Sun, 23 Dec 2012 17:08:22 +0100
8005071: Incremental inlining for JSR 292
Summary: post parse inlining driven by number of live nodes.
Reviewed-by: twisti, kvn, jrose
1 /*
2 * Copyright (c) 1997, 2012, 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/callGenerator.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/escape.hpp"
31 #include "opto/locknode.hpp"
32 #include "opto/machnode.hpp"
33 #include "opto/matcher.hpp"
34 #include "opto/parse.hpp"
35 #include "opto/regalloc.hpp"
36 #include "opto/regmask.hpp"
37 #include "opto/rootnode.hpp"
38 #include "opto/runtime.hpp"
40 // Portions of code courtesy of Clifford Click
42 // Optimization - Graph Style
44 //=============================================================================
45 uint StartNode::size_of() const { return sizeof(*this); }
46 uint StartNode::cmp( const Node &n ) const
47 { return _domain == ((StartNode&)n)._domain; }
48 const Type *StartNode::bottom_type() const { return _domain; }
49 const Type *StartNode::Value(PhaseTransform *phase) const { return _domain; }
50 #ifndef PRODUCT
51 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);}
52 #endif
54 //------------------------------Ideal------------------------------------------
55 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){
56 return remove_dead_region(phase, can_reshape) ? this : NULL;
57 }
59 //------------------------------calling_convention-----------------------------
60 void StartNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
61 Matcher::calling_convention( sig_bt, parm_regs, argcnt, false );
62 }
64 //------------------------------Registers--------------------------------------
65 const RegMask &StartNode::in_RegMask(uint) const {
66 return RegMask::Empty;
67 }
69 //------------------------------match------------------------------------------
70 // Construct projections for incoming parameters, and their RegMask info
71 Node *StartNode::match( const ProjNode *proj, const Matcher *match ) {
72 switch (proj->_con) {
73 case TypeFunc::Control:
74 case TypeFunc::I_O:
75 case TypeFunc::Memory:
76 return new (match->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
77 case TypeFunc::FramePtr:
78 return new (match->C) MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP);
79 case TypeFunc::ReturnAdr:
80 return new (match->C) MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP);
81 case TypeFunc::Parms:
82 default: {
83 uint parm_num = proj->_con - TypeFunc::Parms;
84 const Type *t = _domain->field_at(proj->_con);
85 if (t->base() == Type::Half) // 2nd half of Longs and Doubles
86 return new (match->C) ConNode(Type::TOP);
87 uint ideal_reg = t->ideal_reg();
88 RegMask &rm = match->_calling_convention_mask[parm_num];
89 return new (match->C) MachProjNode(this,proj->_con,rm,ideal_reg);
90 }
91 }
92 return NULL;
93 }
95 //------------------------------StartOSRNode----------------------------------
96 // The method start node for an on stack replacement adapter
98 //------------------------------osr_domain-----------------------------
99 const TypeTuple *StartOSRNode::osr_domain() {
100 const Type **fields = TypeTuple::fields(2);
101 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer
103 return TypeTuple::make(TypeFunc::Parms+1, fields);
104 }
106 //=============================================================================
107 const char * const ParmNode::names[TypeFunc::Parms+1] = {
108 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms"
109 };
111 #ifndef PRODUCT
112 void ParmNode::dump_spec(outputStream *st) const {
113 if( _con < TypeFunc::Parms ) {
114 st->print(names[_con]);
115 } else {
116 st->print("Parm%d: ",_con-TypeFunc::Parms);
117 // Verbose and WizardMode dump bottom_type for all nodes
118 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st);
119 }
120 }
121 #endif
123 uint ParmNode::ideal_reg() const {
124 switch( _con ) {
125 case TypeFunc::Control : // fall through
126 case TypeFunc::I_O : // fall through
127 case TypeFunc::Memory : return 0;
128 case TypeFunc::FramePtr : // fall through
129 case TypeFunc::ReturnAdr: return Op_RegP;
130 default : assert( _con > TypeFunc::Parms, "" );
131 // fall through
132 case TypeFunc::Parms : {
133 // Type of argument being passed
134 const Type *t = in(0)->as_Start()->_domain->field_at(_con);
135 return t->ideal_reg();
136 }
137 }
138 ShouldNotReachHere();
139 return 0;
140 }
142 //=============================================================================
143 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) {
144 init_req(TypeFunc::Control,cntrl);
145 init_req(TypeFunc::I_O,i_o);
146 init_req(TypeFunc::Memory,memory);
147 init_req(TypeFunc::FramePtr,frameptr);
148 init_req(TypeFunc::ReturnAdr,retadr);
149 }
151 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){
152 return remove_dead_region(phase, can_reshape) ? this : NULL;
153 }
155 const Type *ReturnNode::Value( PhaseTransform *phase ) const {
156 return ( phase->type(in(TypeFunc::Control)) == Type::TOP)
157 ? Type::TOP
158 : Type::BOTTOM;
159 }
161 // Do we Match on this edge index or not? No edges on return nodes
162 uint ReturnNode::match_edge(uint idx) const {
163 return 0;
164 }
167 #ifndef PRODUCT
168 void ReturnNode::dump_req() const {
169 // Dump the required inputs, enclosed in '(' and ')'
170 uint i; // Exit value of loop
171 for( i=0; i<req(); i++ ) { // For all required inputs
172 if( i == TypeFunc::Parms ) tty->print("returns");
173 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
174 else tty->print("_ ");
175 }
176 }
177 #endif
179 //=============================================================================
180 RethrowNode::RethrowNode(
181 Node* cntrl,
182 Node* i_o,
183 Node* memory,
184 Node* frameptr,
185 Node* ret_adr,
186 Node* exception
187 ) : Node(TypeFunc::Parms + 1) {
188 init_req(TypeFunc::Control , cntrl );
189 init_req(TypeFunc::I_O , i_o );
190 init_req(TypeFunc::Memory , memory );
191 init_req(TypeFunc::FramePtr , frameptr );
192 init_req(TypeFunc::ReturnAdr, ret_adr);
193 init_req(TypeFunc::Parms , exception);
194 }
196 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){
197 return remove_dead_region(phase, can_reshape) ? this : NULL;
198 }
200 const Type *RethrowNode::Value( PhaseTransform *phase ) const {
201 return (phase->type(in(TypeFunc::Control)) == Type::TOP)
202 ? Type::TOP
203 : Type::BOTTOM;
204 }
206 uint RethrowNode::match_edge(uint idx) const {
207 return 0;
208 }
210 #ifndef PRODUCT
211 void RethrowNode::dump_req() const {
212 // Dump the required inputs, enclosed in '(' and ')'
213 uint i; // Exit value of loop
214 for( i=0; i<req(); i++ ) { // For all required inputs
215 if( i == TypeFunc::Parms ) tty->print("exception");
216 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
217 else tty->print("_ ");
218 }
219 }
220 #endif
222 //=============================================================================
223 // Do we Match on this edge index or not? Match only target address & method
224 uint TailCallNode::match_edge(uint idx) const {
225 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
