author | kvn |
Thu, 06 Mar 2008 10:53:33 -0800 | |
changeset 212 | cd4963e67949 |
parent 1 | 489c9b5090e2 |
child 236 | 9a04268c8eea |
permissions | -rw-r--r-- |
1 | 1 |
/* |
2 |
* Copyright 1997-2006 Sun Microsystems, Inc. All Rights Reserved. |
|
3 |
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
|
4 |
* |
|
5 |
* This code is free software; you can redistribute it and/or modify it |
|
6 |
* under the terms of the GNU General Public License version 2 only, as |
|
7 |
* published by the Free Software Foundation. |
|
8 |
* |
|
9 |
* This code is distributed in the hope that it will be useful, but WITHOUT |
|
10 |
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
|
11 |
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
|
12 |
* version 2 for more details (a copy is included in the LICENSE file that |
|
13 |
* accompanied this code). |
|
14 |
* |
|
15 |
* You should have received a copy of the GNU General Public License version |
|
16 |
* 2 along with this work; if not, write to the Free Software Foundation, |
|
17 |
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
|
18 |
* |
|
19 |
* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
|
20 |
* CA 95054 USA or visit www.sun.com if you need additional information or |
|
21 |
* have any questions. |
|
22 |
* |
|
23 |
*/ |
|
24 |
||
25 |
// Portions of code courtesy of Clifford Click |
|
26 |
||
27 |
// Optimization - Graph Style |
|
28 |
||
29 |
#include "incls/_precompiled.incl" |
|
30 |
#include "incls/_callnode.cpp.incl" |
|
31 |
||
32 |
//============================================================================= |
|
33 |
uint StartNode::size_of() const { return sizeof(*this); } |
|
34 |
uint StartNode::cmp( const Node &n ) const |
|
35 |
{ return _domain == ((StartNode&)n)._domain; } |
|
36 |
const Type *StartNode::bottom_type() const { return _domain; } |
|
37 |
const Type *StartNode::Value(PhaseTransform *phase) const { return _domain; } |
|
38 |
#ifndef PRODUCT |
|
39 |
void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);} |
|
40 |
#endif |
|
41 |
||
42 |
//------------------------------Ideal------------------------------------------ |
|
43 |
Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
|
44 |
return remove_dead_region(phase, can_reshape) ? this : NULL; |
|
45 |
} |
|
46 |
||
47 |
//------------------------------calling_convention----------------------------- |
|
48 |
void StartNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { |
|
49 |
Matcher::calling_convention( sig_bt, parm_regs, argcnt, false ); |
|
50 |
} |
|
51 |
||
52 |
//------------------------------Registers-------------------------------------- |
|
53 |
const RegMask &StartNode::in_RegMask(uint) const { |
|
54 |
return RegMask::Empty; |
|
55 |
} |
|
56 |
||
57 |
//------------------------------match------------------------------------------ |
|
58 |
// Construct projections for incoming parameters, and their RegMask info |
|
59 |
Node *StartNode::match( const ProjNode *proj, const Matcher *match ) { |
|
60 |
switch (proj->_con) { |
|
61 |
case TypeFunc::Control: |
|
62 |
case TypeFunc::I_O: |
|
63 |
case TypeFunc::Memory: |
|
64 |
return new (match->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); |
|
65 |
case TypeFunc::FramePtr: |
|
66 |
return new (match->C, 1) MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP); |
|
67 |
case TypeFunc::ReturnAdr: |
|
68 |
return new (match->C, 1) MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP); |
|
69 |
case TypeFunc::Parms: |
|
70 |
default: { |
|
71 |
uint parm_num = proj->_con - TypeFunc::Parms; |
|
72 |
const Type *t = _domain->field_at(proj->_con); |
|
73 |
if (t->base() == Type::Half) // 2nd half of Longs and Doubles |
|
74 |
return new (match->C, 1) ConNode(Type::TOP); |
|
75 |
uint ideal_reg = Matcher::base2reg[t->base()]; |
|
76 |
RegMask &rm = match->_calling_convention_mask[parm_num]; |
|
77 |
return new (match->C, 1) MachProjNode(this,proj->_con,rm,ideal_reg); |
|
78 |
} |
|
79 |
} |
|
80 |
return NULL; |
|
81 |
} |
|
82 |
||
83 |
//------------------------------StartOSRNode---------------------------------- |
|
84 |
// The method start node for an on stack replacement adapter |
|
85 |
||
86 |
//------------------------------osr_domain----------------------------- |
|
87 |
const TypeTuple *StartOSRNode::osr_domain() { |
|
88 |
const Type **fields = TypeTuple::fields(2); |
|
89 |
fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer |
|
90 |
||
91 |
return TypeTuple::make(TypeFunc::Parms+1, fields); |
|
92 |
} |
|
93 |
||
94 |
//============================================================================= |
|
95 |
const char * const ParmNode::names[TypeFunc::Parms+1] = { |
|
96 |
"Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms" |
|
97 |
}; |
|
98 |
||
99 |
#ifndef PRODUCT |
|
100 |
void ParmNode::dump_spec(outputStream *st) const { |
|
101 |
if( _con < TypeFunc::Parms ) { |
|
102 |
st->print(names[_con]); |
|
103 |
} else { |
|
104 |
st->print("Parm%d: ",_con-TypeFunc::Parms); |
|
105 |
// Verbose and WizardMode dump bottom_type for all nodes |
|
106 |
if( !Verbose && !WizardMode ) bottom_type()->dump_on(st); |
|
107 |
} |
|
108 |
} |
|
109 |
#endif |
|
110 |
||
111 |
uint ParmNode::ideal_reg() const { |
|
112 |
switch( _con ) { |
|
113 |
case TypeFunc::Control : // fall through |
|
114 |
case TypeFunc::I_O : // fall through |
|
115 |
case TypeFunc::Memory : return 0; |
|
116 |
case TypeFunc::FramePtr : // fall through |
|
117 |
case TypeFunc::ReturnAdr: return Op_RegP; |
|
118 |
default : assert( _con > TypeFunc::Parms, "" ); |
|
119 |
// fall through |
|
120 |
case TypeFunc::Parms : { |
|
121 |
// Type of argument being passed |
|
122 |
const Type *t = in(0)->as_Start()->_domain->field_at(_con); |
|
123 |
return Matcher::base2reg[t->base()]; |
|
124 |
} |
|
125 |
} |
|
126 |
ShouldNotReachHere(); |
|
127 |
return 0; |
|
128 |
} |
|
129 |
||
130 |
//============================================================================= |
|
131 |
ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) { |
|
132 |
init_req(TypeFunc::Control,cntrl); |
|
133 |
init_req(TypeFunc::I_O,i_o); |
|
134 |
init_req(TypeFunc::Memory,memory); |
|
135 |
init_req(TypeFunc::FramePtr,frameptr); |
|
136 |
init_req(TypeFunc::ReturnAdr,retadr); |
|
137 |
} |
|
138 |
||
139 |
Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
|
140 |
return remove_dead_region(phase, can_reshape) ? this : NULL; |
|
141 |
} |
|
142 |
||
143 |
const Type *ReturnNode::Value( PhaseTransform *phase ) const { |
|
144 |
return ( phase->type(in(TypeFunc::Control)) == Type::TOP) |
|
145 |
? Type::TOP |
|
146 |
: Type::BOTTOM; |
|
147 |
} |
|
148 |
||
149 |
