Tue, 09 Mar 2010 20:16:19 +0100
6919934: JSR 292 needs to support x86 C1
Summary: This implements JSR 292 support for C1 x86.
Reviewed-by: never, jrose, kvn
1 /*
2 * Copyright 2001-2010 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_graphKit.cpp.incl"
28 //----------------------------GraphKit-----------------------------------------
29 // Main utility constructor.
30 GraphKit::GraphKit(JVMState* jvms)
31 : Phase(Phase::Parser),
32 _env(C->env()),
33 _gvn(*C->initial_gvn())
34 {
35 _exceptions = jvms->map()->next_exception();
36 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
37 set_jvms(jvms);
38 }
40 // Private constructor for parser.
41 GraphKit::GraphKit()
42 : Phase(Phase::Parser),
43 _env(C->env()),
44 _gvn(*C->initial_gvn())
45 {
46 _exceptions = NULL;
47 set_map(NULL);
48 debug_only(_sp = -99);
49 debug_only(set_bci(-99));
50 }
54 //---------------------------clean_stack---------------------------------------
55 // Clear away rubbish from the stack area of the JVM state.
56 // This destroys any arguments that may be waiting on the stack.
57 void GraphKit::clean_stack(int from_sp) {
58 SafePointNode* map = this->map();
59 JVMState* jvms = this->jvms();
60 int stk_size = jvms->stk_size();
61 int stkoff = jvms->stkoff();
62 Node* top = this->top();
63 for (int i = from_sp; i < stk_size; i++) {
64 if (map->in(stkoff + i) != top) {
65 map->set_req(stkoff + i, top);
66 }
67 }
68 }
71 //--------------------------------sync_jvms-----------------------------------
72 // Make sure our current jvms agrees with our parse state.
73 JVMState* GraphKit::sync_jvms() const {
74 JVMState* jvms = this->jvms();
75 jvms->set_bci(bci()); // Record the new bci in the JVMState
76 jvms->set_sp(sp()); // Record the new sp in the JVMState
77 assert(jvms_in_sync(), "jvms is now in sync");
78 return jvms;
79 }
81 #ifdef ASSERT
82 bool GraphKit::jvms_in_sync() const {
83 Parse* parse = is_Parse();
84 if (parse == NULL) {
85 if (bci() != jvms()->bci()) return false;
86 if (sp() != (int)jvms()->sp()) return false;
87 return true;
88 }
89 if (jvms()->method() != parse->method()) return false;
90 if (jvms()->bci() != parse->bci()) return false;
91 int jvms_sp = jvms()->sp();
92 if (jvms_sp != parse->sp()) return false;
93 int jvms_depth = jvms()->depth();
94 if (jvms_depth != parse->depth()) return false;
95 return true;
96 }
98 // Local helper checks for special internal merge points
99 // used to accumulate and merge exception states.
100 // They are marked by the region's in(0) edge being the map itself.
101 // Such merge points must never "escape" into the parser at large,
102 // until they have been handed to gvn.transform.
103 static bool is_hidden_merge(Node* reg) {
104 if (reg == NULL) return false;
105 if (reg->is_Phi()) {
106 reg = reg->in(0);
107 if (reg == NULL) return false;
108 }
109 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
110 }
112 void GraphKit::verify_map() const {
113 if (map() == NULL) return; // null map is OK
114 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
115 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
116 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
117 }
119 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
120 assert(ex_map->next_exception() == NULL, "not already part of a chain");
121 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
122 }
123 #endif
125 //---------------------------stop_and_kill_map---------------------------------
126 // Set _map to NULL, signalling a stop to further bytecode execution.
127 // First smash the current map's control to a constant, to mark it dead.
128 void GraphKit::stop_and_kill_map() {
129 SafePointNode* dead_map = stop();
130 if (dead_map != NULL) {
131 dead_map->disconnect_inputs(NULL); // Mark the map as killed.
132 assert(dead_map->is_killed(), "must be so marked");
133 }
134 }
137 //--------------------------------stopped--------------------------------------
138 // Tell if _map is NULL, or control is top.
139 bool GraphKit::stopped() {
140 if (map() == NULL) return true;
141 else if (control() == top()) return true;
142 else return false;
143 }
146 //-----------------------------has_ex_handler----------------------------------
147 // Tell if this method or any caller method has exception handlers.
148 bool GraphKit::has_ex_handler() {
149 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
150 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
151 return true;
152 }
153 }
154 return false;
155 }
157 //------------------------------save_ex_oop------------------------------------
158 // Save an exception without blowing stack contents or other JVM state.
159 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
160 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
161 ex_map->add_req(ex_oop);
162 debug_only(verify_exception_state(ex_map));
163 }
165 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
166 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
167 Node* ex_oop = ex_map->in(ex_map->req()-1);
168 if (clear_it) ex_map->del_req(ex_map->req()-1);
169 return ex_oop;
170 }
172 //-----------------------------saved_ex_oop------------------------------------
173 // Recover a saved exception from its map.
174 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
175 return common_saved_ex_oop(ex_map, false);
176 }
178 //--------------------------clear_saved_ex_oop---------------------------------
179 // Erase a previously saved exception from its map.
180 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
181 return common_saved_ex_oop(ex_map, true);
182 }
184 #ifdef ASSERT
185 //---------------------------has_saved_ex_oop----------------------------------
186 // Erase a previously saved exception from its map.
187 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
188 return ex_map->req() == ex_map->jvms()->endoff()+1;
189 }
190 #endif
192 //-------------------------make_exception_state--------------------------------
193 // Turn the current JVM state into an exception state, appending the ex_oop.
194 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
195 sync_jvms();
196 SafePointNode* ex_map = stop(); // do not manipulate this map any more
197 set_saved_ex_oop(ex_map, ex_oop);
198 return ex_map;
199 }
202 //--------------------------add_exception_state--------------------------------
203 // Add an exception to my list of exceptions.
204 void GraphKit::add_exception_state(SafePointNode* ex_map) {
205 if (ex_map == NULL || ex_map->control() == top()) {
206 return;
207 }
208 #ifdef ASSERT
209 verify_exception_state(ex_map);
210 if (has_exceptions()) {
211 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
212 }
213 #endif
215 // If there is already an exception of exactly this type, merge with it.
216 // In particular, null-checks and other low-level exceptions common up here.
217 Node* ex_oop = saved_ex_oop(ex_map);
218 const Type* ex_type = _gvn.type(ex_oop);
219 if (ex_oop == top()) {
220 // No action needed.
221 return;
222 }
223 assert(ex_type->isa_instptr(), "exception must be an instance");
224 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
225 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
226 // We check sp also because call bytecodes can generate exceptions
227 // both before and after arguments are popped!
228 if (ex_type2 == ex_type
229 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
230 combine_exception_states(ex_map, e2);
231 return;
232 }
233 }
235 // No pre-existing exception of the same type. Chain it on the list.
236 push_exception_state(ex_map);
237 }
239 //-----------------------add_exception_states_from-----------------------------
240 void GraphKit::add_exception_states_from(JVMState* jvms) {
241 SafePointNode* ex_map = jvms->map()->next_exception();
242 if (ex_map != NULL) {
243 jvms->map()->set_next_exception(NULL);
244 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
245 next_map = ex_map->next_exception();
246 ex_map->set_next_exception(NULL);
247 add_exception_state(ex_map);
248 }
249 }
250 }
252 //-----------------------transfer_exceptions_into_jvms-------------------------
253 JVMState* GraphKit::transfer_exceptions_into_jvms() {
254 if (map() == NULL) {
255 // We need a JVMS to carry the exceptions, but the map has gone away.
256 // Create a scratch JVMS, cloned from any of the exception states...
257 if (has_exceptions()) {
258 _map = _exceptions;
259 _map = clone_map();
260 _map->set_next_exception(NULL);
261 clear_saved_ex_oop(_map);
262 debug_only(verify_map());
263 } else {
264 // ...or created from scratch
265 JVMState* jvms = new (C) JVMState(_method, NULL);
266 jvms->set_bci(_bci);
267 jvms->set_sp(_sp);
268 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms));
269 set_jvms(jvms);
270 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
271 set_all_memory(top());
272 while (map()->req() < jvms->endoff()) map()->add_req(top());
273 }
274 // (This is a kludge, in case you didn't notice.)
275 set_control(top());
276 }
277 JVMState* jvms = sync_jvms();
278 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
279 jvms->map()->set_next_exception(_exceptions);
280 _exceptions = NULL; // done with this set of exceptions
281 return jvms;
282 }
284 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
285 assert(is_hidden_merge(dstphi), "must be a special merge node");
286 assert(is_hidden_merge(srcphi), "must be a special merge node");
287 uint limit = srcphi->req();
288 for (uint i = PhiNode::Input; i < limit; i++) {
289 dstphi->add_req(srcphi->in(i));
290 }
291 }
292 static inline void add_one_req(Node* dstphi, Node* src) {
293 assert(is_hidden_merge(dstphi), "must be a special merge node");
294 assert(!is_hidden_merge(src), "must not be a special merge node");
295 dstphi->add_req(src);
296 }
298 //-----------------------combine_exception_states------------------------------
299 // This helper function combines exception states by building phis on a
300 // specially marked state-merging region. These regions and phis are
301 // untransformed, and can build up gradually. The region is marked by
302 // having a control input of its exception map, rather than NULL. Such
303 // regions do not appear except in this function, and in use_exception_state.
304 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
305 if (failing()) return; // dying anyway...
306 JVMState* ex_jvms = ex_map->_jvms;
307 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
308 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
309 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
310 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
311 assert(ex_map->req() == phi_map->req(), "matching maps");
312 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
313 Node* hidden_merge_mark = root();
314 Node* region = phi_map->control();
315 MergeMemNode* phi_mem = phi_map->merged_memory();
316 MergeMemNode* ex_mem = ex_map->merged_memory();
317 if (region->in(0) != hidden_merge_mark) {
318 // The control input is not (yet) a specially-marked region in phi_map.
319 // Make it so, and build some phis.