226 }
228 //=============================================================================
229 // Do we Match on this edge index or not? Match only target address & oop
230 uint TailJumpNode::match_edge(uint idx) const {
231 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
232 }
234 //=============================================================================
235 JVMState::JVMState(ciMethod* method, JVMState* caller) :
236 _method(method) {
237 assert(method != NULL, "must be valid call site");
238 _reexecute = Reexecute_Undefined;
239 debug_only(_bci = -99); // random garbage value
240 debug_only(_map = (SafePointNode*)-1);
241 _caller = caller;
242 _depth = 1 + (caller == NULL ? 0 : caller->depth());
243 _locoff = TypeFunc::Parms;
244 _stkoff = _locoff + _method->max_locals();
245 _monoff = _stkoff + _method->max_stack();
246 _scloff = _monoff;
247 _endoff = _monoff;
248 _sp = 0;
249 }
250 JVMState::JVMState(int stack_size) :
251 _method(NULL) {
252 _bci = InvocationEntryBci;
253 _reexecute = Reexecute_Undefined;
254 debug_only(_map = (SafePointNode*)-1);
255 _caller = NULL;
256 _depth = 1;
257 _locoff = TypeFunc::Parms;
258 _stkoff = _locoff;
259 _monoff = _stkoff + stack_size;
260 _scloff = _monoff;
261 _endoff = _monoff;
262 _sp = 0;
263 }
265 //--------------------------------of_depth-------------------------------------
266 JVMState* JVMState::of_depth(int d) const {
267 const JVMState* jvmp = this;
268 assert(0 < d && (uint)d <= depth(), "oob");
269 for (int skip = depth() - d; skip > 0; skip--) {
270 jvmp = jvmp->caller();
271 }
272 assert(jvmp->depth() == (uint)d, "found the right one");
273 return (JVMState*)jvmp;
274 }
276 //-----------------------------same_calls_as-----------------------------------
277 bool JVMState::same_calls_as(const JVMState* that) const {
278 if (this == that) return true;
279 if (this->depth() != that->depth()) return false;
280 const JVMState* p = this;
281 const JVMState* q = that;
282 for (;;) {
283 if (p->_method != q->_method) return false;
284 if (p->_method == NULL) return true; // bci is irrelevant
285 if (p->_bci != q->_bci) return false;
286 if (p->_reexecute != q->_reexecute) return false;
287 p = p->caller();
288 q = q->caller();
289 if (p == q) return true;
290 assert(p != NULL && q != NULL, "depth check ensures we don't run off end");
291 }
292 }
294 //------------------------------debug_start------------------------------------
295 uint JVMState::debug_start() const {
296 debug_only(JVMState* jvmroot = of_depth(1));
297 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
298 return of_depth(1)->locoff();
299 }
301 //-------------------------------debug_end-------------------------------------
302 uint JVMState::debug_end() const {
303 debug_only(JVMState* jvmroot = of_depth(1));
304 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
305 return endoff();
306 }
308 //------------------------------debug_depth------------------------------------
309 uint JVMState::debug_depth() const {
310 uint total = 0;
311 for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) {
312 total += jvmp->debug_size();
313 }
314 return total;
315 }
317 #ifndef PRODUCT
319 //------------------------------format_helper----------------------------------
320 // Given an allocation (a Chaitin object) and a Node decide if the Node carries
321 // any defined value or not. If it does, print out the register or constant.
322 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
323 if (n == NULL) { st->print(" NULL"); return; }
324 if (n->is_SafePointScalarObject()) {
325 // Scalar replacement.
326 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
327 scobjs->append_if_missing(spobj);
328 int sco_n = scobjs->find(spobj);
329 assert(sco_n >= 0, "");
330 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
331 return;
332 }
333 if( OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
334 char buf[50];
335 regalloc->dump_register(n,buf);
336 st->print(" %s%d]=%s",msg,i,buf);
337 } else { // No register, but might be constant
338 const Type *t = n->bottom_type();
339 switch (t->base()) {
340 case Type::Int:
341 st->print(" %s%d]=#"INT32_FORMAT,msg,i,t->is_int()->get_con());
342 break;
343 case Type::AnyPtr:
344 assert( t == TypePtr::NULL_PTR, "" );
345 st->print(" %s%d]=#NULL",msg,i);
346 break;
347 case Type::AryPtr:
348 case Type::InstPtr:
349 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->isa_oopptr()->const_oop());
350 break;
351 case Type::KlassPtr:
352 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->make_ptr()->isa_klassptr()->klass());
353 break;
354 case Type::MetadataPtr:
355 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->make_ptr()->isa_metadataptr()->metadata());
356 break;
357 case Type::NarrowOop:
358 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->make_ptr()->isa_oopptr()->const_oop());
359 break;
360 case Type::RawPtr:
361 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,t->is_rawptr());
362 break;
363 case Type::DoubleCon:
364 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
365 break;
366 case Type::FloatCon:
367 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
368 break;
369 case Type::Long:
370 st->print(" %s%d]=#"INT64_FORMAT,msg,i,t->is_long()->get_con());
371 break;
372 case Type::Half:
373 case Type::Top:
374 st->print(" %s%d]=_",msg,i);
375 break;
376 default: ShouldNotReachHere();
377 }
378 }
379 }
381 //------------------------------format-----------------------------------------
382 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
383 st->print(" #");
384 if( _method ) {
385 _method->print_short_name(st);
386 st->print(" @ bci:%d ",_bci);
387 } else {
388 st->print_cr(" runtime stub ");
389 return;
390 }
391 if (n->is_MachSafePoint()) {
392 GrowableArray<SafePointScalarObjectNode*> scobjs;
393 MachSafePointNode *mcall = n->as_MachSafePoint();
394 uint i;
395 // Print locals
396 for( i = 0; i < (uint)loc_size(); i++ )
397 format_helper( regalloc, st, mcall->local(this, i), "L[", i, &scobjs );
398 // Print stack
399 for (i = 0; i < (uint)stk_size(); i++) {
400 if ((uint)(_stkoff + i) >= mcall->len())
401 st->print(" oob ");
402 else
403 format_helper( regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs );
404 }
405 for (i = 0; (int)i < nof_monitors(); i++) {
406 Node *box = mcall->monitor_box(this, i);
407 Node *obj = mcall->monitor_obj(this, i);
408 if ( OptoReg::is_valid(regalloc->get_reg_first(box)) ) {
409 box = BoxLockNode::box_node(box);
410 format_helper( regalloc, st, box, "MON-BOX[", i, &scobjs );
411 } else {
412 OptoReg::Name box_reg = BoxLockNode::reg(box);
413 st->print(" MON-BOX%d=%s+%d",
414 i,
415 OptoReg::regname(OptoReg::c_frame_pointer),
416 regalloc->reg2offset(box_reg));
417 }
418 const char* obj_msg = "MON-OBJ[";
419 if (EliminateLocks) {
420 if (BoxLockNode::box_node(box)->is_eliminated())
421 obj_msg = "MON-OBJ(LOCK ELIMINATED)[";
422 }
423 format_helper( regalloc, st, obj, obj_msg, i, &scobjs );
424 }
426 for (i = 0; i < (uint)scobjs.length(); i++) {
427 // Scalar replaced objects.