// Do we Match on this edge index or not? No edges on return nodes |
|
150 |
uint ReturnNode::match_edge(uint idx) const { |
|
151 |
return 0; |
|
152 |
} |
|
153 |
||
154 |
||
155 |
#ifndef PRODUCT |
|
156 |
void ReturnNode::dump_req() const { |
|
157 |
// Dump the required inputs, enclosed in '(' and ')' |
|
158 |
uint i; // Exit value of loop |
|
159 |
for( i=0; i<req(); i++ ) { // For all required inputs |
|
160 |
if( i == TypeFunc::Parms ) tty->print("returns"); |
|
161 |
if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); |
|
162 |
else tty->print("_ "); |
|
163 |
} |
|
164 |
} |
|
165 |
#endif |
|
166 |
||
167 |
//============================================================================= |
|
168 |
RethrowNode::RethrowNode( |
|
169 |
Node* cntrl, |
|
170 |
Node* i_o, |
|
171 |
Node* memory, |
|
172 |
Node* frameptr, |
|
173 |
Node* ret_adr, |
|
174 |
Node* exception |
|
175 |
) : Node(TypeFunc::Parms + 1) { |
|
176 |
init_req(TypeFunc::Control , cntrl ); |
|
177 |
init_req(TypeFunc::I_O , i_o ); |
|
178 |
init_req(TypeFunc::Memory , memory ); |
|
179 |
init_req(TypeFunc::FramePtr , frameptr ); |
|
180 |
init_req(TypeFunc::ReturnAdr, ret_adr); |
|
181 |
init_req(TypeFunc::Parms , exception); |
|
182 |
} |
|
183 |
||
184 |
Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
|
185 |
return remove_dead_region(phase, can_reshape) ? this : NULL; |
|
186 |
} |
|
187 |
||
188 |
const Type *RethrowNode::Value( PhaseTransform *phase ) const { |
|
189 |
return (phase->type(in(TypeFunc::Control)) == Type::TOP) |
|
190 |
? Type::TOP |
|
191 |
: Type::BOTTOM; |
|
192 |
} |
|
193 |
||
194 |
uint RethrowNode::match_edge(uint idx) const { |
|
195 |
return 0; |
|
196 |
} |
|
197 |
||
198 |
#ifndef PRODUCT |
|
199 |
void RethrowNode::dump_req() const { |
|
200 |
// Dump the required inputs, enclosed in '(' and ')' |
|
201 |
uint i; // Exit value of loop |
|
202 |
for( i=0; i<req(); i++ ) { // For all required inputs |
|
203 |
if( i == TypeFunc::Parms ) tty->print("exception"); |
|
204 |
if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); |
|
205 |
else tty->print("_ "); |
|
206 |
} |
|
207 |
} |
|
208 |
#endif |
|
209 |
||
210 |
//============================================================================= |
|
211 |
// Do we Match on this edge index or not? Match only target address & method |
|
212 |
uint TailCallNode::match_edge(uint idx) const { |
|
213 |
return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; |
|
214 |
} |
|
215 |
||
216 |
//============================================================================= |
|
217 |
// Do we Match on this edge index or not? Match only target address & oop |
|
218 |
uint TailJumpNode::match_edge(uint idx) const { |
|
219 |
return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; |
|
220 |
} |
|
221 |
||
222 |
//============================================================================= |
|
223 |
JVMState::JVMState(ciMethod* method, JVMState* caller) { |
|
224 |
assert(method != NULL, "must be valid call site"); |
|
225 |
_method = method; |
|
226 |
debug_only(_bci = -99); // random garbage value |
|
227 |
debug_only(_map = (SafePointNode*)-1); |
|
228 |
_caller = caller; |
|
229 |
_depth = 1 + (caller == NULL ? 0 : caller->depth()); |
|
230 |
_locoff = TypeFunc::Parms; |
|
231 |
_stkoff = _locoff + _method->max_locals(); |
|
232 |
_monoff = _stkoff + _method->max_stack(); |
|
233 |
_endoff = _monoff; |
|
234 |
_sp = 0; |
|
235 |
} |
|
236 |
JVMState::JVMState(int stack_size) { |
|
237 |
_method = NULL; |
|
238 |
_bci = InvocationEntryBci; |
|
239 |
debug_only(_map = (SafePointNode*)-1); |
|
240 |
_caller = NULL; |
|
241 |
_depth = 1; |
|
242 |
_locoff = TypeFunc::Parms; |
|
243 |
_stkoff = _locoff; |
|
244 |
_monoff = _stkoff + stack_size; |
|
245 |
_endoff = _monoff; |
|
246 |
_sp = 0; |
|
247 |
} |
|
248 |
||
249 |
//--------------------------------of_depth------------------------------------- |
|
250 |
JVMState* JVMState::of_depth(int d) const { |
|
251 |
const JVMState* jvmp = this; |
|
252 |
assert(0 < d && (uint)d <= depth(), "oob"); |
|
253 |
for (int skip = depth() - d; skip > 0; skip--) { |
|
254 |
jvmp = jvmp->caller(); |
|
255 |
} |
|
256 |
assert(jvmp->depth() == (uint)d, "found the right one"); |
|
257 |
return (JVMState*)jvmp; |
|
258 |
} |
|
259 |
||
260 |
//-----------------------------same_calls_as----------------------------------- |
|
261 |
bool JVMState::same_calls_as(const JVMState* that) const { |
|
262 |
if (this == that) return true; |
|
263 |
if (this->depth() != that->depth()) return false; |
|
264 |
const JVMState* p = this; |
|
265 |
const JVMState* q = that; |
|
266 |
for (;;) { |
|
267 |
if (p->_method != q->_method) return false; |
|
268 |
if (p->_method == NULL) return true; // bci is irrelevant |
|
269 |
if (p->_bci != q->_bci) return false; |
|
270 |
p = p->caller(); |
|
271 |
q = q->caller(); |
|
272 |
if (p == q) return true; |
|
273 |
assert(p != NULL && q != NULL, "depth check ensures we don't run off end"); |
|
274 |
} |
|
275 |
} |
|
276 |
||
277 |
//------------------------------debug_start------------------------------------ |
|
278 |
uint JVMState::debug_start() const { |
|
279 |
debug_only(JVMState* jvmroot = of_depth(1)); |
|
280 |
assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last"); |
|
281 |
return of_depth(1)->locoff(); |
|
282 |
} |
|
283 |
||
284 |
//-------------------------------debug_end------------------------------------- |
|
285 |
uint JVMState::debug_end() const { |
|
286 |
debug_only(JVMState* jvmroot = of_depth(1)); |
|
287 |
assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last"); |
|
288 |
return endoff(); |
|
289 |
} |
|
290 |
||
291 |
//------------------------------debug_depth------------------------------------ |
|
292 |
uint JVMState::debug_depth() const { |
|
293 |
uint total = 0; |
|
294 |
for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) { |
|
295 |
total += jvmp->debug_size(); |
|
296 |
} |
|
297 |
return total; |
|
298 |
} |
|
299 |
||
300 |
//------------------------------format_helper---------------------------------- |
|
301 |
// Given an allocation (a Chaitin object) and a Node decide if the Node carries |
|
302 |
// any defined value or not. If it does, print out the register or constant. |
|
303 |
#ifndef PRODUCT |
|
304 |
static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i ) { |
|
305 |
if (n == NULL) { st->print(" NULL"); return; } |
|
306 |
if( OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined |
|
307 |
char buf[50]; |
|
308 |
regalloc->dump_register(n,buf); |
|
309 |
st->print(" %s%d]=%s",msg,i,buf); |
|
310 |
} else { // No register, but might be constant |
|
311 |