320 region = new (C, 2) RegionNode(2);
321 _gvn.set_type(region, Type::CONTROL);
322 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
323 region->init_req(1, phi_map->control());
324 phi_map->set_control(region);
325 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
326 record_for_igvn(io_phi);
327 _gvn.set_type(io_phi, Type::ABIO);
328 phi_map->set_i_o(io_phi);
329 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
330 Node* m = mms.memory();
331 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
332 record_for_igvn(m_phi);
333 _gvn.set_type(m_phi, Type::MEMORY);
334 mms.set_memory(m_phi);
335 }
336 }
338 // Either or both of phi_map and ex_map might already be converted into phis.
339 Node* ex_control = ex_map->control();
340 // if there is special marking on ex_map also, we add multiple edges from src
341 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
342 // how wide was the destination phi_map, originally?
343 uint orig_width = region->req();
345 if (add_multiple) {
346 add_n_reqs(region, ex_control);
347 add_n_reqs(phi_map->i_o(), ex_map->i_o());
348 } else {
349 // ex_map has no merges, so we just add single edges everywhere
350 add_one_req(region, ex_control);
351 add_one_req(phi_map->i_o(), ex_map->i_o());
352 }
353 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
354 if (mms.is_empty()) {
355 // get a copy of the base memory, and patch some inputs into it
356 const TypePtr* adr_type = mms.adr_type(C);
357 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
358 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
359 mms.set_memory(phi);
360 // Prepare to append interesting stuff onto the newly sliced phi:
361 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
362 }
363 // Append stuff from ex_map:
364 if (add_multiple) {
365 add_n_reqs(mms.memory(), mms.memory2());
366 } else {
367 add_one_req(mms.memory(), mms.memory2());
368 }
369 }
370 uint limit = ex_map->req();
371 for (uint i = TypeFunc::Parms; i < limit; i++) {
372 // Skip everything in the JVMS after tos. (The ex_oop follows.)
373 if (i == tos) i = ex_jvms->monoff();
374 Node* src = ex_map->in(i);
375 Node* dst = phi_map->in(i);
376 if (src != dst) {
377 PhiNode* phi;
378 if (dst->in(0) != region) {
379 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
380 record_for_igvn(phi);
381 _gvn.set_type(phi, phi->type());
382 phi_map->set_req(i, dst);
383 // Prepare to append interesting stuff onto the new phi:
384 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
385 } else {
386 assert(dst->is_Phi(), "nobody else uses a hidden region");
387 phi = (PhiNode*)dst;
388 }
389 if (add_multiple && src->in(0) == ex_control) {
390 // Both are phis.
391 add_n_reqs(dst, src);
392 } else {
393 while (dst->req() < region->req()) add_one_req(dst, src);
394 }
395 const Type* srctype = _gvn.type(src);
396 if (phi->type() != srctype) {
397 const Type* dsttype = phi->type()->meet(srctype);
398 if (phi->type() != dsttype) {
399 phi->set_type(dsttype);
400 _gvn.set_type(phi, dsttype);
401 }
402 }
403 }
404 }
405 }
407 //--------------------------use_exception_state--------------------------------
408 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
409 if (failing()) { stop(); return top(); }
410 Node* region = phi_map->control();
411 Node* hidden_merge_mark = root();
412 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
413 Node* ex_oop = clear_saved_ex_oop(phi_map);
414 if (region->in(0) == hidden_merge_mark) {
415 // Special marking for internal ex-states. Process the phis now.
416 region->set_req(0, region); // now it's an ordinary region
417 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
418 // Note: Setting the jvms also sets the bci and sp.
419 set_control(_gvn.transform(region));
420 uint tos = jvms()->stkoff() + sp();
421 for (uint i = 1; i < tos; i++) {
422 Node* x = phi_map->in(i);
423 if (x->in(0) == region) {
424 assert(x->is_Phi(), "expected a special phi");
425 phi_map->set_req(i, _gvn.transform(x));
426 }
427 }
428 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
429 Node* x = mms.memory();
430 if (x->in(0) == region) {
431 assert(x->is_Phi(), "nobody else uses a hidden region");
432 mms.set_memory(_gvn.transform(x));
433 }
434 }
435 if (ex_oop->in(0) == region) {
436 assert(ex_oop->is_Phi(), "expected a special phi");
437 ex_oop = _gvn.transform(ex_oop);
438 }
439 } else {
440 set_jvms(phi_map->jvms());
441 }
443 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
444 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
445 return ex_oop;
446 }
448 //---------------------------------java_bc-------------------------------------
449 Bytecodes::Code GraphKit::java_bc() const {
450 ciMethod* method = this->method();
451 int bci = this->bci();
452 if (method != NULL && bci != InvocationEntryBci)
453 return method->java_code_at_bci(bci);
454 else
455 return Bytecodes::_illegal;
456 }
458 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
459 bool must_throw) {
460 // if the exception capability is set, then we will generate code
461 // to check the JavaThread.should_post_on_exceptions flag to see
462 // if we actually need to report exception events (for this
463 // thread). If we don't need to report exception events, we will
464 // take the normal fast path provided by add_exception_events. If
465 // exception event reporting is enabled for this thread, we will
466 // take the uncommon_trap in the BuildCutout below.
468 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
469 Node* jthread = _gvn.transform(new (C, 1) ThreadLocalNode());
470 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
471 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false);
473 // Test the should_post_on_exceptions_flag vs. 0
474 Node* chk = _gvn.transform( new (C, 3) CmpINode(should_post_flag, intcon(0)) );
475 Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, BoolTest::eq) );
477 // Branch to slow_path if should_post_on_exceptions_flag was true
478 { BuildCutout unless(this, tst, PROB_MAX);
479 // Do not try anything fancy if we're notifying the VM on every throw.
480 // Cf. case Bytecodes::_athrow in parse2.cpp.
481 uncommon_trap(reason, Deoptimization::Action_none,
482 (ciKlass*)NULL, (char*)NULL, must_throw);
483 }
485 }
487 //------------------------------builtin_throw----------------------------------
488 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
489 bool must_throw = true;
491 if (env()->jvmti_can_post_on_exceptions()) {
492 // check if we must post exception events, take uncommon trap if so
493 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
494 // here if should_post_on_exceptions is false
495 // continue on with the normal codegen
496 }
498 // If this particular condition has not yet happened at this
499 // bytecode, then use the uncommon trap mechanism, and allow for
500 // a future recompilation if several traps occur here.
501 // If the throw is hot, try to use a more complicated inline mechanism
502 // which keeps execution inside the compiled code.
503 bool treat_throw_as_hot = false;
504 ciMethodData* md = method()->method_data();
506 if (ProfileTraps) {
507 if (too_many_traps(reason)) {
508 treat_throw_as_hot = true;
509 }
510 // (If there is no MDO at all, assume it is early in
511 // execution, and that any deopts are part of the
512 // startup transient, and don't need to be remembered.)
514 // Also, if there is a local exception handler, treat all throws
515 // as hot if there has been at least one in this method.
516 if (C->trap_count(reason) != 0
517 && method()->method_data()->trap_count(reason) != 0
518 && has_ex_handler()) {
519 treat_throw_as_hot = true;
520 }
521 }
523 // If this throw happens frequently, an uncommon trap might cause
524 // a performance pothole. If there is a local exception handler,
525 // and if this particular bytecode appears to be deoptimizing often,
526 // let us handle the throw inline, with a preconstructed instance.
527 // Note: If the deopt count has blown up, the uncommon trap
528 // runtime is going to flush this nmethod, not matter what.
529 if (treat_throw_as_hot
530 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
531 // If the throw is local, we use a pre-existing instance and
532 // punt on the backtrace. This would lead to a missing backtrace
533 // (a repeat of 4292742) if the backtrace object is ever asked
534 // for its backtrace.
535 // Fixing this remaining case of 4292742 requires some flavor of
536 // escape analysis. Leave that for the future.
537 ciInstance* ex_obj = NULL;
538 switch (reason) {
539 case Deoptimization::Reason_null_check:
540 ex_obj = env()->NullPointerException_instance();
541 break;
542 case Deoptimization::Reason_div0_check:
543 ex_obj = env()->ArithmeticException_instance();
544 break;
545 case Deoptimization::Reason_range_check:
546 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
547 break;
548 case Deoptimization::Reason_class_check:
549 if (java_bc() == Bytecodes::_aastore) {
550 ex_obj = env()->ArrayStoreException_instance();
551 } else {
552 ex_obj = env()->ClassCastException_instance();
553 }
554 break;
555 }
556 if (failing()) { stop(); return; } // exception allocation might fail
557 if (ex_obj != NULL) {
558 // Cheat with a preallocated exception object.
559 if (C->log() != NULL)
560 C->log()->elem("hot_throw preallocated='1' reason='%s'",
561 Deoptimization::trap_reason_name(reason));
562 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
563 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
565 // Clear the detail message of the preallocated exception object.
566 // Weblogic sometimes mutates the detail message of exceptions
567 // using reflection.
568 int offset = java_lang_Throwable::get_detailMessage_offset();
569 const TypePtr* adr_typ = ex_con->add_offset(offset);
571 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
572 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT);
574 add_exception_state(make_exception_state(ex_node));
575 return;
576 }
577 }
579 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
580 // It won't be much cheaper than bailing to the interp., since we'll
581 // have to pass up all the debug-info, and the runtime will have to
582 // create the stack trace.
584 // Usual case: Bail to interpreter.
585 // Reserve the right to recompile if we haven't seen anything yet.
587 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
588 if (treat_throw_as_hot
589 && (method()->method_data()->trap_recompiled_at(bci())
590 || C->too_many_traps(reason))) {
591 // We cannot afford to take more traps here. Suffer in the interpreter.
592 if (C->log() != NULL)
593 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
594 Deoptimization::trap_reason_name(reason),
595 C->trap_count(reason));
596 action = Deoptimization::Action_none;
597 }
599 // "must_throw" prunes the JVM state to include only the stack, if there
600 // are no local exception handlers. This should cut down on register
601 // allocation time and code size, by drastically reducing the number
602 // of in-edges on the call to the uncommon trap.