428 st->print_cr("");
429 st->print(" # ScObj" INT32_FORMAT " ", i);
430 SafePointScalarObjectNode* spobj = scobjs.at(i);
431 ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass();
432 assert(cik->is_instance_klass() ||
433 cik->is_array_klass(), "Not supported allocation.");
434 ciInstanceKlass *iklass = NULL;
435 if (cik->is_instance_klass()) {
436 cik->print_name_on(st);
437 iklass = cik->as_instance_klass();
438 } else if (cik->is_type_array_klass()) {
439 cik->as_array_klass()->base_element_type()->print_name_on(st);
440 st->print("[%d]", spobj->n_fields());
441 } else if (cik->is_obj_array_klass()) {
442 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass();
443 if (cie->is_instance_klass()) {
444 cie->print_name_on(st);
445 } else if (cie->is_type_array_klass()) {
446 cie->as_array_klass()->base_element_type()->print_name_on(st);
447 } else {
448 ShouldNotReachHere();
449 }
450 st->print("[%d]", spobj->n_fields());
451 int ndim = cik->as_array_klass()->dimension() - 1;
452 while (ndim-- > 0) {
453 st->print("[]");
454 }
455 }
456 st->print("={");
457 uint nf = spobj->n_fields();
458 if (nf > 0) {
459 uint first_ind = spobj->first_index();
460 Node* fld_node = mcall->in(first_ind);
461 ciField* cifield;
462 if (iklass != NULL) {
463 st->print(" [");
464 cifield = iklass->nonstatic_field_at(0);
465 cifield->print_name_on(st);
466 format_helper( regalloc, st, fld_node, ":", 0, &scobjs );
467 } else {
468 format_helper( regalloc, st, fld_node, "[", 0, &scobjs );
469 }
470 for (uint j = 1; j < nf; j++) {
471 fld_node = mcall->in(first_ind+j);
472 if (iklass != NULL) {
473 st->print(", [");
474 cifield = iklass->nonstatic_field_at(j);
475 cifield->print_name_on(st);
476 format_helper( regalloc, st, fld_node, ":", j, &scobjs );
477 } else {
478 format_helper( regalloc, st, fld_node, ", [", j, &scobjs );
479 }
480 }
481 }
482 st->print(" }");
483 }
484 }
485 st->print_cr("");
486 if (caller() != NULL) caller()->format(regalloc, n, st);
487 }
490 void JVMState::dump_spec(outputStream *st) const {
491 if (_method != NULL) {
492 bool printed = false;
493 if (!Verbose) {
494 // The JVMS dumps make really, really long lines.
495 // Take out the most boring parts, which are the package prefixes.
496 char buf[500];
497 stringStream namest(buf, sizeof(buf));
498 _method->print_short_name(&namest);
499 if (namest.count() < sizeof(buf)) {
500 const char* name = namest.base();
501 if (name[0] == ' ') ++name;
502 const char* endcn = strchr(name, ':'); // end of class name
503 if (endcn == NULL) endcn = strchr(name, '(');
504 if (endcn == NULL) endcn = name + strlen(name);
505 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
506 --endcn;
507 st->print(" %s", endcn);
508 printed = true;
509 }
510 }
511 if (!printed)
512 _method->print_short_name(st);
513 st->print(" @ bci:%d",_bci);
514 if(_reexecute == Reexecute_True)
515 st->print(" reexecute");
516 } else {
517 st->print(" runtime stub");
518 }
519 if (caller() != NULL) caller()->dump_spec(st);
520 }
523 void JVMState::dump_on(outputStream* st) const {
524 if (_map && !((uintptr_t)_map & 1)) {
525 if (_map->len() > _map->req()) { // _map->has_exceptions()
526 Node* ex = _map->in(_map->req()); // _map->next_exception()
527 // skip the first one; it's already being printed
528 while (ex != NULL && ex->len() > ex->req()) {
529 ex = ex->in(ex->req()); // ex->next_exception()
530 ex->dump(1);
531 }
532 }
533 _map->dump(2);
534 }
535 st->print("JVMS depth=%d loc=%d stk=%d arg=%d mon=%d scalar=%d end=%d mondepth=%d sp=%d bci=%d reexecute=%s method=",
536 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false");
537 if (_method == NULL) {
538 st->print_cr("(none)");
539 } else {
540 _method->print_name(st);
541 st->cr();
542 if (bci() >= 0 && bci() < _method->code_size()) {
543 st->print(" bc: ");
544 _method->print_codes_on(bci(), bci()+1, st);
545 }
546 }
547 if (caller() != NULL) {
548 caller()->dump_on(st);
549 }
550 }
552 // Extra way to dump a jvms from the debugger,
553 // to avoid a bug with C++ member function calls.
554 void dump_jvms(JVMState* jvms) {
555 jvms->dump();
556 }
557 #endif
559 //--------------------------clone_shallow--------------------------------------
560 JVMState* JVMState::clone_shallow(Compile* C) const {
561 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
562 n->set_bci(_bci);
563 n->_reexecute = _reexecute;
564 n->set_locoff(_locoff);
565 n->set_stkoff(_stkoff);
566 n->set_monoff(_monoff);
567 n->set_scloff(_scloff);
568 n->set_endoff(_endoff);
569 n->set_sp(_sp);
570 n->set_map(_map);
571 return n;
572 }
574 //---------------------------clone_deep----------------------------------------
575 JVMState* JVMState::clone_deep(Compile* C) const {
576 JVMState* n = clone_shallow(C);
577 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) {
578 p->_caller = p->_caller->clone_shallow(C);
579 }
580 assert(n->depth() == depth(), "sanity");
581 assert(n->debug_depth() == debug_depth(), "sanity");
582 return n;
583 }
585 //=============================================================================
586 uint CallNode::cmp( const Node &n ) const
587 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
588 #ifndef PRODUCT
589 void CallNode::dump_req() const {
590 // Dump the required inputs, enclosed in '(' and ')'
591 uint i; // Exit value of loop
592 for( i=0; i<req(); i++ ) { // For all required inputs
593 if( i == TypeFunc::Parms ) tty->print("(");
594 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
595 else tty->print("_ ");
596 }
597 tty->print(")");
598 }
600 void CallNode::dump_spec(outputStream *st) const {
601 st->print(" ");
602 tf()->dump_on(st);
603 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt);
604 if (jvms() != NULL) jvms()->dump_spec(st);
605 }
606 #endif
608 const Type *CallNode::bottom_type() const { return tf()->range(); }
609 const Type *CallNode::Value(PhaseTransform *phase) const {
610 if (phase->type(in(0)) == Type::TOP) return Type::TOP;
611 return tf()->range();
612 }
614 //------------------------------calling_convention-----------------------------
615 void CallNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
616 // Use the standard compiler calling convention
617 Matcher::calling_convention( sig_bt, parm_regs, argcnt, true );
618 }
621 //------------------------------match------------------------------------------
622 // Construct projections for control, I/O, memory-fields, ..., and
623 // return result(s) along with their RegMask info
624 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
625 switch (proj->_con) {
626 case TypeFunc::Control:
627 case TypeFunc::I_O:
628 case TypeFunc::Memory:
629 return new (match->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
631 case TypeFunc::Parms+1: // For LONG & DOUBLE returns
632 assert(tf()->_range->field_at(TypeFunc::Parms+1) == Type::HALF, "");
633 // 2nd half of doubles and longs
634 return new (match->C) MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad);
636 case TypeFunc::Parms: { // Normal returns
637 uint ideal_reg = tf()->range()->field_at(TypeFunc::Parms)->ideal_reg();
638 OptoRegPair regs = is_CallRuntime()
639 ? match->c_return_value(ideal_reg,true) // Calls into C runtime
640 : match-> return_value(ideal_reg,true); // Calls into compiled Java code
641 RegMask rm = RegMask(regs.first());
642 if( OptoReg::is_valid(regs.second()) )
643 rm.Insert( regs.second() );
644 return new (match->C) MachProjNode(this,proj->_con,rm,ideal_reg);
645 }
647 case TypeFunc::ReturnAdr:
648 case TypeFunc::FramePtr:
649 default:
650 ShouldNotReachHere();
651 }
652 return NULL;
653 }
655 // Do we Match on this edge index or not? Match no edges
656 uint CallNode::match_edge(uint idx) const {
657 return 0;
658 }
660 //
661 // Determine whether the call could modify the field of the specified
662 // instance at the specified offset.