const Type *t = n->bottom_type(); |
|
312 |
switch (t->base()) { |
|
313 |
case Type::Int: |
|
314 |
st->print(" %s%d]=#"INT32_FORMAT,msg,i,t->is_int()->get_con()); |
|
315 |
break; |
|
316 |
case Type::AnyPtr: |
|
317 |
assert( t == TypePtr::NULL_PTR, "" ); |
|
318 |
st->print(" %s%d]=#NULL",msg,i); |
|
319 |
break; |
|
320 |
case Type::AryPtr: |
|
321 |
case Type::KlassPtr: |
|
322 |
case Type::InstPtr: |
|
323 |
st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->isa_oopptr()->const_oop()); |
|
324 |
break; |
|
325 |
case Type::RawPtr: |
|
326 |
st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,t->is_rawptr()); |
|
327 |
break; |
|
328 |
case Type::DoubleCon: |
|
329 |
st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d); |
|
330 |
break; |
|
331 |
case Type::FloatCon: |
|
332 |
st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f); |
|
333 |
break; |
|
334 |
case Type::Long: |
|
335 |
st->print(" %s%d]=#"INT64_FORMAT,msg,i,t->is_long()->get_con()); |
|
336 |
break; |
|
337 |
case Type::Half: |
|
338 |
case Type::Top: |
|
339 |
st->print(" %s%d]=_",msg,i); |
|
340 |
break; |
|
341 |
default: ShouldNotReachHere(); |
|
342 |
} |
|
343 |
} |
|
344 |
} |
|
345 |
#endif |
|
346 |
||
347 |
//------------------------------format----------------------------------------- |
|
348 |
#ifndef PRODUCT |
|
349 |
void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const { |
|
350 |
st->print(" #"); |
|
351 |
if( _method ) { |
|
352 |
_method->print_short_name(st); |
|
353 |
st->print(" @ bci:%d ",_bci); |
|
354 |
} else { |
|
355 |
st->print_cr(" runtime stub "); |
|
356 |
return; |
|
357 |
} |
|
358 |
if (n->is_MachSafePoint()) { |
|
359 |
MachSafePointNode *mcall = n->as_MachSafePoint(); |
|
360 |
uint i; |
|
361 |
// Print locals |
|
362 |
for( i = 0; i < (uint)loc_size(); i++ ) |
|
363 |
format_helper( regalloc, st, mcall->local(this, i), "L[", i ); |
|
364 |
// Print stack |
|
365 |
for (i = 0; i < (uint)stk_size(); i++) { |
|
366 |
if ((uint)(_stkoff + i) >= mcall->len()) |
|
367 |
st->print(" oob "); |
|
368 |
else |
|
369 |
format_helper( regalloc, st, mcall->stack(this, i), "STK[", i ); |
|
370 |
} |
|
371 |
for (i = 0; (int)i < nof_monitors(); i++) { |
|
372 |
Node *box = mcall->monitor_box(this, i); |
|
373 |
Node *obj = mcall->monitor_obj(this, i); |
|
374 |
if ( OptoReg::is_valid(regalloc->get_reg_first(box)) ) { |
|
375 |
while( !box->is_BoxLock() ) box = box->in(1); |
|
376 |
format_helper( regalloc, st, box, "MON-BOX[", i ); |
|
377 |
} else { |
|
378 |
OptoReg::Name box_reg = BoxLockNode::stack_slot(box); |
|
379 |
st->print(" MON-BOX%d=%s+%d", |
|
380 |
i, |
|
381 |
OptoReg::regname(OptoReg::c_frame_pointer), |
|
382 |
regalloc->reg2offset(box_reg)); |
|
383 |
} |
|
384 |
format_helper( regalloc, st, obj, "MON-OBJ[", i ); |
|
385 |
} |
|
386 |
} |
|
387 |
st->print_cr(""); |
|
388 |
if (caller() != NULL) caller()->format(regalloc, n, st); |
|
389 |
} |
|
390 |
#endif |
|
391 |
||
392 |
#ifndef PRODUCT |
|
393 |
void JVMState::dump_spec(outputStream *st) const { |
|
394 |
if (_method != NULL) { |
|
395 |
bool printed = false; |
|
396 |
if (!Verbose) { |
|
397 |
// The JVMS dumps make really, really long lines. |
|
398 |
// Take out the most boring parts, which are the package prefixes. |
|
399 |
char buf[500]; |
|
400 |
stringStream namest(buf, sizeof(buf)); |
|
401 |
_method->print_short_name(&namest); |
|
402 |
if (namest.count() < sizeof(buf)) { |
|
403 |
const char* name = namest.base(); |
|
404 |
if (name[0] == ' ') ++name; |
|
405 |
const char* endcn = strchr(name, ':'); // end of class name |
|
406 |
if (endcn == NULL) endcn = strchr(name, '('); |
|
407 |
if (endcn == NULL) endcn = name + strlen(name); |
|
408 |
while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/') |
|
409 |
--endcn; |
|
410 |
st->print(" %s", endcn); |
|
411 |
printed = true; |
|
412 |
} |
|
413 |
} |
|
414 |
if (!printed) |
|
415 |
_method->print_short_name(st); |
|
416 |
st->print(" @ bci:%d",_bci); |
|
417 |
} else { |
|
418 |
st->print(" runtime stub"); |
|
419 |
} |
|
420 |
if (caller() != NULL) caller()->dump_spec(st); |
|
421 |
} |
|
422 |
#endif |
|
423 |
||
424 |
#ifndef PRODUCT |
|
425 |
void JVMState::dump_on(outputStream* st) const { |
|
426 |
if (_map && !((uintptr_t)_map & 1)) { |
|
427 |
if (_map->len() > _map->req()) { // _map->has_exceptions() |
|
428 |
Node* ex = _map->in(_map->req()); // _map->next_exception() |
|
429 |
// skip the first one; it's already being printed |
|
430 |
while (ex != NULL && ex->len() > ex->req()) { |
|
431 |
ex = ex->in(ex->req()); // ex->next_exception() |
|
432 |
ex->dump(1); |
|
433 |
} |
|
434 |
} |
|
435 |
_map->dump(2); |
|
436 |
} |
|
437 |
st->print("JVMS depth=%d loc=%d stk=%d mon=%d end=%d mondepth=%d sp=%d bci=%d method=", |
|
438 |
depth(), locoff(), stkoff(), monoff(), endoff(), monitor_depth(), sp(), bci()); |
|
439 |
if (_method == NULL) { |
|
440 |
st->print_cr("(none)"); |
|
441 |
} else { |
|
442 |
_method->print_name(st); |
|
443 |
st->cr(); |
|
444 |
if (bci() >= 0 && bci() < _method->code_size()) { |
|
445 |
st->print(" bc: "); |
|
446 |
_method->print_codes_on(bci(), bci()+1, st); |
|
447 |
} |
|
448 |
} |
|
449 |
if (caller() != NULL) { |
|
450 |
caller()->dump_on(st); |
|
451 |
} |
|
452 |
} |
|
453 |
||
454 |
// Extra way to dump a jvms from the debugger, |
|
455 |
// to avoid a bug with C++ member function calls. |
|
456 |
void dump_jvms(JVMState* jvms) { |
|
457 |
jvms->dump(); |
|
458 |
} |
|
459 |
#endif |
|
460 |
||
461 |
//--------------------------clone_shallow-------------------------------------- |
|
462 |
JVMState* JVMState::clone_shallow(Compile* C) const { |
|
463 |
JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0); |
|
464 |
n->set_bci(_bci); |
|
465 |
n->set_locoff(_locoff); |
|
466 |
n->set_stkoff(_stkoff); |
|
467 |
n->set_monoff(_monoff); |
|
468 |
n->set_endoff(_endoff); |
|
469 |
n->set_sp(_sp); |
|
470 |
n->set_map(_map); |
|
471 |
return n; |
|
472 |
} |
|
473 |
||
474 |
//---------------------------clone_deep---------------------------------------- |
|
475 |
JVMState* JVMState::clone_deep(Compile* C) const { |
|
476 |
JVMState* n = clone_shallow(C); |
|
477 |
for (JVMState* p = n; p->_caller != NULL; p = p->_caller) { |
|
478 |
p->_caller = p->_caller->clone_shallow(C); |
|
479 |
} |
|
480 |
assert(n->depth() == depth(), "sanity"); |
|
481 |
assert(n->debug_depth() == debug_depth(), "sanity"); |
|
482 |
return n; |
|
483 |
} |
|
484 |
||
485 |
//============================================================================= |
|
486 |
uint CallNode::cmp( const Node &n ) const |
|
487 |
{ return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; } |