604 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
605 }
608 //----------------------------PreserveJVMState---------------------------------
609 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
610 debug_only(kit->verify_map());
611 _kit = kit;
612 _map = kit->map(); // preserve the map
613 _sp = kit->sp();
614 kit->set_map(clone_map ? kit->clone_map() : NULL);
615 #ifdef ASSERT
616 _bci = kit->bci();
617 Parse* parser = kit->is_Parse();
618 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
619 _block = block;
620 #endif
621 }
622 PreserveJVMState::~PreserveJVMState() {
623 GraphKit* kit = _kit;
624 #ifdef ASSERT
625 assert(kit->bci() == _bci, "bci must not shift");
626 Parse* parser = kit->is_Parse();
627 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
628 assert(block == _block, "block must not shift");
629 #endif
630 kit->set_map(_map);
631 kit->set_sp(_sp);
632 }
635 //-----------------------------BuildCutout-------------------------------------
636 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
637 : PreserveJVMState(kit)
638 {
639 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
640 SafePointNode* outer_map = _map; // preserved map is caller's
641 SafePointNode* inner_map = kit->map();
642 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
643 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
644 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
645 }
646 BuildCutout::~BuildCutout() {
647 GraphKit* kit = _kit;
648 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
649 }
651 //---------------------------PreserveReexecuteState----------------------------
652 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
653 assert(!kit->stopped(), "must call stopped() before");
654 _kit = kit;
655 _sp = kit->sp();
656 _reexecute = kit->jvms()->_reexecute;
657 }
658 PreserveReexecuteState::~PreserveReexecuteState() {
659 if (_kit->stopped()) return;
660 _kit->jvms()->_reexecute = _reexecute;
661 _kit->set_sp(_sp);
662 }
664 //------------------------------clone_map--------------------------------------
665 // Implementation of PreserveJVMState
666 //
667 // Only clone_map(...) here. If this function is only used in the
668 // PreserveJVMState class we may want to get rid of this extra
669 // function eventually and do it all there.
671 SafePointNode* GraphKit::clone_map() {
672 if (map() == NULL) return NULL;
674 // Clone the memory edge first
675 Node* mem = MergeMemNode::make(C, map()->memory());
676 gvn().set_type_bottom(mem);
678 SafePointNode *clonemap = (SafePointNode*)map()->clone();
679 JVMState* jvms = this->jvms();
680 JVMState* clonejvms = jvms->clone_shallow(C);
681 clonemap->set_memory(mem);
682 clonemap->set_jvms(clonejvms);
683 clonejvms->set_map(clonemap);
684 record_for_igvn(clonemap);
685 gvn().set_type_bottom(clonemap);
686 return clonemap;
687 }
690 //-----------------------------set_map_clone-----------------------------------
691 void GraphKit::set_map_clone(SafePointNode* m) {
692 _map = m;
693 _map = clone_map();
694 _map->set_next_exception(NULL);
695 debug_only(verify_map());
696 }
699 //----------------------------kill_dead_locals---------------------------------
700 // Detect any locals which are known to be dead, and force them to top.
701 void GraphKit::kill_dead_locals() {
702 // Consult the liveness information for the locals. If any
703 // of them are unused, then they can be replaced by top(). This
704 // should help register allocation time and cut down on the size
705 // of the deoptimization information.
707 // This call is made from many of the bytecode handling
708 // subroutines called from the Big Switch in do_one_bytecode.
709 // Every bytecode which might include a slow path is responsible
710 // for killing its dead locals. The more consistent we
711 // are about killing deads, the fewer useless phis will be
712 // constructed for them at various merge points.
714 // bci can be -1 (InvocationEntryBci). We return the entry
715 // liveness for the method.
717 if (method() == NULL || method()->code_size() == 0) {
718 // We are building a graph for a call to a native method.
719 // All locals are live.
720 return;
721 }
723 ResourceMark rm;
725 // Consult the liveness information for the locals. If any
726 // of them are unused, then they can be replaced by top(). This
727 // should help register allocation time and cut down on the size
728 // of the deoptimization information.
729 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
731 int len = (int)live_locals.size();
732 assert(len <= jvms()->loc_size(), "too many live locals");
733 for (int local = 0; local < len; local++) {
734 if (!live_locals.at(local)) {
735 set_local(local, top());
736 }
737 }
738 }
740 #ifdef ASSERT
741 //-------------------------dead_locals_are_killed------------------------------
742 // Return true if all dead locals are set to top in the map.
743 // Used to assert "clean" debug info at various points.
744 bool GraphKit::dead_locals_are_killed() {
745 if (method() == NULL || method()->code_size() == 0) {
746 // No locals need to be dead, so all is as it should be.
747 return true;
748 }
750 // Make sure somebody called kill_dead_locals upstream.
751 ResourceMark rm;
752 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
753 if (jvms->loc_size() == 0) continue; // no locals to consult
754 SafePointNode* map = jvms->map();
755 ciMethod* method = jvms->method();
756 int bci = jvms->bci();
757 if (jvms == this->jvms()) {
758 bci = this->bci(); // it might not yet be synched
759 }
760 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
761 int len = (int)live_locals.size();
762 if (!live_locals.is_valid() || len == 0)
763 // This method is trivial, or is poisoned by a breakpoint.
764 return true;
765 assert(len == jvms->loc_size(), "live map consistent with locals map");
766 for (int local = 0; local < len; local++) {
767 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
768 if (PrintMiscellaneous && (Verbose || WizardMode)) {
769 tty->print_cr("Zombie local %d: ", local);
770 jvms->dump();
771 }
772 return false;
773 }
774 }
775 }
776 return true;
777 }
779 #endif //ASSERT
781 // Helper function for enforcing certain bytecodes to reexecute if
782 // deoptimization happens
783 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
784 ciMethod* cur_method = jvms->method();
785 int cur_bci = jvms->bci();
786 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
787 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
788 return Interpreter::bytecode_should_reexecute(code) ||
789 is_anewarray && code == Bytecodes::_multianewarray;
790 // Reexecute _multianewarray bytecode which was replaced with
791 // sequence of [a]newarray. See Parse::do_multianewarray().
792 //
793 // Note: interpreter should not have it set since this optimization
794 // is limited by dimensions and guarded by flag so in some cases
795 // multianewarray() runtime calls will be generated and
796 // the bytecode should not be reexecutes (stack will not be reset).
797 } else
798 return false;
799 }
801 // Helper function for adding JVMState and debug information to node
802 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
803 // Add the safepoint edges to the call (or other safepoint).
805 // Make sure dead locals are set to top. This
806 // should help register allocation time and cut down on the size
807 // of the deoptimization information.
808 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
810 // Walk the inline list to fill in the correct set of JVMState's
811 // Also fill in the associated edges for each JVMState.
813 JVMState* youngest_jvms = sync_jvms();
815 // Do we need debug info here? If it is a SafePoint and this method
816 // cannot de-opt, then we do NOT need any debug info.
817 bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint);
819 // If we are guaranteed to throw, we can prune everything but the
820 // input to the current bytecode.
821 bool can_prune_locals = false;
822 uint stack_slots_not_pruned = 0;
823 int inputs = 0, depth = 0;
824 if (must_throw) {
825 assert(method() == youngest_jvms->method(), "sanity");
826 if (compute_stack_effects(inputs, depth)) {
827 can_prune_locals = true;
828 stack_slots_not_pruned = inputs;
829 }
830 }
832 if (env()->jvmti_can_examine_or_deopt_anywhere()) {
833 // At any safepoint, this method can get breakpointed, which would
834 // then require an immediate deoptimization.
835 full_info = true;
836 can_prune_locals = false; // do not prune locals
837 stack_slots_not_pruned = 0;
838 }
840 // do not scribble on the input jvms
841 JVMState* out_jvms = youngest_jvms->clone_deep(C);
842 call->set_jvms(out_jvms); // Start jvms list for call node
844 // For a known set of bytecodes, the interpreter should reexecute them if
845 // deoptimization happens. We set the reexecute state for them here
846 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
847 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
848 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
849 }
851 // Presize the call:
852 debug_only(uint non_debug_edges = call->req());
853 call->add_req_batch(top(), youngest_jvms->debug_depth());
854 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
856 // Set up edges so that the call looks like this:
857 // Call [state:] ctl io mem fptr retadr
858 // [parms:] parm0 ... parmN
859 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
860 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
861 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
862 // Note that caller debug info precedes callee debug info.
864 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
865 uint debug_ptr = call->req();
867 // Loop over the map input edges associated with jvms, add them
868 // to the call node, & reset all offsets to match call node array.
869 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
870 uint debug_end = debug_ptr;
871 uint debug_start = debug_ptr - in_jvms->debug_size();
872 debug_ptr = debug_start; // back up the ptr
874 uint p = debug_start; // walks forward in [debug_start, debug_end)
875 uint j, k, l;
876 SafePointNode* in_map = in_jvms->map();
877 out_jvms->set_map(call);
879 if (can_prune_locals) {
880 assert(in_jvms->method() == out_jvms->method(), "sanity");
881 // If the current throw can reach an exception handler in this JVMS,
882 // then we must keep everything live that can reach that handler.
883 // As a quick and dirty approximation, we look for any handlers at all.
884 if (in_jvms->method()->has_exception_handlers()) {
885 can_prune_locals = false;
886 }
887 }
889 // Add the Locals
890 k = in_jvms->locoff();
891 l = in_jvms->loc_size();
892 out_jvms->set_locoff(p);
893 if (full_info && !can_prune_locals) {
894 for (j = 0; j < l; j++)
895 call->set_req(p++, in_map->in(k+j));
896 } else {
897 p += l; // already set to top above by add_req_batch
898 }
900 // Add the Expression Stack
901 k = in_jvms->stkoff();
902 l = in_jvms->sp();
903 out_jvms->set_stkoff(p);
904 if (full_info && !can_prune_locals) {
905 for (j = 0; j < l; j++)
906 call->set_req(p++, in_map->in(k+j));
907 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
908 // Divide stack into {S0,...,S1}, where S0 is set to top.
909 uint s1 = stack_slots_not_pruned;
910 stack_slots_not_pruned = 0; // for next iteration
911 if (s1 > l) s1 = l;
912 uint s0 = l - s1;
913 p += s0; // skip the tops preinstalled by add_req_batch
914 for (j = s0; j < l; j++)
915 call->set_req(p++, in_map->in(k+j));
916 } else {
917 p += l; // already set to top above by add_req_batch
918 }
920 // Add the Monitors
921 k = in_jvms->monoff();
922 l = in_jvms->mon_size();
923 out_jvms->set_monoff(p);
924 for (j = 0; j < l; j++)
925 call->set_req(p++, in_map->in(k+j));
927 // Copy any scalar object fields.