663 //
664 bool CallNode::may_modify(const TypePtr *addr_t, PhaseTransform *phase) {
665 const TypeOopPtr *adrInst_t = addr_t->isa_oopptr();
667 // If not an OopPtr or not an instance type, assume the worst.
668 // Note: currently this method is called only for instance types.
669 if (adrInst_t == NULL || !adrInst_t->is_known_instance()) {
670 return true;
671 }
672 // The instance_id is set only for scalar-replaceable allocations which
673 // are not passed as arguments according to Escape Analysis.
674 return false;
675 }
677 // Does this call have a direct reference to n other than debug information?
678 bool CallNode::has_non_debug_use(Node *n) {
679 const TypeTuple * d = tf()->domain();
680 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
681 Node *arg = in(i);
682 if (arg == n) {
683 return true;
684 }
685 }
686 return false;
687 }
689 // Returns the unique CheckCastPP of a call
690 // or 'this' if there are several CheckCastPP
691 // or returns NULL if there is no one.
692 Node *CallNode::result_cast() {
693 Node *cast = NULL;
695 Node *p = proj_out(TypeFunc::Parms);
696 if (p == NULL)
697 return NULL;
699 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
700 Node *use = p->fast_out(i);
701 if (use->is_CheckCastPP()) {
702 if (cast != NULL) {
703 return this; // more than 1 CheckCastPP
704 }
705 cast = use;
706 }
707 }
708 return cast;
709 }
712 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj) {
713 projs->fallthrough_proj = NULL;
714 projs->fallthrough_catchproj = NULL;
715 projs->fallthrough_ioproj = NULL;
716 projs->catchall_ioproj = NULL;
717 projs->catchall_catchproj = NULL;
718 projs->fallthrough_memproj = NULL;
719 projs->catchall_memproj = NULL;
720 projs->resproj = NULL;
721 projs->exobj = NULL;
723 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
724 ProjNode *pn = fast_out(i)->as_Proj();
725 if (pn->outcnt() == 0) continue;
726 switch (pn->_con) {
727 case TypeFunc::Control:
728 {
729 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
730 projs->fallthrough_proj = pn;
731 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
732 const Node *cn = pn->fast_out(j);
733 if (cn->is_Catch()) {
734 ProjNode *cpn = NULL;
735 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
736 cpn = cn->fast_out(k)->as_Proj();
737 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
738 if (cpn->_con == CatchProjNode::fall_through_index)
739 projs->fallthrough_catchproj = cpn;
740 else {
741 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
742 projs->catchall_catchproj = cpn;
743 }
744 }
745 }
746 break;
747 }
748 case TypeFunc::I_O:
749 if (pn->_is_io_use)
750 projs->catchall_ioproj = pn;
751 else
752 projs->fallthrough_ioproj = pn;
753 for (DUIterator j = pn->outs(); pn->has_out(j); j++) {
754 Node* e = pn->out(j);
755 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) {
756 assert(projs->exobj == NULL, "only one");
757 projs->exobj = e;
758 }
759 }
760 break;
761 case TypeFunc::Memory:
762 if (pn->_is_io_use)
763 projs->catchall_memproj = pn;
764 else
765 projs->fallthrough_memproj = pn;
766 break;
767 case TypeFunc::Parms:
768 projs->resproj = pn;
769 break;
770 default:
771 assert(false, "unexpected projection from allocation node.");
772 }
773 }
775 // The resproj may not exist because the result couuld be ignored
776 // and the exception object may not exist if an exception handler
777 // swallows the exception but all the other must exist and be found.
778 assert(projs->fallthrough_proj != NULL, "must be found");
779 assert(Compile::current()->inlining_incrementally() || projs->fallthrough_catchproj != NULL, "must be found");
780 assert(Compile::current()->inlining_incrementally() || projs->fallthrough_memproj != NULL, "must be found");
781 assert(Compile::current()->inlining_incrementally() || projs->fallthrough_ioproj != NULL, "must be found");
782 assert(Compile::current()->inlining_incrementally() || projs->catchall_catchproj != NULL, "must be found");
783 if (separate_io_proj) {
784 assert(Compile::current()->inlining_incrementally() || projs->catchall_memproj != NULL, "must be found");
785 assert(Compile::current()->inlining_incrementally() || projs->catchall_ioproj != NULL, "must be found");
786 }
787 }
789 Node *CallNode::Ideal(PhaseGVN *phase, bool can_reshape) {
790 CallGenerator* cg = generator();
791 if (can_reshape && cg != NULL && cg->is_mh_late_inline() && !cg->already_attempted()) {
792 // Check whether this MH handle call becomes a candidate for inlining
793 ciMethod* callee = cg->method();
794 vmIntrinsics::ID iid = callee->intrinsic_id();
795 if (iid == vmIntrinsics::_invokeBasic) {
796 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) {
797 phase->C->prepend_late_inline(cg);
798 set_generator(NULL);
799 }
800 } else {
801 assert(callee->has_member_arg(), "wrong type of call?");
802 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) {
803 phase->C->prepend_late_inline(cg);
804 set_generator(NULL);
805 }
806 }
807 }
808 return SafePointNode::Ideal(phase, can_reshape);
809 }
812 //=============================================================================
813 uint CallJavaNode::size_of() const { return sizeof(*this); }
814 uint CallJavaNode::cmp( const Node &n ) const {
815 CallJavaNode &call = (CallJavaNode&)n;
816 return CallNode::cmp(call) && _method == call._method;
817 }
818 #ifndef PRODUCT
819 void CallJavaNode::dump_spec(outputStream *st) const {
820 if( _method ) _method->print_short_name(st);
821 CallNode::dump_spec(st);
822 }
823 #endif
825 //=============================================================================
826 uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
827 uint CallStaticJavaNode::cmp( const Node &n ) const {
828 CallStaticJavaNode &call = (CallStaticJavaNode&)n;
829 return CallJavaNode::cmp(call);
830 }
832 //----------------------------uncommon_trap_request----------------------------
833 // If this is an uncommon trap, return the request code, else zero.