|
488 |
#ifndef PRODUCT |
|
489 |
void CallNode::dump_req() const { |
|
490 |
// Dump the required inputs, enclosed in '(' and ')' |
|
491 |
uint i; // Exit value of loop |
|
492 |
for( i=0; i<req(); i++ ) { // For all required inputs |
|
493 |
if( i == TypeFunc::Parms ) tty->print("("); |
|
494 |
if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); |
|
495 |
else tty->print("_ "); |
|
496 |
} |
|
497 |
tty->print(")"); |
|
498 |
} |
|
499 |
||
500 |
void CallNode::dump_spec(outputStream *st) const { |
|
501 |
st->print(" "); |
|
502 |
tf()->dump_on(st); |
|
503 |
if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt); |
|
504 |
if (jvms() != NULL) jvms()->dump_spec(st); |
|
505 |
} |
|
506 |
#endif |
|
507 |
||
508 |
const Type *CallNode::bottom_type() const { return tf()->range(); } |
|
509 |
const Type *CallNode::Value(PhaseTransform *phase) const { |
|
510 |
if (phase->type(in(0)) == Type::TOP) return Type::TOP; |
|
511 |
return tf()->range(); |
|
512 |
} |
|
513 |
||
514 |
//------------------------------calling_convention----------------------------- |
|
515 |
void CallNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { |
|
516 |
// Use the standard compiler calling convention |
|
517 |
Matcher::calling_convention( sig_bt, parm_regs, argcnt, true ); |
|
518 |
} |
|
519 |
||
520 |
||
521 |
//------------------------------match------------------------------------------ |
|
522 |
// Construct projections for control, I/O, memory-fields, ..., and |
|
523 |
// return result(s) along with their RegMask info |
|
524 |
Node *CallNode::match( const ProjNode *proj, const Matcher *match ) { |
|
525 |
switch (proj->_con) { |
|
526 |
case TypeFunc::Control: |
|
527 |
case TypeFunc::I_O: |
|
528 |
case TypeFunc::Memory: |
|
529 |
return new (match->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); |
|
530 |
||
531 |
case TypeFunc::Parms+1: // For LONG & DOUBLE returns |
|
532 |
assert(tf()->_range->field_at(TypeFunc::Parms+1) == Type::HALF, ""); |
|
533 |
// 2nd half of doubles and longs |
|
534 |
return new (match->C, 1) MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad); |
|
535 |
||
536 |
case TypeFunc::Parms: { // Normal returns |
|
537 |
uint ideal_reg = Matcher::base2reg[tf()->range()->field_at(TypeFunc::Parms)->base()]; |
|
538 |
OptoRegPair regs = is_CallRuntime() |
|
539 |
? match->c_return_value(ideal_reg,true) // Calls into C runtime |
|
540 |
: match-> return_value(ideal_reg,true); // Calls into compiled Java code |
|
541 |
RegMask rm = RegMask(regs.first()); |
|
542 |
if( OptoReg::is_valid(regs.second()) ) |
|
543 |
rm.Insert( regs.second() ); |
|
544 |
return new (match->C, 1) MachProjNode(this,proj->_con,rm,ideal_reg); |
|
545 |
} |
|
546 |
||
547 |
case TypeFunc::ReturnAdr: |
|
548 |
case TypeFunc::FramePtr: |
|
549 |
default: |
|
550 |
ShouldNotReachHere(); |
|
551 |
} |
|
552 |
return NULL; |
|
553 |
} |
|
554 |
||
555 |
// Do we Match on this edge index or not? Match no edges |
|
556 |
uint CallNode::match_edge(uint idx) const { |
|
557 |
return 0; |
|
558 |
} |
|
559 |
||
560 |
//============================================================================= |
|
561 |
uint CallJavaNode::size_of() const { return sizeof(*this); } |
|
562 |
uint CallJavaNode::cmp( const Node &n ) const { |
|
563 |
CallJavaNode &call = (CallJavaNode&)n; |
|
564 |
return CallNode::cmp(call) && _method == call._method; |
|
565 |
} |
|
566 |
#ifndef PRODUCT |
|
567 |
void CallJavaNode::dump_spec(outputStream *st) const { |
|
568 |
if( _method ) _method->print_short_name(st); |
|
569 |
CallNode::dump_spec(st); |
|
570 |
} |
|
571 |
#endif |
|
572 |
||
573 |
//============================================================================= |
|
574 |
uint CallStaticJavaNode::size_of() const { return sizeof(*this); } |
|
575 |
uint CallStaticJavaNode::cmp( const Node &n ) const { |
|
576 |
CallStaticJavaNode &call = (CallStaticJavaNode&)n; |
|
577 |
return CallJavaNode::cmp(call); |
|
578 |
} |
|
579 |
||
580 |
//----------------------------uncommon_trap_request---------------------------- |
|
581 |
// If this is an uncommon trap, return the request code, else zero. |
|
582 |
int CallStaticJavaNode::uncommon_trap_request() const { |
|
583 |
if (_name != NULL && !strcmp(_name, "uncommon_trap")) { |
|
584 |
return extract_uncommon_trap_request(this); |
|
585 |
} |
|
586 |
return 0; |
|
587 |
} |
|
588 |
int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) { |
|
589 |
#ifndef PRODUCT |
|
590 |
if (!(call->req() > TypeFunc::Parms && |
|
591 |
call->in(TypeFunc::Parms) != NULL && |
|
592 |
call->in(TypeFunc::Parms)->is_Con())) { |
|
593 |
assert(_in_dump_cnt != 0, "OK if dumping"); |
|
594 |
tty->print("[bad uncommon trap]"); |
|
595 |
return 0; |
|
596 |
} |
|
597 |
#endif |
|
598 |
return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con(); |
|
599 |
} |
|
600 |
||
601 |
#ifndef PRODUCT |
|
602 |
void CallStaticJavaNode::dump_spec(outputStream *st) const { |
|
603 |
st->print("# Static "); |
|
604 |
if (_name != NULL) { |
|
605 |
st->print("%s", _name); |
|
606 |
int trap_req = uncommon_trap_request(); |
|
607 |
if (trap_req != 0) { |
|
608 |
char buf[100]; |
|
609 |
st->print("(%s)", |
|
610 |
Deoptimization::format_trap_request(buf, sizeof(buf), |
|
611 |
trap_req)); |
|
612 |
} |
|
613 |
st->print(" "); |
|
614 |
} |
|
615 |
CallJavaNode::dump_spec(st); |
|
616 |
} |
|
617 |
#endif |
|
618 |
||
619 |
//============================================================================= |
|
620 |
uint CallDynamicJavaNode::size_of() const { return sizeof(*this); } |
|
621 |
uint CallDynamicJavaNode::cmp( const Node &n ) const { |
|
622 |
CallDynamicJavaNode &call = (CallDynamicJavaNode&)n; |
|
623 |
return CallJavaNode::cmp(call); |
|
624 |
} |
|
625 |
#ifndef PRODUCT |
|
626 |
void CallDynamicJavaNode::dump_spec(outputStream *st) const { |
|
627 |
st->print("# Dynamic "); |
|
628 |
CallJavaNode::dump_spec(st); |
|
629 |
} |
|
630 |
#endif |
|
631 |
||
632 |
//============================================================================= |
|
633 |
uint CallRuntimeNode::size_of() const { return sizeof(*this); } |
|
634 |
uint CallRuntimeNode::cmp( const Node &n ) const { |
|
635 |
CallRuntimeNode &call = (CallRuntimeNode&)n; |
|
636 |
return CallNode::cmp(call) && !strcmp(_name,call._name); |
|
637 |
} |
|
638 |
#ifndef PRODUCT |
|
639 |
void CallRuntimeNode::dump_spec(outputStream *st) const { |
|
640 |
st->print("# "); |
|
641 |
st->print(_name); |
|
642 |
CallNode::dump_spec(st); |
|
643 |
} |
|
644 |
#endif |
|
645 |
||
646 |
//------------------------------calling_convention----------------------------- |
|
647 |
