1398 }
1400 //
1401 // Find the lock matching an unlock. Returns null if a safepoint
1402 // or complicated control is encountered first.
1403 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
1404 LockNode *lock_result = NULL;
1405 // find the matching lock, or an intervening safepoint
1406 Node *ctrl = next_control(unlock->in(0));
1407 while (1) {
1408 assert(ctrl != NULL, "invalid control graph");
1409 assert(!ctrl->is_Start(), "missing lock for unlock");
1410 if (ctrl->is_top()) break; // dead control path
1411 if (ctrl->is_Proj()) ctrl = ctrl->in(0);
1412 if (ctrl->is_SafePoint()) {
1413 break; // found a safepoint (may be the lock we are searching for)
1414 } else if (ctrl->is_Region()) {
1415 // Check for a simple diamond pattern. Punt on anything more complicated
1416 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) {
1417 Node *in1 = next_control(ctrl->in(1));
1418 Node *in2 = next_control(ctrl->in(2));
1419 if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
1420 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
1421 ctrl = next_control(in1->in(0)->in(0));
1422 } else {
1423 break;
1424 }
1425 } else {
1426 break;
1427 }
1428 } else {
1429 ctrl = next_control(ctrl->in(0)); // keep searching
1430 }
1431 }
1432 if (ctrl->is_Lock()) {
1433 LockNode *lock = ctrl->as_Lock();
1434 if (lock->obj_node()->eqv_uncast(unlock->obj_node()) &&
1435 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) {
1436 lock_result = lock;
1437 }
1438 }
1439 return lock_result;
1440 }
1442 // This code corresponds to case 3 above.
1444 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
1445 GrowableArray<AbstractLockNode*> &lock_ops) {
1446 Node* if_node = node->in(0);
1447 bool if_true = node->is_IfTrue();
1449 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
1450 Node *lock_ctrl = next_control(if_node->in(0));
1451 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
1452 Node* lock1_node = NULL;
1453 ProjNode* proj = if_node->as_If()->proj_out(!if_true);
1454 if (if_true) {
1455 if (proj->is_IfFalse() && proj->outcnt() == 1) {
1456 lock1_node = proj->unique_out();
1457 }
1458 } else {
1459 if (proj->is_IfTrue() && proj->outcnt() == 1) {
1460 lock1_node = proj->unique_out();
1461 }
1462 }
1463 if (lock1_node != NULL && lock1_node->is_Lock()) {
1464 LockNode *lock1 = lock1_node->as_Lock();
1465 if (lock->obj_node()->eqv_uncast(lock1->obj_node()) &&
1466 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) &&
1467 !lock1->is_eliminated()) {
1468 lock_ops.append(lock1);
1469 return true;
1470 }
1471 }
1472 }
1473 }
1475 lock_ops.trunc_to(0);
1476 return false;
1477 }
1479 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
1480 GrowableArray<AbstractLockNode*> &lock_ops) {
1481 // check each control merging at this point for a matching unlock.
1482 // in(0) should be self edge so skip it.
1483 for (int i = 1; i < (int)region->req(); i++) {
1484 Node *in_node = next_control(region->in(i));
1485 if (in_node != NULL) {
1486 if (find_matching_unlock(in_node, lock, lock_ops)) {
1487 // found a match so keep on checking.
1488 continue;
1489 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
1490 continue;
1491 }
1493 // If we fall through to here then it was some kind of node we
1494 // don't understand or there wasn't a matching unlock, so give
1495 // up trying to merge locks.
1496 lock_ops.trunc_to(0);
1497 return false;
1498 }
1499 }
1500 return true;
1502 }
1504 #ifndef PRODUCT
1505 //
1506 // Create a counter which counts the number of times this lock is acquired
1507 //
1508 void AbstractLockNode::create_lock_counter(JVMState* state) {
1509 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
1510 }
1512 void AbstractLockNode::set_eliminated_lock_counter() {
1513 if (_counter) {
1514 // Update the counter to indicate that this lock was eliminated.
1515 // The counter update code will stay around even though the
1516 // optimizer will eliminate the lock operation itself.
1517 _counter->set_tag(NamedCounter::EliminatedLockCounter);
1518 }
1519 }
1520 #endif
1522 //=============================================================================
1523 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1525 // perform any generic optimizations first (returns 'this' or NULL)
1526 Node *result = SafePointNode::Ideal(phase, can_reshape);
1527 if (result != NULL) return result;
1528 // Don't bother trying to transform a dead node
1529 if (in(0) && in(0)->is_top()) return NULL;
1531 // Now see if we can optimize away this lock. We don't actually
1532 // remove the locking here, we simply set the _eliminate flag which
1533 // prevents macro expansion from expanding the lock. Since we don't
1534 // modify the graph, the value returned from this function is the
1535 // one computed above.
1536 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
1537 //
1538 // If we are locking an unescaped object, the lock/unlock is unnecessary
1539 //
1540 ConnectionGraph *cgr = phase->C->congraph();
1541 PointsToNode::EscapeState es = PointsToNode::GlobalEscape;
1542 if (cgr != NULL)
1543 es = cgr->escape_state(obj_node());
1544 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1545 assert(!is_eliminated() || is_coarsened(), "sanity");
1546 // The lock could be marked eliminated by lock coarsening
1547 // code during first IGVN before EA. Replace coarsened flag
1548 // to eliminate all associated locks/unlocks.