834 int CallStaticJavaNode::uncommon_trap_request() const {
835 if (_name != NULL && !strcmp(_name, "uncommon_trap")) {
836 return extract_uncommon_trap_request(this);
837 }
838 return 0;
839 }
840 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
841 #ifndef PRODUCT
842 if (!(call->req() > TypeFunc::Parms &&
843 call->in(TypeFunc::Parms) != NULL &&
844 call->in(TypeFunc::Parms)->is_Con())) {
845 assert(_in_dump_cnt != 0, "OK if dumping");
846 tty->print("[bad uncommon trap]");
847 return 0;
848 }
849 #endif
850 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
851 }
853 #ifndef PRODUCT
854 void CallStaticJavaNode::dump_spec(outputStream *st) const {
855 st->print("# Static ");
856 if (_name != NULL) {
857 st->print("%s", _name);
858 int trap_req = uncommon_trap_request();
859 if (trap_req != 0) {
860 char buf[100];
861 st->print("(%s)",
862 Deoptimization::format_trap_request(buf, sizeof(buf),
863 trap_req));
864 }
865 st->print(" ");
866 }
867 CallJavaNode::dump_spec(st);
868 }
869 #endif
871 //=============================================================================
872 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
873 uint CallDynamicJavaNode::cmp( const Node &n ) const {
874 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
875 return CallJavaNode::cmp(call);
876 }
877 #ifndef PRODUCT
878 void CallDynamicJavaNode::dump_spec(outputStream *st) const {
879 st->print("# Dynamic ");
880 CallJavaNode::dump_spec(st);
881 }
882 #endif
884 //=============================================================================
885 uint CallRuntimeNode::size_of() const { return sizeof(*this); }
886 uint CallRuntimeNode::cmp( const Node &n ) const {
887 CallRuntimeNode &call = (CallRuntimeNode&)n;
888 return CallNode::cmp(call) && !strcmp(_name,call._name);
889 }
890 #ifndef PRODUCT
891 void CallRuntimeNode::dump_spec(outputStream *st) const {
892 st->print("# ");
893 st->print(_name);
894 CallNode::dump_spec(st);
895 }
896 #endif
898 //------------------------------calling_convention-----------------------------
899 void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
900 Matcher::c_calling_convention( sig_bt, parm_regs, argcnt );
901 }
903 //=============================================================================
904 //------------------------------calling_convention-----------------------------
907 //=============================================================================
908 #ifndef PRODUCT
909 void CallLeafNode::dump_spec(outputStream *st) const {
910 st->print("# ");
911 st->print(_name);
912 CallNode::dump_spec(st);
913 }
914 #endif
916 //=============================================================================
918 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) {
919 assert(verify_jvms(jvms), "jvms must match");
920 int loc = jvms->locoff() + idx;
921 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
922 // If current local idx is top then local idx - 1 could
923 // be a long/double that needs to be killed since top could
924 // represent the 2nd half ofthe long/double.
925 uint ideal = in(loc -1)->ideal_reg();
926 if (ideal == Op_RegD || ideal == Op_RegL) {
927 // set other (low index) half to top
928 set_req(loc - 1, in(loc));
929 }
930 }
931 set_req(loc, c);
932 }
934 uint SafePointNode::size_of() const { return sizeof(*this); }
935 uint SafePointNode::cmp( const Node &n ) const {
936 return (&n == this); // Always fail except on self
937 }
939 //-------------------------set_next_exception----------------------------------
940 void SafePointNode::set_next_exception(SafePointNode* n) {
941 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception");
942 if (len() == req()) {
943 if (n != NULL) add_prec(n);
944 } else {
945 set_prec(req(), n);
946 }
947 }
950 //----------------------------next_exception-----------------------------------
951 SafePointNode* SafePointNode::next_exception() const {
952 if (len() == req()) {
953 return NULL;
954 } else {
955 Node* n = in(req());
956 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
957 return (SafePointNode*) n;
958 }
959 }
962 //------------------------------Ideal------------------------------------------
963 // Skip over any collapsed Regions
964 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
965 return remove_dead_region(phase, can_reshape) ? this : NULL;
966 }
968 //------------------------------Identity---------------------------------------
969 // Remove obviously duplicate safepoints
970 Node *SafePointNode::Identity( PhaseTransform *phase ) {
972 // If you have back to back safepoints, remove one
973 if( in(TypeFunc::Control)->is_SafePoint() )
974 return in(TypeFunc::Control);
976 if( in(0)->is_Proj() ) {
977 Node *n0 = in(0)->in(0);
978 // Check if he is a call projection (except Leaf Call)
979 if( n0->is_Catch() ) {
980 n0 = n0->in(0)->in(0);
981 assert( n0->is_Call(), "expect a call here" );
982 }
983 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
984 // Useless Safepoint, so remove it
985 return in(TypeFunc::Control);
986 }
987 }
989 return this;
990 }
992 //------------------------------Value------------------------------------------
993 const Type *SafePointNode::Value( PhaseTransform *phase ) const {
994 if( phase->type(in(0)) == Type::TOP ) return Type::TOP;
995 if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop
996 return Type::CONTROL;
997 }
999 #ifndef PRODUCT
1000 void SafePointNode::dump_spec(outputStream *st) const {
1001 st->print(" SafePoint ");
1002 }
1003 #endif
1005 const RegMask &SafePointNode::in_RegMask(uint idx) const {
1006 if( idx < TypeFunc::Parms ) return RegMask::Empty;
1007 // Values outside the domain represent debug info
1008 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1009 }
1010 const RegMask &SafePointNode::out_RegMask() const {
1011 return RegMask::Empty;
1012 }
1015 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
1016 assert((int)grow_by > 0, "sanity");
1017 int monoff = jvms->monoff();
1018 int scloff = jvms->scloff();
1019 int endoff = jvms->endoff();
1020 assert(endoff == (int)req(), "no other states or debug info after me");
1021 Node* top = Compile::current()->top();
1022 for (uint i = 0; i < grow_by; i++) {
1023 ins_req(monoff, top);
1024 }
1025 jvms->set_monoff(monoff + grow_by);
1026 jvms->set_scloff(scloff + grow_by);
1027 jvms->set_endoff(endoff + grow_by);
1028 }
1030 void SafePointNode::push_monitor(const FastLockNode *lock) {
1031 // Add a LockNode, which points to both the original BoxLockNode (the
1032 // stack space for the monitor) and the Object being locked.
1033 const int MonitorEdges = 2;
1034 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1035 assert(req() == jvms()->endoff(), "correct sizing");
1036 int nextmon = jvms()->scloff();
1037 if (GenerateSynchronizationCode) {
1038 add_req(lock->box_node());
1039 add_req(lock->obj_node());
1040 } else {
1041 Node* top = Compile::current()->top();
1042 add_req(top);
1043 add_req(top);
1044 }
1045 jvms()->set_scloff(nextmon+MonitorEdges);
1046 jvms()->set_endoff(req());
1047 }
1049 void SafePointNode::pop_monitor() {
1050 // Delete last monitor from debug info
1051 debug_only(int num_before_pop = jvms()->nof_monitors());
1052 const int MonitorEdges = (1<<JVMState::logMonitorEdges);
1053 int scloff = jvms()->scloff();
1054 int endoff = jvms()->endoff();
1055 int new_scloff = scloff - MonitorEdges;
1056 int new_endoff = endoff - MonitorEdges;
1057 jvms()->set_scloff(new_scloff);
1058 jvms()->set_endoff(new_endoff);
1059 while (scloff > new_scloff) del_req(--scloff);
1060 assert(jvms()->nof_monitors() == num_before_pop-1, "");
1061 }
1063 Node *SafePointNode::peek_monitor_box() const {
1064 int mon = jvms()->nof_monitors() - 1;
1065 assert(mon >= 0, "most have a monitor");
1066 return monitor_box(jvms(), mon);
1067 }
1069 Node *SafePointNode::peek_monitor_obj() const {
1070 int mon = jvms()->nof_monitors() - 1;
1071 assert(mon >= 0, "most have a monitor");
1072 return monitor_obj(jvms(), mon);
1073 }
1075 // Do we Match on this edge index or not? Match no edges
1076 uint SafePointNode::match_edge(uint idx) const {
1077 if( !needs_polling_address_input() )
1078 return 0;
1080 return (TypeFunc::Parms == idx);
1081 }
1083 //============== SafePointScalarObjectNode ==============
1085 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp,
1086 #ifdef ASSERT
1087 AllocateNode* alloc,
1088 #endif
1089 uint first_index,
1090 uint n_fields) :
1091 TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1092 #ifdef ASSERT
1093 _alloc(alloc),
1094 #endif
1095 _first_index(first_index),
1096 _n_fields(n_fields)
1097 {
1098 init_class_id(Class_SafePointScalarObject);
1099 }
1101 // Do not allow value-numbering for SafePointScalarObject node.