void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { |
|
648 |
Matcher::c_calling_convention( sig_bt, parm_regs, argcnt ); |
|
649 |
} |
|
650 |
||
651 |
//============================================================================= |
|
652 |
//------------------------------calling_convention----------------------------- |
|
653 |
||
654 |
||
655 |
//============================================================================= |
|
656 |
#ifndef PRODUCT |
|
657 |
void CallLeafNode::dump_spec(outputStream *st) const { |
|
658 |
st->print("# "); |
|
659 |
st->print(_name); |
|
660 |
CallNode::dump_spec(st); |
|
661 |
} |
|
662 |
#endif |
|
663 |
||
664 |
//============================================================================= |
|
665 |
||
666 |
void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) { |
|
667 |
assert(verify_jvms(jvms), "jvms must match"); |
|
668 |
int loc = jvms->locoff() + idx; |
|
669 |
if (in(loc)->is_top() && idx > 0 && !c->is_top() ) { |
|
670 |
// If current local idx is top then local idx - 1 could |
|
671 |
// be a long/double that needs to be killed since top could |
|
672 |
// represent the 2nd half ofthe long/double. |
|
673 |
uint ideal = in(loc -1)->ideal_reg(); |
|
674 |
if (ideal == Op_RegD || ideal == Op_RegL) { |
|
675 |
// set other (low index) half to top |
|
676 |
set_req(loc - 1, in(loc)); |
|
677 |
} |
|
678 |
} |
|
679 |
set_req(loc, c); |
|
680 |
} |
|
681 |
||
682 |
uint SafePointNode::size_of() const { return sizeof(*this); } |
|
683 |
uint SafePointNode::cmp( const Node &n ) const { |
|
684 |
return (&n == this); // Always fail except on self |
|
685 |
} |
|
686 |
||
687 |
//-------------------------set_next_exception---------------------------------- |
|
688 |
void SafePointNode::set_next_exception(SafePointNode* n) { |
|
689 |
assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception"); |
|
690 |
if (len() == req()) { |
|
691 |
if (n != NULL) add_prec(n); |
|
692 |
} else { |
|
693 |
set_prec(req(), n); |
|
694 |
} |
|
695 |
} |
|
696 |
||
697 |
||
698 |
//----------------------------next_exception----------------------------------- |
|
699 |
SafePointNode* SafePointNode::next_exception() const { |
|
700 |
if (len() == req()) { |
|
701 |
return NULL; |
|
702 |
} else { |
|
703 |
Node* n = in(req()); |
|
704 |
assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges"); |
|
705 |
return (SafePointNode*) n; |
|
706 |
} |
|
707 |
} |
|
708 |
||
709 |
||
710 |
//------------------------------Ideal------------------------------------------ |
|
711 |
// Skip over any collapsed Regions |
|
712 |
Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
|
713 |
if (remove_dead_region(phase, can_reshape)) return this; |
|
714 |
||
715 |
return NULL; |
|
716 |
} |
|
717 |
||
718 |
//------------------------------Identity--------------------------------------- |
|
719 |
// Remove obviously duplicate safepoints |
|
720 |
Node *SafePointNode::Identity( PhaseTransform *phase ) { |
|
721 |
||
722 |
// If you have back to back safepoints, remove one |
|
723 |
if( in(TypeFunc::Control)->is_SafePoint() ) |
|
724 |
return in(TypeFunc::Control); |
|
725 |
||
726 |
if( in(0)->is_Proj() ) { |
|
727 |
Node *n0 = in(0)->in(0); |
|
728 |
// Check if he is a call projection (except Leaf Call) |
|
729 |
if( n0->is_Catch() ) { |
|
730 |
n0 = n0->in(0)->in(0); |
|
731 |
assert( n0->is_Call(), "expect a call here" ); |
|
732 |
} |
|
733 |
if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) { |
|
734 |
// Useless Safepoint, so remove it |
|
735 |
return in(TypeFunc::Control); |
|
736 |
} |
|
737 |
} |
|
738 |
||
739 |
return this; |
|
740 |
} |
|
741 |
||
742 |
//------------------------------Value------------------------------------------ |
|
743 |
const Type *SafePointNode::Value( PhaseTransform *phase ) const { |
|
744 |
if( phase->type(in(0)) == Type::TOP ) return Type::TOP; |
|
745 |
if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop |
|
746 |
return Type::CONTROL; |
|
747 |
} |
|
748 |
||
749 |
#ifndef PRODUCT |
|
750 |
void SafePointNode::dump_spec(outputStream *st) const { |
|
751 |
st->print(" SafePoint "); |
|
752 |
} |
|
753 |
#endif |
|
754 |
||
755 |
const RegMask &SafePointNode::in_RegMask(uint idx) const { |
|
756 |
if( idx < TypeFunc::Parms ) return RegMask::Empty; |
|
757 |
// Values outside the domain represent debug info |
|
758 |
return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); |
|
759 |
} |
|
760 |
const RegMask &SafePointNode::out_RegMask() const { |
|
761 |
return RegMask::Empty; |
|
762 |
} |
|
763 |
||
764 |
||
765 |
void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) { |
|
766 |
assert((int)grow_by > 0, "sanity"); |
|
767 |
int monoff = jvms->monoff(); |
|
768 |
int endoff = jvms->endoff(); |
|
769 |
assert(endoff == (int)req(), "no other states or debug info after me"); |
|
770 |
Node* top = Compile::current()->top(); |
|
771 |
for (uint i = 0; i < grow_by; i++) { |
|
772 |
ins_req(monoff, top); |
|
773 |
} |
|
774 |
jvms->set_monoff(monoff + grow_by); |
|
775 |
jvms->set_endoff(endoff + grow_by); |
|
776 |
} |
|
777 |
||
778 |
void SafePointNode::push_monitor(const FastLockNode *lock) { |
|
779 |
// Add a LockNode, which points to both the original BoxLockNode (the |
|
780 |
// stack space for the monitor) and the Object being locked. |
|
781 |
const int MonitorEdges = 2; |
|
782 |
assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); |
|
783 |
assert(req() == jvms()->endoff(), "correct sizing"); |
|
784 |
if (GenerateSynchronizationCode) { |
|
785 |
add_req(lock->box_node()); |
|
786 |
add_req(lock->obj_node()); |
|
787 |
} else { |
|
788 |
add_req(NULL); |
|
789 |
add_req(NULL); |
|
790 |
} |
|
791 |
jvms()->set_endoff(req()); |
|
792 |
} |
|
793 |
||
794 |
void SafePointNode::pop_monitor() { |
|
795 |
// Delete last monitor from debug info |
|
796 |
debug_only(int num_before_pop = jvms()->nof_monitors()); |
|
797 |
const int MonitorEdges = (1<<JVMState::logMonitorEdges); |
|
798 |
int endoff = jvms()->endoff(); |
|
799 |
int new_endoff = endoff - MonitorEdges; |
|
800 |
jvms()->set_endoff(new_endoff); |
|
801 |
while (endoff > new_endoff) del_req(--endoff); |
|
802 |
assert(jvms()->nof_monitors() == num_before_pop-1, ""); |
|
803 |
} |
|
804 |
||
805 |
Node *SafePointNode::peek_monitor_box() const { |
|
806 |
int mon = jvms()->nof_monitors() - 1; |
|
807 |
assert(mon >= 0, "most have a monitor"); |
|
808 |
return monitor_box(jvms(), mon); |
|
809 |
} |
|
810 |
||
811 |
Node *SafePointNode::peek_monitor_obj() const { |
|
812 |
int mon = jvms()->nof_monitors() - 1; |
|
813 |
assert(mon >= 0, "most have a monitor"); |
|
814 |
return monitor_obj(jvms(), mon); |
|
815 |
} |
|
816 |