928 k = in_jvms->scloff();
929 l = in_jvms->scl_size();
930 out_jvms->set_scloff(p);
931 for (j = 0; j < l; j++)
932 call->set_req(p++, in_map->in(k+j));
934 // Finish the new jvms.
935 out_jvms->set_endoff(p);
937 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
938 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
939 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
940 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
941 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
942 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
944 // Update the two tail pointers in parallel.
945 out_jvms = out_jvms->caller();
946 in_jvms = in_jvms->caller();
947 }
949 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
951 // Test the correctness of JVMState::debug_xxx accessors:
952 assert(call->jvms()->debug_start() == non_debug_edges, "");
953 assert(call->jvms()->debug_end() == call->req(), "");
954 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
955 }
957 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
958 Bytecodes::Code code = java_bc();
959 if (code == Bytecodes::_wide) {
960 code = method()->java_code_at_bci(bci() + 1);
961 }
963 BasicType rtype = T_ILLEGAL;
964 int rsize = 0;
966 if (code != Bytecodes::_illegal) {
967 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
968 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
969 if (rtype < T_CONFLICT)
970 rsize = type2size[rtype];
971 }
973 switch (code) {
974 case Bytecodes::_illegal:
975 return false;
977 case Bytecodes::_ldc:
978 case Bytecodes::_ldc_w:
979 case Bytecodes::_ldc2_w:
980 inputs = 0;
981 break;
983 case Bytecodes::_dup: inputs = 1; break;
984 case Bytecodes::_dup_x1: inputs = 2; break;
985 case Bytecodes::_dup_x2: inputs = 3; break;
986 case Bytecodes::_dup2: inputs = 2; break;
987 case Bytecodes::_dup2_x1: inputs = 3; break;
988 case Bytecodes::_dup2_x2: inputs = 4; break;
989 case Bytecodes::_swap: inputs = 2; break;
990 case Bytecodes::_arraylength: inputs = 1; break;
992 case Bytecodes::_getstatic:
993 case Bytecodes::_putstatic:
994 case Bytecodes::_getfield:
995 case Bytecodes::_putfield:
996 {
997 bool is_get = (depth >= 0), is_static = (depth & 1);
998 bool ignore;
999 ciBytecodeStream iter(method());
1000 iter.reset_to_bci(bci());
1001 iter.next();
1002 ciField* field = iter.get_field(ignore);
1003 int size = field->type()->size();
1004 inputs = (is_static ? 0 : 1);
1005 if (is_get) {
1006 depth = size - inputs;
1007 } else {
1008 inputs += size; // putxxx pops the value from the stack
1009 depth = - inputs;
1010 }
1011 }
1012 break;
1014 case Bytecodes::_invokevirtual:
1015 case Bytecodes::_invokespecial:
1016 case Bytecodes::_invokestatic:
1017 case Bytecodes::_invokedynamic:
1018 case Bytecodes::_invokeinterface:
1019 {
1020 bool ignore;
1021 ciBytecodeStream iter(method());
1022 iter.reset_to_bci(bci());
1023 iter.next();
1024 ciMethod* method = iter.get_method(ignore);
1025 inputs = method->arg_size_no_receiver();
1026 // Add a receiver argument, maybe:
1027 if (code != Bytecodes::_invokestatic &&
1028 code != Bytecodes::_invokedynamic)
1029 inputs += 1;
1030 // (Do not use ciMethod::arg_size(), because
1031 // it might be an unloaded method, which doesn't
1032 // know whether it is static or not.)
1033 int size = method->return_type()->size();
1034 depth = size - inputs;
1035 }
1036 break;
1038 case Bytecodes::_multianewarray:
1039 {
1040 ciBytecodeStream iter(method());
1041 iter.reset_to_bci(bci());
1042 iter.next();
1043 inputs = iter.get_dimensions();
1044 assert(rsize == 1, "");
1045 depth = rsize - inputs;
1046 }
1047 break;
1049 case Bytecodes::_ireturn:
1050 case Bytecodes::_lreturn:
1051 case Bytecodes::_freturn:
1052 case Bytecodes::_dreturn:
1053 case Bytecodes::_areturn:
1054 assert(rsize = -depth, "");
1055 inputs = rsize;
1056 break;
1058 case Bytecodes::_jsr:
1059 case Bytecodes::_jsr_w:
1060 inputs = 0;
1061 depth = 1; // S.B. depth=1, not zero
1062 break;
1064 default:
1065 // bytecode produces a typed result
1066 inputs = rsize - depth;
1067 assert(inputs >= 0, "");
1068 break;
1069 }
1071 #ifdef ASSERT
1072 // spot check
1073 int outputs = depth + inputs;
1074 assert(outputs >= 0, "sanity");
1075 switch (code) {
1076 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1077 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1078 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1079 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1080 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1081 }
1082 #endif //ASSERT
1084 return true;
1085 }
1089 //------------------------------basic_plus_adr---------------------------------
1090 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1091 // short-circuit a common case
1092 if (offset == intcon(0)) return ptr;
1093 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1094 }
1096 Node* GraphKit::ConvI2L(Node* offset) {
1097 // short-circuit a common case
1098 jint offset_con = find_int_con(offset, Type::OffsetBot);
1099 if (offset_con != Type::OffsetBot) {
1100 return longcon((long) offset_con);
1101 }
1102 return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1103 }
1104 Node* GraphKit::ConvL2I(Node* offset) {
1105 // short-circuit a common case
1106 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1107 if (offset_con != (jlong)Type::OffsetBot) {
1108 return intcon((int) offset_con);
1109 }
1110 return _gvn.transform( new (C, 2) ConvL2INode(offset));
1111 }
1113 //-------------------------load_object_klass-----------------------------------
1114 Node* GraphKit::load_object_klass(Node* obj) {
1115 // Special-case a fresh allocation to avoid building nodes:
1116 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1117 if (akls != NULL) return akls;
1118 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1119 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1120 }
1122 //-------------------------load_array_length-----------------------------------
1123 Node* GraphKit::load_array_length(Node* array) {
1124 // Special-case a fresh allocation to avoid building nodes:
1125 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1126 Node *alen;
1127 if (alloc == NULL) {
1128 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1129 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1130 } else {
1131 alen = alloc->Ideal_length();
1132 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1133 if (ccast != alen) {
1134 alen = _gvn.transform(ccast);
1135 }
1136 }
1137 return alen;
1138 }
1140 //------------------------------do_null_check----------------------------------
1141 // Helper function to do a NULL pointer check. Returned value is
1142 // the incoming address with NULL casted away. You are allowed to use the
1143 // not-null value only if you are control dependent on the test.
1144 extern int explicit_null_checks_inserted,
1145 explicit_null_checks_elided;
1146 Node* GraphKit::null_check_common(Node* value, BasicType type,
1147 // optional arguments for variations:
1148 bool assert_null,
1149 Node* *null_control) {
1150 assert(!assert_null || null_control == NULL, "not both at once");
1151 if (stopped()) return top();
1152 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1153 // For some performance testing, we may wish to suppress null checking.
1154 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1155 return value;
1156 }
1157 explicit_null_checks_inserted++;
1159 // Construct NULL check
1160 Node *chk = NULL;
1161 switch(type) {
1162 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1163 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1164 case T_ARRAY : // fall through
1165 type = T_OBJECT; // simplify further tests
1166 case T_OBJECT : {
1167 const Type *t = _gvn.type( value );
1169 const TypeOopPtr* tp = t->isa_oopptr();
1170 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1171 // Only for do_null_check, not any of its siblings:
1172 && !assert_null && null_control == NULL) {
1173 // Usually, any field access or invocation on an unloaded oop type
1174 // will simply fail to link, since the statically linked class is
1175 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1176 // the static class is loaded but the sharper oop type is not.
1177 // Rather than checking for this obscure case in lots of places,
1178 // we simply observe that a null check on an unloaded class
1179 // will always be followed by a nonsense operation, so we
1180 // can just issue the uncommon trap here.
1181 // Our access to the unloaded class will only be correct
1182 // after it has been loaded and initialized, which requires
1183 // a trip through the interpreter.
1184 #ifndef PRODUCT
1185 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1186 #endif
1187 uncommon_trap(Deoptimization::Reason_unloaded,
1188 Deoptimization::Action_reinterpret,
1189 tp->klass(), "!loaded");
1190 return top();
1191 }
1193 if (assert_null) {
1194 // See if the type is contained in NULL_PTR.
1195 // If so, then the value is already null.
1196 if (t->higher_equal(TypePtr::NULL_PTR)) {
1197 explicit_null_checks_elided++;
1198 return value; // Elided null assert quickly!
1199 }
1200 } else {
1201 // See if mixing in the NULL pointer changes type.
1202 // If so, then the NULL pointer was not allowed in the original
1203 // type. In other words, "value" was not-null.
1204 if (t->meet(TypePtr::NULL_PTR) != t) {
1205 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1206 explicit_null_checks_elided++;
1207 return value; // Elided null check quickly!
1208 }
1209 }
1210 chk = new (C, 3) CmpPNode( value, null() );
1211 break;
1212 }
1214 default : ShouldNotReachHere();
1215 }
1216 assert(chk != NULL, "sanity check");
1217 chk = _gvn.transform(chk);
1219 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1220 BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1221 Node *tst = _gvn.transform( btst );
1223 //-----------
1224 // if peephole optimizations occurred, a prior test existed.
1225 // If a prior test existed, maybe it dominates as we can avoid this test.
1226 if (tst != btst && type == T_OBJECT) {
1227 // At this point we want to scan up the CFG to see if we can
1228 // find an identical test (and so avoid this test altogether).
1229 Node *cfg = control();
1230 int depth = 0;
1231 while( depth < 16 ) { // Limit search depth for speed
1232 if( cfg->Opcode() == Op_IfTrue &&
1233 cfg->in(0)->in(1) == tst ) {
1234 // Found prior test. Use "cast_not_null" to construct an identical
1235 // CastPP (and hence hash to) as already exists for the prior test.
1236 // Return that casted value.