1102 uint SafePointScalarObjectNode::hash() const { return NO_HASH; }
1103 uint SafePointScalarObjectNode::cmp( const Node &n ) const {
1104 return (&n == this); // Always fail except on self
1105 }
1107 uint SafePointScalarObjectNode::ideal_reg() const {
1108 return 0; // No matching to machine instruction
1109 }
1111 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1112 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1113 }
1115 const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1116 return RegMask::Empty;
1117 }
1119 uint SafePointScalarObjectNode::match_edge(uint idx) const {
1120 return 0;
1121 }
1123 SafePointScalarObjectNode*
1124 SafePointScalarObjectNode::clone(int jvms_adj, Dict* sosn_map) const {
1125 void* cached = (*sosn_map)[(void*)this];
1126 if (cached != NULL) {
1127 return (SafePointScalarObjectNode*)cached;
1128 }
1129 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1130 res->_first_index += jvms_adj;
1131 sosn_map->Insert((void*)this, (void*)res);
1132 return res;
1133 }
1136 #ifndef PRODUCT
1137 void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1138 st->print(" # fields@[%d..%d]", first_index(),
1139 first_index() + n_fields() - 1);
1140 }
1142 #endif
1144 //=============================================================================
1145 uint AllocateNode::size_of() const { return sizeof(*this); }
1147 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1148 Node *ctrl, Node *mem, Node *abio,
1149 Node *size, Node *klass_node, Node *initial_test)
1150 : CallNode(atype, NULL, TypeRawPtr::BOTTOM)
1151 {
1152 init_class_id(Class_Allocate);
1153 init_flags(Flag_is_macro);
1154 _is_scalar_replaceable = false;
1155 Node *topnode = C->top();
1157 init_req( TypeFunc::Control , ctrl );
1158 init_req( TypeFunc::I_O , abio );
1159 init_req( TypeFunc::Memory , mem );
1160 init_req( TypeFunc::ReturnAdr, topnode );
1161 init_req( TypeFunc::FramePtr , topnode );
1162 init_req( AllocSize , size);
1163 init_req( KlassNode , klass_node);
1164 init_req( InitialTest , initial_test);
1165 init_req( ALength , topnode);
1166 C->add_macro_node(this);
1167 }
1169 //=============================================================================
1170 uint AllocateArrayNode::size_of() const { return sizeof(*this); }
1172 Node* AllocateArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1173 if (remove_dead_region(phase, can_reshape)) return this;
1174 // Don't bother trying to transform a dead node
1175 if (in(0) && in(0)->is_top()) return NULL;
1177 const Type* type = phase->type(Ideal_length());
1178 if (type->isa_int() && type->is_int()->_hi < 0) {
1179 if (can_reshape) {
1180 PhaseIterGVN *igvn = phase->is_IterGVN();
1181 // Unreachable fall through path (negative array length),
1182 // the allocation can only throw so disconnect it.
1183 Node* proj = proj_out(TypeFunc::Control);
1184 Node* catchproj = NULL;
1185 if (proj != NULL) {
1186 for (DUIterator_Fast imax, i = proj->fast_outs(imax); i < imax; i++) {
1187 Node *cn = proj->fast_out(i);
1188 if (cn->is_Catch()) {
1189 catchproj = cn->as_Multi()->proj_out(CatchProjNode::fall_through_index);
1190 break;
1191 }
1192 }
1193 }
1194 if (catchproj != NULL && catchproj->outcnt() > 0 &&
1195 (catchproj->outcnt() > 1 ||
1196 catchproj->unique_out()->Opcode() != Op_Halt)) {
1197 assert(catchproj->is_CatchProj(), "must be a CatchProjNode");
1198 Node* nproj = catchproj->clone();
1199 igvn->register_new_node_with_optimizer(nproj);
1201 Node *frame = new (phase->C) ParmNode( phase->C->start(), TypeFunc::FramePtr );
1202 frame = phase->transform(frame);
1203 // Halt & Catch Fire
1204 Node *halt = new (phase->C) HaltNode( nproj, frame );
1205 phase->C->root()->add_req(halt);
1206 phase->transform(halt);
1208 igvn->replace_node(catchproj, phase->C->top());
1209 return this;
1210 }
1211 } else {
1212 // Can't correct it during regular GVN so register for IGVN
1213 phase->C->record_for_igvn(this);
1214 }
1215 }
1216 return NULL;
1217 }
1219 // Retrieve the length from the AllocateArrayNode. Narrow the type with a
1220 // CastII, if appropriate. If we are not allowed to create new nodes, and
1221 // a CastII is appropriate, return NULL.
1222 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) {
1223 Node *length = in(AllocateNode::ALength);
1224 assert(length != NULL, "length is not null");
1226 const TypeInt* length_type = phase->find_int_type(length);
1227 const TypeAryPtr* ary_type = oop_type->isa_aryptr();
1229 if (ary_type != NULL && length_type != NULL) {
1230 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type);
1231 if (narrow_length_type != length_type) {
1232 // Assert one of:
1233 // - the narrow_length is 0
1234 // - the narrow_length is not wider than length
1235 assert(narrow_length_type == TypeInt::ZERO ||
1236 (narrow_length_type->_hi <= length_type->_hi &&
1237 narrow_length_type->_lo >= length_type->_lo),
1238 "narrow type must be narrower than length type");
1240 // Return NULL if new nodes are not allowed
1241 if (!allow_new_nodes) return NULL;
1242 // Create a cast which is control dependent on the initialization to
1243 // propagate the fact that the array length must be positive.
1244 length = new (phase->C) CastIINode(length, narrow_length_type);
1245 length->set_req(0, initialization()->proj_out(0));
1246 }
1247 }
1249 return length;
1250 }
1252 //=============================================================================
1253 uint LockNode::size_of() const { return sizeof(*this); }
1255 // Redundant lock elimination
1256 //
1257 // There are various patterns of locking where we release and
1258 // immediately reacquire a lock in a piece of code where no operations
1259 // occur in between that would be observable. In those cases we can
1260 // skip releasing and reacquiring the lock without violating any
1261 // fairness requirements. Doing this around a loop could cause a lock
1262 // to be held for a very long time so we concentrate on non-looping
1263 // control flow. We also require that the operations are fully
1264 // redundant meaning that we don't introduce new lock operations on
1265 // some paths so to be able to eliminate it on others ala PRE. This
1266 // would probably require some more extensive graph manipulation to
1267 // guarantee that the memory edges were all handled correctly.
1268 //
1269 // Assuming p is a simple predicate which can't trap in any way and s
1270 // is a synchronized method consider this code:
1271 //
1272 // s();
1273 // if (p)
1274 // s();
1275 // else
1276 // s();
1277 // s();
1278 //
1279 // 1. The unlocks of the first call to s can be eliminated if the
1280 // locks inside the then and else branches are eliminated.
1281 //
1282 // 2. The unlocks of the then and else branches can be eliminated if
1283 // the lock of the final call to s is eliminated.