||
817 |
// Do we Match on this edge index or not? Match no edges |
|
818 |
uint SafePointNode::match_edge(uint idx) const { |
|
819 |
if( !needs_polling_address_input() ) |
|
820 |
return 0; |
|
821 |
||
822 |
return (TypeFunc::Parms == idx); |
|
823 |
} |
|
824 |
||
825 |
//============================================================================= |
|
826 |
uint AllocateNode::size_of() const { return sizeof(*this); } |
|
827 |
||
828 |
AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype, |
|
829 |
Node *ctrl, Node *mem, Node *abio, |
|
830 |
Node *size, Node *klass_node, Node *initial_test) |
|
831 |
: CallNode(atype, NULL, TypeRawPtr::BOTTOM) |
|
832 |
{ |
|
833 |
init_class_id(Class_Allocate); |
|
834 |
init_flags(Flag_is_macro); |
|
212
cd4963e67949
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
1
diff
changeset
|
835 |
_is_scalar_replaceable = false; |
1 | 836 |
Node *topnode = C->top(); |
837 |
||
838 |
init_req( TypeFunc::Control , ctrl ); |
|
839 |
init_req( TypeFunc::I_O , abio ); |
|
840 |
init_req( TypeFunc::Memory , mem ); |
|
841 |
init_req( TypeFunc::ReturnAdr, topnode ); |
|
842 |
init_req( TypeFunc::FramePtr , topnode ); |
|
843 |
init_req( AllocSize , size); |
|
844 |
init_req( KlassNode , klass_node); |
|
845 |
init_req( InitialTest , initial_test); |
|
846 |
init_req( ALength , topnode); |
|
847 |
C->add_macro_node(this); |
|
848 |
} |
|
849 |
||
850 |
//============================================================================= |
|
851 |
uint AllocateArrayNode::size_of() const { return sizeof(*this); } |
|
852 |
||
853 |
//============================================================================= |
|
854 |
uint LockNode::size_of() const { return sizeof(*this); } |
|
855 |
||
856 |
// Redundant lock elimination |
|
857 |
// |
|
858 |
// There are various patterns of locking where we release and |
|
859 |
// immediately reacquire a lock in a piece of code where no operations |
|
860 |
// occur in between that would be observable. In those cases we can |
|
861 |
// skip releasing and reacquiring the lock without violating any |
|
862 |
// fairness requirements. Doing this around a loop could cause a lock |
|
863 |
// to be held for a very long time so we concentrate on non-looping |
|
864 |
// control flow. We also require that the operations are fully |
|
865 |
// redundant meaning that we don't introduce new lock operations on |
|
866 |
// some paths so to be able to eliminate it on others ala PRE. This |
|
867 |
// would probably require some more extensive graph manipulation to |
|
868 |
// guarantee that the memory edges were all handled correctly. |
|
869 |
// |
|
870 |
// Assuming p is a simple predicate which can't trap in any way and s |
|
871 |
// is a synchronized method consider this code: |
|
872 |
// |
|
873 |
// s(); |
|
874 |
// if (p) |
|
875 |
// s(); |
|
876 |
// else |
|
877 |
// s(); |
|
878 |
// s(); |
|
879 |
// |
|
880 |
// 1. The unlocks of the first call to s can be eliminated if the |
|
881 |
// locks inside the then and else branches are eliminated. |
|
882 |
// |
|
883 |
// 2. The unlocks of the then and else branches can be eliminated if |
|
884 |
// the lock of the final call to s is eliminated. |
|
885 |
// |
|
886 |
// Either of these cases subsumes the simple case of sequential control flow |
|
887 |
// |
|
888 |
// Addtionally we can eliminate versions without the else case: |
|
889 |
// |
|
890 |
// s(); |
|
891 |
// if (p) |
|
892 |
// s(); |
|
893 |
// s(); |
|
894 |
// |
|
895 |
// 3. In this case we eliminate the unlock of the first s, the lock |
|
896 |
// and unlock in the then case and the lock in the final s. |
|
897 |
// |
|
898 |
// Note also that in all these cases the then/else pieces don't have |
|
899 |
// to be trivial as long as they begin and end with synchronization |
|
900 |
// operations. |
|
901 |
// |
|
902 |
// s(); |
|
903 |
// if (p) |
|
904 |
// s(); |
|
905 |
// f(); |
|
906 |
// s(); |
|
907 |
// s(); |
|
908 |
// |
|
909 |
// The code will work properly for this case, leaving in the unlock |
|
910 |
// before the call to f and the relock after it. |
|
911 |
// |
|
912 |
// A potentially interesting case which isn't handled here is when the |
|
913 |
// locking is partially redundant. |
|
914 |
// |
|
915 |
// s(); |
|
916 |
// if (p) |
|
917 |
// s(); |
|
918 |
// |
|
919 |
// This could be eliminated putting unlocking on the else case and |
|
920 |
// eliminating the first unlock and the lock in the then side. |
|
921 |
// Alternatively the unlock could be moved out of the then side so it |
|
922 |
// was after the merge and the first unlock and second lock |
|
923 |
// eliminated. This might require less manipulation of the memory |
|
924 |
// state to get correct. |
|
925 |
// |
|
926 |
// Additionally we might allow work between a unlock and lock before |
|
927 |
// giving up eliminating the locks. The current code disallows any |
|
928 |
// conditional control flow between these operations. A formulation |
|
929 |
// similar to partial redundancy elimination computing the |
|
930 |
// availability of unlocking and the anticipatability of locking at a |
|
931 |
// program point would allow detection of fully redundant locking with |
|
932 |
// some amount of work in between. I'm not sure how often I really |
|
933 |
// think that would occur though. Most of the cases I've seen |
|
934 |
// indicate it's likely non-trivial work would occur in between. |
|
935 |
// There may be other more complicated constructs where we could |
|
936 |
// eliminate locking but I haven't seen any others appear as hot or |
|
937 |
// interesting. |
|
938 |
// |
|
939 |
// Locking and unlocking have a canonical form in ideal that looks |
|
940 |
// roughly like this: |
|
941 |
// |
|
942 |
// <obj> |
|
943 |
// | \\------+ |
|
944 |
// | \ \ |
|
945 |
// | BoxLock \ |
|
946 |
// | | | \ |
|
947 |
// | | \ \ |
|
948 |
// | | FastLock |
|
949 |
// | | / |
|
950 |
// | | / |
|
951 |
// | | | |
|
952 |
// |
|
953 |
// Lock |
|
954 |
// | |
|
955 |
// Proj #0 |
|
956 |
// | |
|
957 |
// MembarAcquire |
|
958 |
// | |
|
959 |
// Proj #0 |
|
960 |
// |
|
961 |
// MembarRelease |
|
962 |
// | |
|
963 |
// Proj #0 |
|
964 |
// | |
|
965 |
// Unlock |
|
966 |
// | |
|
967 |
// Proj #0 |
|
968 |
// |
|
969 |
// |
|
970 |
// This code proceeds by processing Lock nodes during PhaseIterGVN |
|
971 |
// and searching back through its control for the proper code |
|
972 |
// patterns. Once it finds a set of lock and unlock operations to |
|
973 |
// eliminate they are marked as eliminatable which causes the |
|
974 |
// expansion of the Lock and Unlock macro nodes to make the operation a NOP |