1549 this->set_non_esc_obj();
1550 return result;
1551 }
1553 //
1554 // Try lock coarsening
1555 //
1556 PhaseIterGVN* iter = phase->is_IterGVN();
1557 if (iter != NULL && !is_eliminated()) {
1559 GrowableArray<AbstractLockNode*> lock_ops;
1561 Node *ctrl = next_control(in(0));
1563 // now search back for a matching Unlock
1564 if (find_matching_unlock(ctrl, this, lock_ops)) {
1565 // found an unlock directly preceding this lock. This is the
1566 // case of single unlock directly control dependent on a
1567 // single lock which is the trivial version of case 1 or 2.
1568 } else if (ctrl->is_Region() ) {
1569 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
1570 // found lock preceded by multiple unlocks along all paths
1571 // joining at this point which is case 3 in description above.
1572 }
1573 } else {
1574 // see if this lock comes from either half of an if and the
1575 // predecessors merges unlocks and the other half of the if
1576 // performs a lock.
1577 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
1578 // found unlock splitting to an if with locks on both branches.
1579 }
1580 }
1582 if (lock_ops.length() > 0) {
1583 // add ourselves to the list of locks to be eliminated.
1584 lock_ops.append(this);
1586 #ifndef PRODUCT
1587 if (PrintEliminateLocks) {
1588 int locks = 0;
1589 int unlocks = 0;
1590 for (int i = 0; i < lock_ops.length(); i++) {
1591 AbstractLockNode* lock = lock_ops.at(i);
1592 if (lock->Opcode() == Op_Lock)
1593 locks++;
1594 else
1595 unlocks++;
1596 if (Verbose) {
1597 lock->dump(1);
1598 }
1599 }
1600 tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks);
1601 }
1602 #endif
1604 // for each of the identified locks, mark them
1605 // as eliminatable
1606 for (int i = 0; i < lock_ops.length(); i++) {
1607 AbstractLockNode* lock = lock_ops.at(i);
1609 // Mark it eliminated by coarsening and update any counters
1610 lock->set_coarsened();
1611 }
1612 } else if (ctrl->is_Region() &&
1613 iter->_worklist.member(ctrl)) {
1614 // We weren't able to find any opportunities but the region this
1615 // lock is control dependent on hasn't been processed yet so put
1616 // this lock back on the worklist so we can check again once any
1617 // region simplification has occurred.
1618 iter->_worklist.push(this);
1619 }
1620 }
1621 }
1623 return result;
1624 }
1626 //=============================================================================
1627 bool LockNode::is_nested_lock_region() {
1628 BoxLockNode* box = box_node()->as_BoxLock();
1629 int stk_slot = box->stack_slot();
1630 if (stk_slot <= 0)
1631 return false; // External lock or it is not Box (Phi node).
1633 // Ignore complex cases: merged locks or multiple locks.
1634 Node* obj = obj_node();
1635 LockNode* unique_lock = NULL;
1636 if (!box->is_simple_lock_region(&unique_lock, obj) ||
1637 (unique_lock != this)) {
1638 return false;
1639 }
1641 // Look for external lock for the same object.
1642 SafePointNode* sfn = this->as_SafePoint();
1643 JVMState* youngest_jvms = sfn->jvms();
1644 int max_depth = youngest_jvms->depth();
1645 for (int depth = 1; depth <= max_depth; depth++) {
1646 JVMState* jvms = youngest_jvms->of_depth(depth);
1647 int num_mon = jvms->nof_monitors();
1648 // Loop over monitors
1649 for (int idx = 0; idx < num_mon; idx++) {
1650 Node* obj_node = sfn->monitor_obj(jvms, idx);
1651 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock();
1652 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) {
1653 return true;
1654 }
1655 }
1656 }
1657 return false;
1658 }
1660 //=============================================================================
1661 uint UnlockNode::size_of() const { return sizeof(*this); }
1663 //=============================================================================
1664 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1666 // perform any generic optimizations first (returns 'this' or NULL)
1667 Node *result = SafePointNode::Ideal(phase, can_reshape);
1668 if (result != NULL) return result;
1669 // Don't bother trying to transform a dead node
1670 if (in(0) && in(0)->is_top()) return NULL;
1672 // Now see if we can optimize away this unlock. We don't actually
1673 // remove the unlocking here, we simply set the _eliminate flag which
1674 // prevents macro expansion from expanding the unlock. Since we don't
1675 // modify the graph, the value returned from this function is the
1676 // one computed above.
1677 // Escape state is defined after Parse phase.
1678 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
1679 //
1680 // If we are unlocking an unescaped object, the lock/unlock is unnecessary.
1681 //
1682 ConnectionGraph *cgr = phase->C->congraph();
1683 PointsToNode::EscapeState es = PointsToNode::GlobalEscape;
1684 if (cgr != NULL)
1685 es = cgr->escape_state(obj_node());
1686 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1687 assert(!is_eliminated() || is_coarsened(), "sanity");
1688 // The lock could be marked eliminated by lock coarsening
1689 // code during first IGVN before EA. Replace coarsened flag
1690 // to eliminate all associated locks/unlocks.
1691 this->set_non_esc_obj();
1692 }
1693 }
1694 return result;
1695 }