1284 //
1285 // Either of these cases subsumes the simple case of sequential control flow
1286 //
1287 // Addtionally we can eliminate versions without the else case:
1288 //
1289 // s();
1290 // if (p)
1291 // s();
1292 // s();
1293 //
1294 // 3. In this case we eliminate the unlock of the first s, the lock
1295 // and unlock in the then case and the lock in the final s.
1296 //
1297 // Note also that in all these cases the then/else pieces don't have
1298 // to be trivial as long as they begin and end with synchronization
1299 // operations.
1300 //
1301 // s();
1302 // if (p)
1303 // s();
1304 // f();
1305 // s();
1306 // s();
1307 //
1308 // The code will work properly for this case, leaving in the unlock
1309 // before the call to f and the relock after it.
1310 //
1311 // A potentially interesting case which isn't handled here is when the
1312 // locking is partially redundant.
1313 //
1314 // s();
1315 // if (p)
1316 // s();
1317 //
1318 // This could be eliminated putting unlocking on the else case and
1319 // eliminating the first unlock and the lock in the then side.
1320 // Alternatively the unlock could be moved out of the then side so it
1321 // was after the merge and the first unlock and second lock
1322 // eliminated. This might require less manipulation of the memory
1323 // state to get correct.
1324 //
1325 // Additionally we might allow work between a unlock and lock before
1326 // giving up eliminating the locks. The current code disallows any
1327 // conditional control flow between these operations. A formulation
1328 // similar to partial redundancy elimination computing the
1329 // availability of unlocking and the anticipatability of locking at a
1330 // program point would allow detection of fully redundant locking with
1331 // some amount of work in between. I'm not sure how often I really
1332 // think that would occur though. Most of the cases I've seen
1333 // indicate it's likely non-trivial work would occur in between.
1334 // There may be other more complicated constructs where we could
1335 // eliminate locking but I haven't seen any others appear as hot or
1336 // interesting.
1337 //
1338 // Locking and unlocking have a canonical form in ideal that looks
1339 // roughly like this:
1340 //
1341 // <obj>
1342 // | \\------+
1343 // | \ \
1344 // | BoxLock \
1345 // | | | \
1346 // | | \ \
1347 // | | FastLock
1348 // | | /
1349 // | | /
1350 // | | |
1351 //
1352 // Lock
1353 // |
1354 // Proj #0
1355 // |
1356 // MembarAcquire
1357 // |
1358 // Proj #0
1359 //
1360 // MembarRelease
1361 // |
1362 // Proj #0
1363 // |
1364 // Unlock
1365 // |
1366 // Proj #0
1367 //
1368 //
1369 // This code proceeds by processing Lock nodes during PhaseIterGVN
1370 // and searching back through its control for the proper code
1371 // patterns. Once it finds a set of lock and unlock operations to
1372 // eliminate they are marked as eliminatable which causes the
1373 // expansion of the Lock and Unlock macro nodes to make the operation a NOP
1374 //
1375 //=============================================================================
1377 //
1378 // Utility function to skip over uninteresting control nodes. Nodes skipped are:
1379 // - copy regions. (These may not have been optimized away yet.)
1380 // - eliminated locking nodes
1381 //
1382 static Node *next_control(Node *ctrl) {
1383 if (ctrl == NULL)
1384 return NULL;
1385 while (1) {
1386 if (ctrl->is_Region()) {
1387 RegionNode *r = ctrl->as_Region();
1388 Node *n = r->is_copy();
1389 if (n == NULL)
1390 break; // hit a region, return it
1391 else
1392 ctrl = n;
1393 } else if (ctrl->is_Proj()) {
1394 Node *in0 = ctrl->in(0);
1395 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
1396 ctrl = in0->in(0);
1397 } else {
1398 break;
1399 }
1400 } else {
1401 break; // found an interesting control
1402 }
1403 }
1404 return ctrl;
1405 }
1406 //
1407 // Given a control, see if it's the control projection of an Unlock which
1408 // operating on the same object as lock.
1409 //
1410 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
1411 GrowableArray<AbstractLockNode*> &lock_ops) {
1412 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL;
1413 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) {
1414 Node *n = ctrl_proj->in(0);
1415 if (n != NULL && n->is_Unlock()) {
1416 UnlockNode *unlock = n->as_Unlock();
1417 if (lock->obj_node()->eqv_uncast(unlock->obj_node()) &&
1418 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) &&
1419 !unlock->is_eliminated()) {
1420 lock_ops.append(unlock);
1421 return true;
1422 }
1423 }
1424 }
1425 return false;
1426 }
1428 //
1429 // Find the lock matching an unlock. Returns null if a safepoint
1430 // or complicated control is encountered first.
1431 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
1432 LockNode *lock_result = NULL;
1433 // find the matching lock, or an intervening safepoint
1434 Node *ctrl = next_control(unlock->in(0));
1435 while (1) {
1436 assert(ctrl != NULL, "invalid control graph");
1437 assert(!ctrl->is_Start(), "missing lock for unlock");
1438 if (ctrl->is_top()) break; // dead control path
1439 if (ctrl->is_Proj()) ctrl = ctrl->in(0);
1440 if (ctrl->is_SafePoint()) {
1441 break; // found a safepoint (may be the lock we are searching for)
1442 } else if (ctrl->is_Region()) {
1443 // Check for a simple diamond pattern. Punt on anything more complicated
1444 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) {
1445 Node *in1 = next_control(ctrl->in(1));
1446 Node *in2 = next_control(ctrl->in(2));
1447 if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
1448 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
1449 ctrl = next_control(in1->in(0)->in(0));
1450 } else {
1451 break;
1452 }
1453 } else {
1454 break;
1455 }
1456 } else {
1457 ctrl = next_control(ctrl->in(0)); // keep searching
1458 }
1459 }
1460 if (ctrl->is_Lock()) {
1461 LockNode *lock = ctrl->as_Lock();
1462 if (lock->obj_node()->eqv_uncast(unlock->obj_node()) &&
1463 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) {
1464 lock_result = lock;
1465 }
1466 }
1467 return lock_result;
1468 }
1470 // This code corresponds to case 3 above.
1472 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
1473 GrowableArray<AbstractLockNode*> &lock_ops) {
1474 Node* if_node = node->in(0);
1475 bool if_true = node->is_IfTrue();
1477 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
1478 Node *lock_ctrl = next_control(if_node->in(0));
1479 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
1480 Node* lock1_node = NULL;
1481 ProjNode* proj = if_node->as_If()->proj_out(!if_true);
1482 if (if_true) {
1483 if (proj->is_IfFalse() && proj->outcnt() == 1) {
1484 lock1_node = proj->unique_out();
1485 }
1486 } else {
1487 if (proj->is_IfTrue() && proj->outcnt() == 1) {
1488 lock1_node = proj->unique_out();
1489 }
1490 }
1491 if (lock1_node != NULL && lock1_node->is_Lock()) {
1492 LockNode *lock1 = lock1_node->as_Lock();
1493 if (lock->obj_node()->eqv_uncast(lock1->obj_node()) &&
1494 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) &&
1495 !lock1->is_eliminated()) {
1496 lock_ops.append(lock1);
1497 return true;
1498 }
1499 }
1500 }
1501 }
1503 lock_ops.trunc_to(0);
1504 return false;
1505 }
1507 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
1508 GrowableArray<AbstractLockNode*> &lock_ops) {
1509 // check each control merging at this point for a matching unlock.