|
975 |
// |
|
976 |
//============================================================================= |
|
977 |
||
978 |
// |
|
979 |
// Utility function to skip over uninteresting control nodes. Nodes skipped are: |
|
980 |
// - copy regions. (These may not have been optimized away yet.) |
|
981 |
// - eliminated locking nodes |
|
982 |
// |
|
983 |
static Node *next_control(Node *ctrl) { |
|
984 |
if (ctrl == NULL) |
|
985 |
return NULL; |
|
986 |
while (1) { |
|
987 |
if (ctrl->is_Region()) { |
|
988 |
RegionNode *r = ctrl->as_Region(); |
|
989 |
Node *n = r->is_copy(); |
|
990 |
if (n == NULL) |
|
991 |
break; // hit a region, return it |
|
992 |
else |
|
993 |
ctrl = n; |
|
994 |
} else if (ctrl->is_Proj()) { |
|
995 |
Node *in0 = ctrl->in(0); |
|
996 |
if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) { |
|
997 |
ctrl = in0->in(0); |
|
998 |
} else { |
|
999 |
break; |
|
1000 |
} |
|
1001 |
} else { |
|
1002 |
break; // found an interesting control |
|
1003 |
} |
|
1004 |
} |
|
1005 |
return ctrl; |
|
1006 |
} |
|
1007 |
// |
|
1008 |
// Given a control, see if it's the control projection of an Unlock which |
|
1009 |
// operating on the same object as lock. |
|
1010 |
// |
|
1011 |
bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock, |
|
1012 |
GrowableArray<AbstractLockNode*> &lock_ops) { |
|
1013 |
ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL; |
|
1014 |
if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) { |
|
1015 |
Node *n = ctrl_proj->in(0); |
|
1016 |
if (n != NULL && n->is_Unlock()) { |
|
1017 |
UnlockNode *unlock = n->as_Unlock(); |
|
1018 |
if ((lock->obj_node() == unlock->obj_node()) && |
|
1019 |
(lock->box_node() == unlock->box_node()) && !unlock->is_eliminated()) { |
|
1020 |
lock_ops.append(unlock); |
|
1021 |
return true; |
|
1022 |
} |
|
1023 |
} |
|
1024 |
} |
|
1025 |
return false; |
|
1026 |
} |
|
1027 |
||
1028 |
// |
|
1029 |
// Find the lock matching an unlock. Returns null if a safepoint |
|
1030 |
// or complicated control is encountered first. |
|
1031 |
LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) { |
|
1032 |
LockNode *lock_result = NULL; |
|
1033 |
// find the matching lock, or an intervening safepoint |
|
1034 |
Node *ctrl = next_control(unlock->in(0)); |
|
1035 |
while (1) { |
|
1036 |
assert(ctrl != NULL, "invalid control graph"); |
|
1037 |
assert(!ctrl->is_Start(), "missing lock for unlock"); |
|
1038 |
if (ctrl->is_top()) break; // dead control path |
|
1039 |
if (ctrl->is_Proj()) ctrl = ctrl->in(0); |
|
1040 |
if (ctrl->is_SafePoint()) { |
|
1041 |
break; // found a safepoint (may be the lock we are searching for) |
|
1042 |
} else if (ctrl->is_Region()) { |
|
1043 |
// Check for a simple diamond pattern. Punt on anything more complicated |
|
1044 |
if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) { |
|
1045 |
Node *in1 = next_control(ctrl->in(1)); |
|
1046 |
Node *in2 = next_control(ctrl->in(2)); |
|
1047 |
if (((in1->is_IfTrue() && in2->is_IfFalse()) || |
|
1048 |
(in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) { |
|
1049 |
ctrl = next_control(in1->in(0)->in(0)); |
|
1050 |
} else { |
|
1051 |
break; |
|
1052 |
} |
|
1053 |
} else { |
|
1054 |
break; |
|
1055 |
} |
|
1056 |
} else { |
|
1057 |
ctrl = next_control(ctrl->in(0)); // keep searching |
|
1058 |
} |
|
1059 |
} |
|
1060 |
if (ctrl->is_Lock()) { |
|
1061 |
LockNode *lock = ctrl->as_Lock(); |
|
1062 |
if ((lock->obj_node() == unlock->obj_node()) && |
|
1063 |
(lock->box_node() == unlock->box_node())) { |
|
1064 |
lock_result = lock; |
|
1065 |
} |
|
1066 |
} |
|
1067 |
return lock_result; |
|
1068 |
} |
|
1069 |
||
1070 |
// This code corresponds to case 3 above. |
|
1071 |
||
1072 |
bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock, |
|
1073 |
GrowableArray<AbstractLockNode*> &lock_ops) { |
|
1074 |
Node* if_node = node->in(0); |
|
1075 |
bool if_true = node->is_IfTrue(); |
|
1076 |
||
1077 |
if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) { |
|
1078 |
Node *lock_ctrl = next_control(if_node->in(0)); |
|
1079 |
if (find_matching_unlock(lock_ctrl, lock, lock_ops)) { |
|
1080 |
Node* lock1_node = NULL; |
|
1081 |
ProjNode* proj = if_node->as_If()->proj_out(!if_true); |
|
1082 |
if (if_true) { |
|
1083 |
if (proj->is_IfFalse() && proj->outcnt() == 1) { |
|
1084 |
lock1_node = proj->unique_out(); |
|
1085 |
} |
|
1086 |
} else { |
|
1087 |
if (proj->is_IfTrue() && proj->outcnt() == 1) { |
|
1088 |
lock1_node = proj->unique_out(); |
|
1089 |
} |
|
1090 |
} |
|
1091 |
if (lock1_node != NULL && lock1_node->is_Lock()) { |
|
1092 |
LockNode *lock1 = lock1_node->as_Lock(); |
|
1093 |
if ((lock->obj_node() == lock1->obj_node()) && |
|
1094 |
(lock->box_node() == lock1->box_node()) && !lock1->is_eliminated()) { |
|
1095 |
lock_ops.append(lock1); |
|
1096 |
return true; |
|
1097 |
} |
|
1098 |
} |
|
1099 |
} |
|
1100 |
} |
|
1101 |
||
1102 |
lock_ops.trunc_to(0); |
|
1103 |
return false; |
|
1104 |
} |
|
1105 |
||
1106 |
bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock, |
|
1107 |
GrowableArray<AbstractLockNode*> &lock_ops) { |
|
1108 |
// check each control merging at this point for a matching unlock. |
|
1109 |
// in(0) should be self edge so skip it. |
|
1110 |
for (int i = 1; i < (int)region->req(); i++) { |
|
1111 |
Node *in_node = next_control(region->in(i)); |
|
1112 |
if (in_node != NULL) { |
|
1113 |
if (find_matching_unlock(in_node, lock, lock_ops)) { |
|
1114 |
// found a match so keep on checking. |
|
1115 |
continue; |
|
1116 |
} else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) { |
|
1117 |
continue; |
|
1118 |
} |
|
1119 |
||
1120 |
// If we fall through to here then it was some kind of node we |
|
1121 |
// don't understand or there wasn't a matching unlock, so give |
|
1122 |
// up trying to merge locks. |
|
1123 |
lock_ops.trunc_to(0); |
|
1124 |
return false; |
|
1125 |
} |
|
1126 |
} |
|
1127 |
return true; |
|
1128 |
||
1129 |
} |
|
1130 |
||
1131 |
#ifndef PRODUCT |
|
1132 |
// |
|
1133 |
// Create a counter which counts the number of times this lock is acquired |
|
1134 |
// |
|
1135 |
void AbstractLockNode::create_lock_counter(JVMState* state) { |
|
1136 |
_counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter); |
|
1137 |
} |
|
1138 |
#endif |
|
1139 |
||
1140 |
void AbstractLockNode::set_eliminated() { |
|
1141 |
_eliminate = true; |
|
1142 |
#ifndef PRODUCT |
|
1143 |
if (_counter) { |
|
1144 |
// Update the counter to indicate that this lock was eliminated. |
|
1145 |
// The counter update code will stay around even though the |
|
1146 |
// optimizer will eliminate the lock operation itself. |
|
1147 |