1237 if (assert_null) {
1238 replace_in_map(value, null());
1239 return null(); // do not issue the redundant test
1240 }
1241 Node *oldcontrol = control();
1242 set_control(cfg);
1243 Node *res = cast_not_null(value);
1244 set_control(oldcontrol);
1245 explicit_null_checks_elided++;
1246 return res;
1247 }
1248 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1249 if (cfg == NULL) break; // Quit at region nodes
1250 depth++;
1251 }
1252 }
1254 //-----------
1255 // Branch to failure if null
1256 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1257 Deoptimization::DeoptReason reason;
1258 if (assert_null)
1259 reason = Deoptimization::Reason_null_assert;
1260 else if (type == T_OBJECT)
1261 reason = Deoptimization::Reason_null_check;
1262 else
1263 reason = Deoptimization::Reason_div0_check;
1265 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1266 // ciMethodData::has_trap_at will return a conservative -1 if any
1267 // must-be-null assertion has failed. This could cause performance
1268 // problems for a method after its first do_null_assert failure.
1269 // Consider using 'Reason_class_check' instead?
1271 // To cause an implicit null check, we set the not-null probability
1272 // to the maximum (PROB_MAX). For an explicit check the probability
1273 // is set to a smaller value.
1274 if (null_control != NULL || too_many_traps(reason)) {
1275 // probability is less likely
1276 ok_prob = PROB_LIKELY_MAG(3);
1277 } else if (!assert_null &&
1278 (ImplicitNullCheckThreshold > 0) &&
1279 method() != NULL &&
1280 (method()->method_data()->trap_count(reason)
1281 >= (uint)ImplicitNullCheckThreshold)) {
1282 ok_prob = PROB_LIKELY_MAG(3);
1283 }
1285 if (null_control != NULL) {
1286 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1287 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1288 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
1289 if (null_true == top())
1290 explicit_null_checks_elided++;
1291 (*null_control) = null_true;
1292 } else {
1293 BuildCutout unless(this, tst, ok_prob);
1294 // Check for optimizer eliding test at parse time
1295 if (stopped()) {
1296 // Failure not possible; do not bother making uncommon trap.
1297 explicit_null_checks_elided++;
1298 } else if (assert_null) {
1299 uncommon_trap(reason,
1300 Deoptimization::Action_make_not_entrant,
1301 NULL, "assert_null");
1302 } else {
1303 replace_in_map(value, zerocon(type));
1304 builtin_throw(reason);
1305 }
1306 }
1308 // Must throw exception, fall-thru not possible?
1309 if (stopped()) {
1310 return top(); // No result
1311 }
1313 if (assert_null) {
1314 // Cast obj to null on this path.
1315 replace_in_map(value, zerocon(type));
1316 return zerocon(type);
1317 }
1319 // Cast obj to not-null on this path, if there is no null_control.
1320 // (If there is a null_control, a non-null value may come back to haunt us.)
1321 if (type == T_OBJECT) {
1322 Node* cast = cast_not_null(value, false);
1323 if (null_control == NULL || (*null_control) == top())
1324 replace_in_map(value, cast);
1325 value = cast;
1326 }
1328 return value;
1329 }
1332 //------------------------------cast_not_null----------------------------------
1333 // Cast obj to not-null on this path
1334 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1335 const Type *t = _gvn.type(obj);
1336 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1337 // Object is already not-null?
1338 if( t == t_not_null ) return obj;
1340 Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1341 cast->init_req(0, control());
1342 cast = _gvn.transform( cast );
1344 // Scan for instances of 'obj' in the current JVM mapping.
1345 // These instances are known to be not-null after the test.
1346 if (do_replace_in_map)
1347 replace_in_map(obj, cast);
1349 return cast; // Return casted value
1350 }
1353 //--------------------------replace_in_map-------------------------------------
1354 void GraphKit::replace_in_map(Node* old, Node* neww) {
1355 this->map()->replace_edge(old, neww);
1357 // Note: This operation potentially replaces any edge
1358 // on the map. This includes locals, stack, and monitors
1359 // of the current (innermost) JVM state.
1361 // We can consider replacing in caller maps.
1362 // The idea would be that an inlined function's null checks
1363 // can be shared with the entire inlining tree.
1364 // The expense of doing this is that the PreserveJVMState class
1365 // would have to preserve caller states too, with a deep copy.
1366 }
1370 //=============================================================================
1371 //--------------------------------memory---------------------------------------
1372 Node* GraphKit::memory(uint alias_idx) {
1373 MergeMemNode* mem = merged_memory();
1374 Node* p = mem->memory_at(alias_idx);
1375 _gvn.set_type(p, Type::MEMORY); // must be mapped
1376 return p;
1377 }
1379 //-----------------------------reset_memory------------------------------------
1380 Node* GraphKit::reset_memory() {
1381 Node* mem = map()->memory();
1382 // do not use this node for any more parsing!
1383 debug_only( map()->set_memory((Node*)NULL) );
1384 return _gvn.transform( mem );
1385 }
1387 //------------------------------set_all_memory---------------------------------
1388 void GraphKit::set_all_memory(Node* newmem) {
1389 Node* mergemem = MergeMemNode::make(C, newmem);
1390 gvn().set_type_bottom(mergemem);
1391 map()->set_memory(mergemem);
1392 }
1394 //------------------------------set_all_memory_call----------------------------
1395 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1396 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1397 set_all_memory(newmem);
1398 }
1400 //=============================================================================
1401 //
1402 // parser factory methods for MemNodes
1403 //
1404 // These are layered on top of the factory methods in LoadNode and StoreNode,
1405 // and integrate with the parser's memory state and _gvn engine.
1406 //
1408 // factory methods in "int adr_idx"
1409 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1410 int adr_idx,
1411 bool require_atomic_access) {
1412 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1413 const TypePtr* adr_type = NULL; // debug-mode-only argument
1414 debug_only(adr_type = C->get_adr_type(adr_idx));
1415 Node* mem = memory(adr_idx);
1416 Node* ld;
1417 if (require_atomic_access && bt == T_LONG) {
1418 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1419 } else {
1420 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1421 }
1422 return _gvn.transform(ld);
1423 }
1425 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1426 int adr_idx,
1427 bool require_atomic_access) {
1428 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1429 const TypePtr* adr_type = NULL;
1430 debug_only(adr_type = C->get_adr_type(adr_idx));
1431 Node *mem = memory(adr_idx);
1432 Node* st;
1433 if (require_atomic_access && bt == T_LONG) {
1434 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1435 } else {
1436 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1437 }
1438 st = _gvn.transform(st);
1439 set_memory(st, adr_idx);
1440 // Back-to-back stores can only remove intermediate store with DU info
1441 // so push on worklist for optimizer.
1442 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1443 record_for_igvn(st);
1445 return st;
1446 }
1449 void GraphKit::pre_barrier(Node* ctl,
1450 Node* obj,
1451 Node* adr,
1452 uint adr_idx,
1453 Node* val,
1454 const TypeOopPtr* val_type,
1455 BasicType bt) {
1456 BarrierSet* bs = Universe::heap()->barrier_set();
1457 set_control(ctl);
1458 switch (bs->kind()) {
1459 case BarrierSet::G1SATBCT:
1460 case BarrierSet::G1SATBCTLogging:
1461 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt);
1462 break;
1464 case BarrierSet::CardTableModRef:
1465 case BarrierSet::CardTableExtension:
1466 case BarrierSet::ModRef:
1467 break;
1469 case BarrierSet::Other:
1470 default :
1471 ShouldNotReachHere();
1473 }
1474 }
1476 void GraphKit::post_barrier(Node* ctl,
1477 Node* store,
1478 Node* obj,
1479 Node* adr,
1480 uint adr_idx,
1481 Node* val,
1482 BasicType bt,
1483 bool use_precise) {
1484 BarrierSet* bs = Universe::heap()->barrier_set();
1485 set_control(ctl);
1486 switch (bs->kind()) {
1487 case BarrierSet::G1SATBCT:
1488 case BarrierSet::G1SATBCTLogging:
1489 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1490 break;
1492 case BarrierSet::CardTableModRef:
1493 case BarrierSet::CardTableExtension:
1494 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1495 break;
1497 case BarrierSet::ModRef:
1498 break;
1500 case BarrierSet::Other:
1501 default :
1502 ShouldNotReachHere();
1504 }
1505 }
1507 Node* GraphKit::store_oop(Node* ctl,
1508 Node* obj,
1509 Node* adr,
1510 const TypePtr* adr_type,
1511 Node* val,
1512 const TypeOopPtr* val_type,
1513 BasicType bt,
1514 bool use_precise) {
1516 set_control(ctl);
1517 if (stopped()) return top(); // Dead path ?
1519 assert(bt == T_OBJECT, "sanity");
1520 assert(val != NULL, "not dead path");
1521 uint adr_idx = C->get_alias_index(adr_type);
1522 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1524 pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt);
1525 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1526 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1527 return store;
1528 }
1530 // Could be an array or object we don't know at compile time (unsafe ref.)
1531 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1532 Node* obj, // containing obj
1533 Node* adr, // actual adress to store val at
1534 const TypePtr* adr_type,
1535 Node* val,
1536 BasicType bt) {
1537 Compile::AliasType* at = C->alias_type(adr_type);
1538 const TypeOopPtr* val_type = NULL;
1539 if (adr_type->isa_instptr()) {
1540 if (at->field() != NULL) {
1541 // known field. This code is a copy of the do_put_xxx logic.
1542 ciField* field = at->field();
1543 if (!field->type()->is_loaded()) {
1544 val_type = TypeInstPtr::BOTTOM;
1545 } else {
1546 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1547 }
1548 }
1549 } else if (adr_type->isa_aryptr()) {
1550 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1551 }
1552 if (val_type == NULL) {
1553 val_type = TypeInstPtr::BOTTOM;
1554 }
1555 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1556 }
1559 //-------------------------array_element_address-------------------------
1560 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1561 const TypeInt* sizetype) {
1562 uint shift = exact_log2(type2aelembytes(elembt));
1563 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1565 // short-circuit a common case (saves lots of confusing waste motion)
1566 jint idx_con = find_int_con(idx, -1);
1567 if (idx_con >= 0) {
1568 intptr_t offset = header + ((intptr_t)idx_con << shift);
1569 return basic_plus_adr(ary, offset);
1570 }
1572 // must be correct type for alignment purposes
1573 Node* base = basic_plus_adr(ary, header);
1574 #ifdef _LP64
1575 // The scaled index operand to AddP must be a clean 64-bit value.