1510 // in(0) should be self edge so skip it.
1511 for (int i = 1; i < (int)region->req(); i++) {
1512 Node *in_node = next_control(region->in(i));
1513 if (in_node != NULL) {
1514 if (find_matching_unlock(in_node, lock, lock_ops)) {
1515 // found a match so keep on checking.
1516 continue;
1517 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
1518 continue;
1519 }
1521 // If we fall through to here then it was some kind of node we
1522 // don't understand or there wasn't a matching unlock, so give
1523 // up trying to merge locks.
1524 lock_ops.trunc_to(0);
1525 return false;
1526 }
1527 }
1528 return true;
1530 }
1532 #ifndef PRODUCT
1533 //
1534 // Create a counter which counts the number of times this lock is acquired
1535 //
1536 void AbstractLockNode::create_lock_counter(JVMState* state) {
1537 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
1538 }
1540 void AbstractLockNode::set_eliminated_lock_counter() {
1541 if (_counter) {
1542 // Update the counter to indicate that this lock was eliminated.
1543 // The counter update code will stay around even though the
1544 // optimizer will eliminate the lock operation itself.
1545 _counter->set_tag(NamedCounter::EliminatedLockCounter);
1546 }
1547 }
1548 #endif
1550 //=============================================================================
1551 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1553 // perform any generic optimizations first (returns 'this' or NULL)
1554 Node *result = SafePointNode::Ideal(phase, can_reshape);
1555 if (result != NULL) return result;
1556 // Don't bother trying to transform a dead node
1557 if (in(0) && in(0)->is_top()) return NULL;
1559 // Now see if we can optimize away this lock. We don't actually
1560 // remove the locking here, we simply set the _eliminate flag which
1561 // prevents macro expansion from expanding the lock. Since we don't
1562 // modify the graph, the value returned from this function is the
1563 // one computed above.
1564 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
1565 //
1566 // If we are locking an unescaped object, the lock/unlock is unnecessary
1567 //
1568 ConnectionGraph *cgr = phase->C->congraph();
1569 if (cgr != NULL && cgr->not_global_escape(obj_node())) {
1570 assert(!is_eliminated() || is_coarsened(), "sanity");
1571 // The lock could be marked eliminated by lock coarsening
1572 // code during first IGVN before EA. Replace coarsened flag
1573 // to eliminate all associated locks/unlocks.
1574 this->set_non_esc_obj();
1575 return result;
1576 }
1578 //
1579 // Try lock coarsening
1580 //
1581 PhaseIterGVN* iter = phase->is_IterGVN();
1582 if (iter != NULL && !is_eliminated()) {
1584 GrowableArray<AbstractLockNode*> lock_ops;
1586 Node *ctrl = next_control(in(0));
1588 // now search back for a matching Unlock
1589 if (find_matching_unlock(ctrl, this, lock_ops)) {
1590 // found an unlock directly preceding this lock. This is the
1591 // case of single unlock directly control dependent on a
1592 // single lock which is the trivial version of case 1 or 2.
1593 } else if (ctrl->is_Region() ) {
1594 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
1595 // found lock preceded by multiple unlocks along all paths
1596 // joining at this point which is case 3 in description above.
1597 }
1598 } else {
1599 // see if this lock comes from either half of an if and the
1600 // predecessors merges unlocks and the other half of the if
1601 // performs a lock.
1602 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
1603 // found unlock splitting to an if with locks on both branches.
1604 }
1605 }
1607 if (lock_ops.length() > 0) {
1608 // add ourselves to the list of locks to be eliminated.
1609 lock_ops.append(this);
1611 #ifndef PRODUCT
1612 if (PrintEliminateLocks) {
1613 int locks = 0;
1614 int unlocks = 0;
1615 for (int i = 0; i < lock_ops.length(); i++) {
1616 AbstractLockNode* lock = lock_ops.at(i);
1617 if (lock->Opcode() == Op_Lock)
1618 locks++;
1619 else
1620 unlocks++;
1621 if (Verbose) {
1622 lock->dump(1);
1623 }
1624 }
1625 tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks);
1626 }
1627 #endif
1629 // for each of the identified locks, mark them
1630 // as eliminatable
1631 for (int i = 0; i < lock_ops.length(); i++) {
1632 AbstractLockNode* lock = lock_ops.at(i);
1634 // Mark it eliminated by coarsening and update any counters
1635 lock->set_coarsened();
1636 }
1637 } else if (ctrl->is_Region() &&
1638 iter->_worklist.member(ctrl)) {
1639 // We weren't able to find any opportunities but the region this
1640 // lock is control dependent on hasn't been processed yet so put
1641 // this lock back on the worklist so we can check again once any
1642 // region simplification has occurred.
1643 iter->_worklist.push(this);
1644 }
1645 }
1646 }
1648 return result;
1649 }
1651 //=============================================================================
1652 bool LockNode::is_nested_lock_region() {
1653 BoxLockNode* box = box_node()->as_BoxLock();
1654 int stk_slot = box->stack_slot();
1655 if (stk_slot <= 0)
1656 return false; // External lock or it is not Box (Phi node).
1658 // Ignore complex cases: merged locks or multiple locks.
1659 Node* obj = obj_node();
1660 LockNode* unique_lock = NULL;
1661 if (!box->is_simple_lock_region(&unique_lock, obj) ||
1662 (unique_lock != this)) {
1663 return false;
1664 }
1666 // Look for external lock for the same object.
1667 SafePointNode* sfn = this->as_SafePoint();
1668 JVMState* youngest_jvms = sfn->jvms();
1669 int max_depth = youngest_jvms->depth();
1670 for (int depth = 1; depth <= max_depth; depth++) {
1671 JVMState* jvms = youngest_jvms->of_depth(depth);
1672 int num_mon = jvms->nof_monitors();
1673 // Loop over monitors
1674 for (int idx = 0; idx < num_mon; idx++) {
1675 Node* obj_node = sfn->monitor_obj(jvms, idx);
1676 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock();
1677 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) {
1678 return true;
1679 }
1680 }
1681 }
1682 return false;
1683 }
1685 //=============================================================================
1686 uint UnlockNode::size_of() const { return sizeof(*this); }
1688 //=============================================================================
1689 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1691 // perform any generic optimizations first (returns 'this' or NULL)
1692 Node *result = SafePointNode::Ideal(phase, can_reshape);
1693 if (result != NULL) return result;
1694 // Don't bother trying to transform a dead node
1695 if (in(0) && in(0)->is_top()) return NULL;
1697 // Now see if we can optimize away this unlock. We don't actually
1698 // remove the unlocking here, we simply set the _eliminate flag which
1699 // prevents macro expansion from expanding the unlock. Since we don't
1700 // modify the graph, the value returned from this function is the
1701 // one computed above.
1702 // Escape state is defined after Parse phase.
1703 if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
1704 //
1705 // If we are unlocking an unescaped object, the lock/unlock is unnecessary.
1706 //
1707 ConnectionGraph *cgr = phase->C->congraph();
1708 if (cgr != NULL && cgr->not_global_escape(obj_node())) {
1709 assert(!is_eliminated() || is_coarsened(), "sanity");
1710 // The lock could be marked eliminated by lock coarsening
1711 // code during first IGVN before EA. Replace coarsened flag
1712 // to eliminate all associated locks/unlocks.
1713 this->set_non_esc_obj();
1714 }
1715 }
1716 return result;
1717 }