_counter->set_tag(NamedCounter::EliminatedLockCounter); |
|
1148 |
} |
|
1149 |
#endif |
|
1150 |
} |
|
1151 |
||
1152 |
//============================================================================= |
|
1153 |
Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
|
1154 |
||
1155 |
// perform any generic optimizations first |
|
1156 |
Node *result = SafePointNode::Ideal(phase, can_reshape); |
|
1157 |
||
1158 |
// Now see if we can optimize away this lock. We don't actually |
|
1159 |
// remove the locking here, we simply set the _eliminate flag which |
|
1160 |
// prevents macro expansion from expanding the lock. Since we don't |
|
1161 |
// modify the graph, the value returned from this function is the |
|
1162 |
// one computed above. |
|
1163 |
if (EliminateLocks && !is_eliminated()) { |
|
1164 |
// |
|
1165 |
// Try lock coarsening |
|
1166 |
// |
|
1167 |
PhaseIterGVN* iter = phase->is_IterGVN(); |
|
1168 |
if (iter != NULL) { |
|
1169 |
||
1170 |
GrowableArray<AbstractLockNode*> lock_ops; |
|
1171 |
||
1172 |
Node *ctrl = next_control(in(0)); |
|
1173 |
||
1174 |
// now search back for a matching Unlock |
|
1175 |
if (find_matching_unlock(ctrl, this, lock_ops)) { |
|
1176 |
// found an unlock directly preceding this lock. This is the |
|
1177 |
// case of single unlock directly control dependent on a |
|
1178 |
// single lock which is the trivial version of case 1 or 2. |
|
1179 |
} else if (ctrl->is_Region() ) { |
|
1180 |
if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) { |
|
1181 |
// found lock preceded by multiple unlocks along all paths |
|
1182 |
// joining at this point which is case 3 in description above. |
|
1183 |
} |
|
1184 |
} else { |
|
1185 |
// see if this lock comes from either half of an if and the |
|
1186 |
// predecessors merges unlocks and the other half of the if |
|
1187 |
// performs a lock. |
|
1188 |
if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) { |
|
1189 |
// found unlock splitting to an if with locks on both branches. |
|
1190 |
} |
|
1191 |
} |
|
1192 |
||
1193 |
if (lock_ops.length() > 0) { |
|
1194 |
// add ourselves to the list of locks to be eliminated. |
|
1195 |
lock_ops.append(this); |
|
1196 |
||
1197 |
#ifndef PRODUCT |
|
1198 |
if (PrintEliminateLocks) { |
|
1199 |
int locks = 0; |
|
1200 |
int unlocks = 0; |
|
1201 |
for (int i = 0; i < lock_ops.length(); i++) { |
|
1202 |
AbstractLockNode* lock = lock_ops.at(i); |
|
1203 |
if (lock->Opcode() == Op_Lock) locks++; |
|
1204 |
else unlocks++; |
|
1205 |
if (Verbose) { |
|
1206 |
lock->dump(1); |
|
1207 |
} |
|
1208 |
} |
|
1209 |
tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks); |
|
1210 |
} |
|
1211 |
#endif |
|
1212 |
||
1213 |
// for each of the identified locks, mark them |
|
1214 |
// as eliminatable |
|
1215 |
for (int i = 0; i < lock_ops.length(); i++) { |
|
1216 |
AbstractLockNode* lock = lock_ops.at(i); |
|
1217 |
||
1218 |
// Mark it eliminated to update any counters |
|
1219 |
lock->set_eliminated(); |
|
1220 |
} |
|
1221 |
} else if (result != NULL && ctrl->is_Region() && |
|
1222 |
iter->_worklist.member(ctrl)) { |
|
1223 |
// We weren't able to find any opportunities but the region this |
|
1224 |
// lock is control dependent on hasn't been processed yet so put |
|
1225 |
// this lock back on the worklist so we can check again once any |
|
1226 |
// region simplification has occurred. |
|
1227 |
iter->_worklist.push(this); |
|
1228 |
} |
|
1229 |
} |
|
1230 |
} |
|
1231 |
||
1232 |
return result; |
|
1233 |
} |
|
1234 |
||
1235 |
//============================================================================= |
|
1236 |
uint UnlockNode::size_of() const { return sizeof(*this); } |
|
1237 |
||
1238 |
//============================================================================= |
|
1239 |
Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
|
1240 |
||
1241 |
// perform any generic optimizations first |
|
1242 |
Node * result = SafePointNode::Ideal(phase, can_reshape); |
|
1243 |
||
1244 |
// Now see if we can optimize away this unlock. We don't actually |
|
1245 |
// remove the unlocking here, we simply set the _eliminate flag which |
|
1246 |
// prevents macro expansion from expanding the unlock. Since we don't |
|
1247 |
// modify the graph, the value returned from this function is the |
|
1248 |
// one computed above. |
|
1249 |
if (EliminateLocks && !is_eliminated()) { |
|
1250 |
// |
|
1251 |
// If we are unlocking an unescaped object, the lock/unlock is unnecessary |
|
1252 |
// We can eliminate them if there are no safepoints in the locked region. |
|
1253 |
// |
|
1254 |
ConnectionGraph *cgr = Compile::current()->congraph(); |
|
1255 |
if (cgr != NULL && cgr->escape_state(obj_node(), phase) == PointsToNode::NoEscape) { |
|
1256 |
GrowableArray<AbstractLockNode*> lock_ops; |
|
1257 |
LockNode *lock = find_matching_lock(this); |
|
1258 |
if (lock != NULL) { |
|
1259 |
lock_ops.append(this); |
|
1260 |
lock_ops.append(lock); |
|
1261 |
// find other unlocks which pair with the lock we found and add them |
|
1262 |
// to the list |
|
1263 |
Node * box = box_node(); |
|
1264 |
||
1265 |
for (DUIterator_Fast imax, i = box->fast_outs(imax); i < imax; i++) { |
|
1266 |
Node *use = box->fast_out(i); |
|
1267 |
if (use->is_Unlock() && use != this) { |
|
1268 |
UnlockNode *unlock1 = use->as_Unlock(); |
|
1269 |
if (!unlock1->is_eliminated()) { |
|
1270 |
LockNode *lock1 = find_matching_lock(unlock1); |
|
1271 |
if (lock == lock1) |
|
1272 |
lock_ops.append(unlock1); |
|
1273 |
else if (lock1 == NULL) { |
|
1274 |
// we can't find a matching lock, we must assume the worst |
|
1275 |
lock_ops.trunc_to(0); |
|
1276 |
break; |
|
1277 |
} |
|
1278 |
} |
|
1279 |
} |
|
1280 |
} |
|
1281 |
if (lock_ops.length() > 0) { |
|
1282 |
||
1283 |
#ifndef PRODUCT |
|
1284 |
if (PrintEliminateLocks) { |
|
1285 |
int locks = 0; |
|
1286 |
int unlocks = 0; |
|
1287 |
for (int i = 0; i < lock_ops.length(); i++) { |
|
1288 |
AbstractLockNode* lock = lock_ops.at(i); |
|
1289 |
if (lock->Opcode() == Op_Lock) locks++; |
|
1290 |
else unlocks++; |
|
1291 |
if (Verbose) { |
|
1292 |
lock->dump(1); |
|
1293 |
} |
|
1294 |
} |
|
1295 |
tty->print_cr("***Eliminated %d unescaped unlocks and %d unescaped locks", unlocks, locks); |
|
1296 |
} |
|
1297 |
#endif |
|
1298 |
||
1299 |
// for each of the identified locks, mark them |
|
1300 |
// as eliminatable |
|
1301 |
for (int i = 0; i < lock_ops.length(); i++) { |
|
1302 |
AbstractLockNode* lock = lock_ops.at(i); |
|
1303 |
||
1304 |
// Mark it eliminated to update any counters |
|
1305 |
lock->set_eliminated(); |
|
1306 |
} |
|
1307 |
} |
|
1308 |
} |
|
1309 |
} |
|
1310 |
} |
|
1311 |
return result; |
|
1312 |
} |