1576 // Java allows a 32-bit int to be incremented to a negative
1577 // value, which appears in a 64-bit register as a large
1578 // positive number. Using that large positive number as an
1579 // operand in pointer arithmetic has bad consequences.
1580 // On the other hand, 32-bit overflow is rare, and the possibility
1581 // can often be excluded, if we annotate the ConvI2L node with
1582 // a type assertion that its value is known to be a small positive
1583 // number. (The prior range check has ensured this.)
1584 // This assertion is used by ConvI2LNode::Ideal.
1585 int index_max = max_jint - 1; // array size is max_jint, index is one less
1586 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1587 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1588 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1589 #endif
1590 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1591 return basic_plus_adr(ary, base, scale);
1592 }
1594 //-------------------------load_array_element-------------------------
1595 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1596 const Type* elemtype = arytype->elem();
1597 BasicType elembt = elemtype->array_element_basic_type();
1598 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1599 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1600 return ld;
1601 }
1603 //-------------------------set_arguments_for_java_call-------------------------
1604 // Arguments (pre-popped from the stack) are taken from the JVMS.
1605 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1606 // Add the call arguments:
1607 uint nargs = call->method()->arg_size();
1608 for (uint i = 0; i < nargs; i++) {
1609 Node* arg = argument(i);
1610 call->init_req(i + TypeFunc::Parms, arg);
1611 }
1612 }
1614 //---------------------------set_edges_for_java_call---------------------------
1615 // Connect a newly created call into the current JVMS.
1616 // A return value node (if any) is returned from set_edges_for_java_call.
1617 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1619 // Add the predefined inputs:
1620 call->init_req( TypeFunc::Control, control() );
1621 call->init_req( TypeFunc::I_O , i_o() );
1622 call->init_req( TypeFunc::Memory , reset_memory() );
1623 call->init_req( TypeFunc::FramePtr, frameptr() );
1624 call->init_req( TypeFunc::ReturnAdr, top() );
1626 add_safepoint_edges(call, must_throw);
1628 Node* xcall = _gvn.transform(call);
1630 if (xcall == top()) {
1631 set_control(top());
1632 return;
1633 }
1634 assert(xcall == call, "call identity is stable");
1636 // Re-use the current map to produce the result.
1638 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1639 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1640 set_all_memory_call(xcall, separate_io_proj);
1642 //return xcall; // no need, caller already has it
1643 }
1645 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1646 if (stopped()) return top(); // maybe the call folded up?
1648 // Capture the return value, if any.
1649 Node* ret;
1650 if (call->method() == NULL ||
1651 call->method()->return_type()->basic_type() == T_VOID)
1652 ret = top();
1653 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1655 // Note: Since any out-of-line call can produce an exception,
1656 // we always insert an I_O projection from the call into the result.
1658 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1660 if (separate_io_proj) {
1661 // The caller requested separate projections be used by the fall
1662 // through and exceptional paths, so replace the projections for
1663 // the fall through path.
1664 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) ));
1665 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ));
1666 }
1667 return ret;
1668 }
1670 //--------------------set_predefined_input_for_runtime_call--------------------
1671 // Reading and setting the memory state is way conservative here.
1672 // The real problem is that I am not doing real Type analysis on memory,
1673 // so I cannot distinguish card mark stores from other stores. Across a GC
1674 // point the Store Barrier and the card mark memory has to agree. I cannot
1675 // have a card mark store and its barrier split across the GC point from
1676 // either above or below. Here I get that to happen by reading ALL of memory.
1677 // A better answer would be to separate out card marks from other memory.
1678 // For now, return the input memory state, so that it can be reused
1679 // after the call, if this call has restricted memory effects.
1680 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1681 // Set fixed predefined input arguments
1682 Node* memory = reset_memory();
1683 call->init_req( TypeFunc::Control, control() );
1684 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1685 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1686 call->init_req( TypeFunc::FramePtr, frameptr() );
1687 call->init_req( TypeFunc::ReturnAdr, top() );
1688 return memory;
1689 }
1691 //-------------------set_predefined_output_for_runtime_call--------------------
1692 // Set control and memory (not i_o) from the call.
1693 // If keep_mem is not NULL, use it for the output state,
1694 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1695 // If hook_mem is NULL, this call produces no memory effects at all.
1696 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1697 // then only that memory slice is taken from the call.
1698 // In the last case, we must put an appropriate memory barrier before
1699 // the call, so as to create the correct anti-dependencies on loads
1700 // preceding the call.
1701 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1702 Node* keep_mem,
1703 const TypePtr* hook_mem) {
1704 // no i/o
1705 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1706 if (keep_mem) {
1707 // First clone the existing memory state
1708 set_all_memory(keep_mem);
1709 if (hook_mem != NULL) {
1710 // Make memory for the call
1711 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1712 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1713 // We also use hook_mem to extract specific effects from arraycopy stubs.
1714 set_memory(mem, hook_mem);
1715 }
1716 // ...else the call has NO memory effects.
1718 // Make sure the call advertises its memory effects precisely.
1719 // This lets us build accurate anti-dependences in gcm.cpp.
1720 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1721 "call node must be constructed correctly");
1722 } else {
1723 assert(hook_mem == NULL, "");
1724 // This is not a "slow path" call; all memory comes from the call.
1725 set_all_memory_call(call);
1726 }
1727 }
1730 // Replace the call with the current state of the kit.
1731 void GraphKit::replace_call(CallNode* call, Node* result) {
1732 JVMState* ejvms = NULL;
1733 if (has_exceptions()) {
1734 ejvms = transfer_exceptions_into_jvms();
1735 }
1737 SafePointNode* final_state = stop();
1739 // Find all the needed outputs of this call
1740 CallProjections callprojs;
1741 call->extract_projections(&callprojs, true);
1743 // Replace all the old call edges with the edges from the inlining result
1744 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control));
1745 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory));
1746 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O));
1748 // Replace the result with the new result if it exists and is used
1749 if (callprojs.resproj != NULL && result != NULL) {
1750 C->gvn_replace_by(callprojs.resproj, result);
1751 }
1753 if (ejvms == NULL) {
1754 // No exception edges to simply kill off those paths
1755 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1756 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1757 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1759 // Replace the old exception object with top
1760 if (callprojs.exobj != NULL) {
1761 C->gvn_replace_by(callprojs.exobj, C->top());
1762 }
1763 } else {
1764 GraphKit ekit(ejvms);
1766 // Load my combined exception state into the kit, with all phis transformed:
1767 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1769 Node* ex_oop = ekit.use_exception_state(ex_map);
1771 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1772 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1773 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1775 // Replace the old exception object with the newly created one
1776 if (callprojs.exobj != NULL) {
1777 C->gvn_replace_by(callprojs.exobj, ex_oop);
1778 }
1779 }
1781 // Disconnect the call from the graph
1782 call->disconnect_inputs(NULL);
1783 C->gvn_replace_by(call, C->top());
1784 }
1787 //------------------------------increment_counter------------------------------
1788 // for statistics: increment a VM counter by 1
1790 void GraphKit::increment_counter(address counter_addr) {
1791 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1792 increment_counter(adr1);
1793 }
1795 void GraphKit::increment_counter(Node* counter_addr) {
1796 int adr_type = Compile::AliasIdxRaw;
1797 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
1798 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1799 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type );
1800 }
1803 //------------------------------uncommon_trap----------------------------------
1804 // Bail out to the interpreter in mid-method. Implemented by calling the
1805 // uncommon_trap blob. This helper function inserts a runtime call with the
1806 // right debug info.
1807 void GraphKit::uncommon_trap(int trap_request,
1808 ciKlass* klass, const char* comment,
1809 bool must_throw,
1810 bool keep_exact_action) {
1811 if (failing()) stop();
1812 if (stopped()) return; // trap reachable?
1814 // Note: If ProfileTraps is true, and if a deopt. actually
1815 // occurs here, the runtime will make sure an MDO exists. There is
1816 // no need to call method()->build_method_data() at this point.
1818 #ifdef ASSERT
1819 if (!must_throw) {
1820 // Make sure the stack has at least enough depth to execute
1821 // the current bytecode.
1822 int inputs, ignore;
1823 if (compute_stack_effects(inputs, ignore)) {
1824 assert(sp() >= inputs, "must have enough JVMS stack to execute");
1825 // It is a frequent error in library_call.cpp to issue an
1826 // uncommon trap with the _sp value already popped.
1827 }
1828 }
1829 #endif
1831 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1832 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1834 switch (action) {
1835 case Deoptimization::Action_maybe_recompile:
1836 case Deoptimization::Action_reinterpret:
1837 // Temporary fix for 6529811 to allow virtual calls to be sure they
1838 // get the chance to go from mono->bi->mega
1839 if (!keep_exact_action &&
1840 Deoptimization::trap_request_index(trap_request) < 0 &&
1841 too_many_recompiles(reason)) {
1842 // This BCI is causing too many recompilations.
1843 action = Deoptimization::Action_none;
1844 trap_request = Deoptimization::make_trap_request(reason, action);
1845 } else {
1846 C->set_trap_can_recompile(true);
1847 }
1848 break;
1849 case Deoptimization::Action_make_not_entrant:
1850 C->set_trap_can_recompile(true);
1851 break;
1852 #ifdef ASSERT
1853 case Deoptimization::Action_none:
1854 case Deoptimization::Action_make_not_compilable:
1855 break;
1856 default:
1857 assert(false, "bad action");
1858 #endif
1859 }
1861 if (TraceOptoParse) {
1862 char buf[100];
1863 tty->print_cr("Uncommon trap %s at bci:%d",
1864 Deoptimization::format_trap_request(buf, sizeof(buf),
1865 trap_request), bci());
1866 }
1868 CompileLog* log = C->log();
1869 if (log != NULL) {
1870 int kid = (klass == NULL)? -1: log->identify(klass);
1871 log->begin_elem("uncommon_trap bci='%d'", bci());
1872 char buf[100];
1873 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1874 trap_request));
1875 if (kid >= 0) log->print(" klass='%d'", kid);
1876 if (comment != NULL) log->print(" comment='%s'", comment);
1877 log->end_elem();
1878 }
1880 // Make sure any guarding test views this path as very unlikely
1881 Node *i0 = control()->in(0);
1882 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1883 IfNode *iff = i0->as_If();
1884 float f = iff->_prob; // Get prob
1885 if (control()->Opcode() == Op_IfTrue) {
1886 if (f > PROB_UNLIKELY_MAG(4))
1887 iff->_prob = PROB_MIN;
1888 } else {
1889 if (f < PROB_LIKELY_MAG(4))
1890 iff->_prob = PROB_MAX;
1891 }
1892 }
1894 // Clear out dead values from the debug info.
1895 kill_dead_locals();
1897 // Now insert the uncommon trap subroutine call
1898 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin();
1899 const TypePtr* no_memory_effects = NULL;
1900 // Pass the index of the class to be loaded
1901 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1902 (must_throw ? RC_MUST_THROW : 0),
1903 OptoRuntime::uncommon_trap_Type(),
1904 call_addr, "uncommon_trap", no_memory_effects,
1905 intcon(trap_request));
1906 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1907 "must extract request correctly from the graph");
1908 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1910 call->set_req(TypeFunc::ReturnAdr, returnadr());
1911 // The debug info is the only real input to this call.
1913 // Halt-and-catch fire here. The above call should never return!
1914 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1915 _gvn.set_type_bottom(halt);
1916 root()->add_req(halt);
1918 stop_and_kill_map();
1919 }
1922 //--------------------------just_allocated_object------------------------------
1923 // Report the object that was just allocated.
1924 // It must be the case that there are no intervening safepoints.
1925 // We use this to determine if an object is so "fresh" that
1926 // it does not require card marks.
1927 Node* GraphKit::just_allocated_object(Node* current_control) {
1928 if (C->recent_alloc_ctl() == current_control)
1929 return C->recent_alloc_obj();
1930 return NULL;
1931 }
1934 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1935 // (Note: TypeFunc::make has a cache that makes this fast.)
1936 const TypeFunc* tf = TypeFunc::make(dest_method);
1937 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
1938 for (int j = 0; j < nargs; j++) {
1939 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1940 if( targ->basic_type() == T_DOUBLE ) {
1941 // If any parameters are doubles, they must be rounded before
1942 // the call, dstore_rounding does gvn.transform
1943 Node *arg = argument(j);
1944 arg = dstore_rounding(arg);
1945 set_argument(j, arg);
1946 }
1947 }
1948 }
1950 void GraphKit::round_double_result(ciMethod* dest_method) {
1951 // A non-strict method may return a double value which has an extended
1952 // exponent, but this must not be visible in a caller which is 'strict'
1953 // If a strict caller invokes a non-strict callee, round a double result
1955 BasicType result_type = dest_method->return_type()->basic_type();
1956 assert( method() != NULL, "must have caller context");
1957 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
1958 // Destination method's return value is on top of stack
1959 // dstore_rounding() does gvn.transform
1960 Node *result = pop_pair();
1961 result = dstore_rounding(result);
1962 push_pair(result);
1963 }
1964 }
1966 // rounding for strict float precision conformance
1967 Node* GraphKit::precision_rounding(Node* n) {
1968 return UseStrictFP && _method->flags().is_strict()
1969 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
1970 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
1971 : n;
1972 }
1974 // rounding for strict double precision conformance
1975 Node* GraphKit::dprecision_rounding(Node *n) {
1976 return UseStrictFP && _method->flags().is_strict()
1977 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
1978 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1979 : n;
1980 }
1982 // rounding for non-strict double stores
1983 Node* GraphKit::dstore_rounding(Node* n) {
1984 return Matcher::strict_fp_requires_explicit_rounding
1985 && UseSSE <= 1
1986 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1987 : n;
1988 }
1990 //=============================================================================
1991 // Generate a fast path/slow path idiom. Graph looks like:
1992 // [foo] indicates that 'foo' is a parameter
1993 //
1994 // [in] NULL
1995 // \ /
1996 // CmpP
1997 // Bool ne
1998 // If
1999 // / \
2000 // True False-<2>
2001 // / |
2002 // / cast_not_null
2003 // Load | | ^
2004 // [fast_test] | |
2005 // gvn to opt_test | |
2006 // / \ | <1>
2007 // True False |
2008 // | \\ |
2009 // [slow_call] \[fast_result]
2010 // Ctl Val \ \
2011 // | \ \
2012 // Catch <1> \ \
2013 // / \ ^ \ \
2014 // Ex No_Ex | \ \
2015 // | \ \ | \ <2> \
2016 // ... \ [slow_res] | | \ [null_result]
2017 // \ \--+--+--- | |
2018 // \ | / \ | /
2019 // --------Region Phi
2020 //
2021 //=============================================================================
2022 // Code is structured as a series of driver functions all called 'do_XXX' that
2023 // call a set of helper functions. Helper functions first, then drivers.
2025 //------------------------------null_check_oop---------------------------------
2026 // Null check oop. Set null-path control into Region in slot 3.
2027 // Make a cast-not-nullness use the other not-null control. Return cast.
2028 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2029 bool never_see_null) {
2030 // Initial NULL check taken path
2031 (*null_control) = top();
2032 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2034 // Generate uncommon_trap:
2035 if (never_see_null && (*null_control) != top()) {
2036 // If we see an unexpected null at a check-cast we record it and force a
2037 // recompile; the offending check-cast will be compiled to handle NULLs.
2038 // If we see more than one offending BCI, then all checkcasts in the
2039 // method will be compiled to handle NULLs.
2040 PreserveJVMState pjvms(this);
2041 set_control(*null_control);
2042 replace_in_map(value, null());
2043 uncommon_trap(Deoptimization::Reason_null_check,
2044 Deoptimization::Action_make_not_entrant);
2045 (*null_control) = top(); // NULL path is dead
2046 }
2048 // Cast away null-ness on the result
2049 return cast;
2050 }
2052 //------------------------------opt_iff----------------------------------------
2053 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2054 // Return slow-path control.
2055 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2056 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2058 // Fast path taken; set region slot 2
2059 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2060 region->init_req(2,fast_taken); // Capture fast-control
2062 // Fast path not-taken, i.e. slow path
2063 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2064 return slow_taken;
2065 }
2067 //-----------------------------make_runtime_call-------------------------------
2068 Node* GraphKit::make_runtime_call(int flags,
2069 const TypeFunc* call_type, address call_addr,
2070 const char* call_name,
2071 const TypePtr* adr_type,
2072 // The following parms are all optional.
2073 // The first NULL ends the list.
2074 Node* parm0, Node* parm1,
2075 Node* parm2, Node* parm3,
2076 Node* parm4, Node* parm5,
2077 Node* parm6, Node* parm7) {
2078 // Slow-path call
2079 int size = call_type->domain()->cnt();
2080 bool is_leaf = !(flags & RC_NO_LEAF);
2081 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2082 if (call_name == NULL) {
2083 assert(!is_leaf, "must supply name for leaf");
2084 call_name = OptoRuntime::stub_name(call_addr);
2085 }
2086 CallNode* call;
2087 if (!is_leaf) {
2088 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2089 bci(), adr_type);
2090 } else if (flags & RC_NO_FP) {
2091 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2092 } else {
2093 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2094 }
2096 // The following is similar to set_edges_for_java_call,
2097 // except that the memory effects of the call are restricted to AliasIdxRaw.
2099 // Slow path call has no side-effects, uses few values
2100 bool wide_in = !(flags & RC_NARROW_MEM);
2101 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2103 Node* prev_mem = NULL;
2104 if (wide_in) {
2105 prev_mem = set_predefined_input_for_runtime_call(call);
2106 } else {
2107 assert(!wide_out, "narrow in => narrow out");
2108 Node* narrow_mem = memory(adr_type);
2109 prev_mem = reset_memory();
2110 map()->set_memory(narrow_mem);
2111 set_predefined_input_for_runtime_call(call);
2112 }
2114 // Hook each parm in order. Stop looking at the first NULL.
2115 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2116 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2117 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2118 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2119 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2120 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2121 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2122 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2123 /* close each nested if ===> */ } } } } } } } }
2124 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2126 if (!is_leaf) {
2127 // Non-leaves can block and take safepoints:
2128 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2129 }
2130 // Non-leaves can throw exceptions:
2131 if (has_io) {
2132 call->set_req(TypeFunc::I_O, i_o());
2133 }
2135 if (flags & RC_UNCOMMON) {
2136 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2137 // (An "if" probability corresponds roughly to an unconditional count.
2138 // Sort of.)
2139 call->set_cnt(PROB_UNLIKELY_MAG(4));
2140 }
2142 Node* c = _gvn.transform(call);
2143 assert(c == call, "cannot disappear");
2145 if (wide_out) {
2146 // Slow path call has full side-effects.
2147 set_predefined_output_for_runtime_call(call);
2148 } else {
2149 // Slow path call has few side-effects, and/or sets few values.
2150 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2151 }
2153 if (has_io) {
2154 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2155 }
2156 return call;
2158 }
2160 //------------------------------merge_memory-----------------------------------
2161 // Merge memory from one path into the current memory state.
2162 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2163 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2164 Node* old_slice = mms.force_memory();
2165 Node* new_slice = mms.memory2();
2166 if (old_slice != new_slice) {
2167 PhiNode* phi;
2168 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2169 phi = new_slice->as_Phi();
2170 #ifdef ASSERT
2171 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2172 old_slice = old_slice->in(new_path);
2173 // Caller is responsible for ensuring that any pre-existing
2174 // phis are already aware of old memory.
2175 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2176 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2177 #endif
2178 mms.set_memory(phi);
2179 } else {
2180 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2181 _gvn.set_type(phi, Type::MEMORY);
2182 phi->set_req(new_path, new_slice);
2183 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2184 }
2185 }
2186 }
2187 }
2189 //------------------------------make_slow_call_ex------------------------------
2190 // Make the exception handler hookups for the slow call
2191 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2192 if (stopped()) return;
2194 // Make a catch node with just two handlers: fall-through and catch-all
2195 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2196 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2197 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2198 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2200 { PreserveJVMState pjvms(this);
2201 set_control(excp);
2202 set_i_o(i_o);
2204 if (excp != top()) {
2205 // Create an exception state also.
2206 // Use an exact type if the caller has specified a specific exception.
2207 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2208 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
2209 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2210 }
2211 }
2213 // Get the no-exception control from the CatchNode.
2214 set_control(norm);
2215 }
2218 //-------------------------------gen_subtype_check-----------------------------
2219 // Generate a subtyping check. Takes as input the subtype and supertype.
2220 // Returns 2 values: sets the default control() to the true path and returns
2221 // the false path. Only reads invariant memory; sets no (visible) memory.
2222 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2223 // but that's not exposed to the optimizer. This call also doesn't take in an
2224 // Object; if you wish to check an Object you need to load the Object's class
2225 // prior to coming here.
2226 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2227 // Fast check for identical types, perhaps identical constants.
2228 // The types can even be identical non-constants, in cases
2229 // involving Array.newInstance, Object.clone, etc.
2230 if (subklass == superklass)
2231 return top(); // false path is dead; no test needed.
2233 if (_gvn.type(superklass)->singleton()) {
2234 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2235 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2237 // In the common case of an exact superklass, try to fold up the
2238 // test before generating code. You may ask, why not just generate
2239 // the code and then let it fold up? The answer is that the generated
2240 // code will necessarily include null checks, which do not always
2241 // completely fold away. If they are also needless, then they turn
2242 // into a performance loss. Example:
2243 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2244 // Here, the type of 'fa' is often exact, so the store check
2245 // of fa[1]=x will fold up, without testing the nullness of x.
2246 switch (static_subtype_check(superk, subk)) {
2247 case SSC_always_false:
2248 {
2249 Node* always_fail = control();
2250 set_control(top());
2251 return always_fail;
2252 }
2253 case SSC_always_true:
2254 return top();
2255 case SSC_easy_test:
2256 {
2257 // Just do a direct pointer compare and be done.
2258 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2259 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2260 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2261 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2262 return _gvn.transform( new(C, 1) IfFalseNode(iff) );
2263 }
2264 case SSC_full_test:
2265 break;
2266 default:
2267 ShouldNotReachHere();
2268 }
2269 }
2271 // %%% Possible further optimization: Even if the superklass is not exact,
2272 // if the subklass is the unique subtype of the superklass, the check
2273 // will always succeed. We could leave a dependency behind to ensure this.
2275 // First load the super-klass's check-offset
2276 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
2277 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2278 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
2279 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2281 // Load from the sub-klass's super-class display list, or a 1-word cache of
2282 // the secondary superclass list, or a failing value with a sentinel offset
2283 // if the super-klass is an interface or exceptionally deep in the Java
2284 // hierarchy and we have to scan the secondary superclass list the hard way.
2285 // Worst-case type is a little odd: NULL is allowed as a result (usually
2286 // klass loads can never produce a NULL).
2287 Node *chk_off_X = ConvI2X(chk_off);
2288 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2289 // For some types like interfaces the following loadKlass is from a 1-word
2290 // cache which is mutable so can't use immutable memory. Other
2291 // types load from the super-class display table which is immutable.
2292 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2293 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2295 // Compile speed common case: ARE a subtype and we canNOT fail
2296 if( superklass == nkls )
2297 return top(); // false path is dead; no test needed.
2299 // See if we get an immediate positive hit. Happens roughly 83% of the
2300 // time. Test to see if the value loaded just previously from the subklass
2301 // is exactly the superklass.
2302 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2303 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2304 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2305 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2306 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2308 // Compile speed common case: Check for being deterministic right now. If
2309 // chk_off is a constant and not equal to cacheoff then we are NOT a
2310 // subklass. In this case we need exactly the 1 test above and we can
2311 // return those results immediately.
2312 if (!might_be_cache) {
2313 Node* not_subtype_ctrl = control();
2314 set_control(iftrue1); // We need exactly the 1 test above
2315 return not_subtype_ctrl;
2316 }
2318 // Gather the various success & failures here
2319 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2320 record_for_igvn(r_ok_subtype);
2321 RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2322 record_for_igvn(r_not_subtype);
2324 r_ok_subtype->init_req(1, iftrue1);
2326 // Check for immediate negative hit. Happens roughly 11% of the time (which
2327 // is roughly 63% of the remaining cases). Test to see if the loaded
2328 // check-offset points into the subklass display list or the 1-element
2329 // cache. If it points to the display (and NOT the cache) and the display
2330 // missed then it's not a subtype.
2331 Node *cacheoff = _gvn.intcon(cacheoff_con);
2332 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2333 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2334 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2335 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2336 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2338 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2339 // No performance impact (too rare) but allows sharing of secondary arrays
2340 // which has some footprint reduction.
2341 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2342 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2343 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2344 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2345 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2347 // -- Roads not taken here: --
2348 // We could also have chosen to perform the self-check at the beginning
2349 // of this code sequence, as the assembler does. This would not pay off
2350 // the same way, since the optimizer, unlike the assembler, can perform
2351 // static type analysis to fold away many successful self-checks.
2352 // Non-foldable self checks work better here in second position, because
2353 // the initial primary superclass check subsumes a self-check for most
2354 // types. An exception would be a secondary type like array-of-interface,
2355 // which does not appear in its own primary supertype display.
2356 // Finally, we could have chosen to move the self-check into the
2357 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2358 // dependent manner. But it is worthwhile to have the check here,
2359 // where it can be perhaps be optimized. The cost in code space is
2360 // small (register compare, branch).
2362 // Now do a linear scan of the secondary super-klass array. Again, no real
2363 // performance impact (too rare) but it's gotta be done.
2364 // Since the code is rarely used, there is no penalty for moving it
2365 // out of line, and it can only improve I-cache density.
2366 // The decision to inline or out-of-line this final check is platform
2367 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2368 Node* psc = _gvn.transform(
2369 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2371 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2372 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2373 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2374 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2375 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2377 // Return false path; set default control to true path.
2378 set_control( _gvn.transform(r_ok_subtype) );
2379 return _gvn.transform(r_not_subtype);
2380 }
2382 //----------------------------static_subtype_check-----------------------------
2383 // Shortcut important common cases when superklass is exact:
2384 // (0) superklass is java.lang.Object (can occur in reflective code)
2385 // (1) subklass is already limited to a subtype of superklass => always ok
2386 // (2) subklass does not overlap with superklass => always fail
2387 // (3) superklass has NO subtypes and we can check with a simple compare.
2388 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2389 if (StressReflectiveCode) {
2390 return SSC_full_test; // Let caller generate the general case.
2391 }
2393 if (superk == env()->Object_klass()) {
2394 return SSC_always_true; // (0) this test cannot fail
2395 }
2397 ciType* superelem = superk;
2398 if (superelem->is_array_klass())
2399 superelem = superelem->as_array_klass()->base_element_type();
2401 if (!subk->is_interface()) { // cannot trust static interface types yet
2402 if (subk->is_subtype_of(superk)) {
2403 return SSC_always_true; // (1) false path dead; no dynamic test needed
2404 }
2405 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2406 !superk->is_subtype_of(subk)) {
2407 return SSC_always_false;
2408 }
2409 }
2411 // If casting to an instance klass, it must have no subtypes
2412 if (superk->is_interface()) {
2413 // Cannot trust interfaces yet.
2414 // %%% S.B. superk->nof_implementors() == 1
2415 } else if (superelem->is_instance_klass()) {
2416 ciInstanceKlass* ik = superelem->as_instance_klass();
2417 if (!ik->has_subklass() && !ik->is_interface()) {
2418 if (!ik->is_final()) {
2419 // Add a dependency if there is a chance of a later subclass.
2420 C->dependencies()->assert_leaf_type(ik);
2421 }
2422 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2423 }
2424 } else {
2425 // A primitive array type has no subtypes.
2426 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2427 }
2429 return SSC_full_test;
2430 }
2432 // Profile-driven exact type check:
2433 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2434 float prob,
2435 Node* *casted_receiver) {
2436 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2437 Node* recv_klass = load_object_klass(receiver);
2438 Node* want_klass = makecon(tklass);
2439 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2440 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2441 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2442 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2443 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2445 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2446 assert(recv_xtype->klass_is_exact(), "");
2448 // Subsume downstream occurrences of receiver with a cast to
2449 // recv_xtype, since now we know what the type will be.
2450 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2451 (*casted_receiver) = _gvn.transform(cast);
2452 // (User must make the replace_in_map call.)
2454 return fail;
2455 }
2458 //-------------------------------gen_instanceof--------------------------------
2459 // Generate an instance-of idiom. Used by both the instance-of bytecode
2460 // and the reflective instance-of call.
2461 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) {
2462 C->set_has_split_ifs(true); // Has chance for split-if optimization
2463 assert( !stopped(), "dead parse path should be checked in callers" );
2464 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2465 "must check for not-null not-dead klass in callers");
2467 // Make the merge point
2468 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2469 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2470 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2471 C->set_has_split_ifs(true); // Has chance for split-if optimization
2473 // Null check; get casted pointer; set region slot 3
2474 Node* null_ctl = top();
2475 Node* not_null_obj = null_check_oop(subobj, &null_ctl);
2477 // If not_null_obj is dead, only null-path is taken
2478 if (stopped()) { // Doing instance-of on a NULL?
2479 set_control(null_ctl);
2480 return intcon(0);
2481 }
2482 region->init_req(_null_path, null_ctl);
2483 phi ->init_req(_null_path, intcon(0)); // Set null path value
2485 // Load the object's klass
2486 Node* obj_klass = load_object_klass(not_null_obj);
2488 // Generate the subtype check
2489 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2491 // Plug in the success path to the general merge in slot 1.
2492 region->init_req(_obj_path, control());
2493 phi ->init_req(_obj_path, intcon(1));
2495 // Plug in the failing path to the general merge in slot 2.
2496 region->init_req(_fail_path, not_subtype_ctrl);
2497 phi ->init_req(_fail_path, intcon(0));
2499 // Return final merged results
2500 set_control( _gvn.transform(region) );
2501 record_for_igvn(region);
2502 return _gvn.transform(phi);
2503 }
2505 //-------------------------------gen_checkcast---------------------------------
2506 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2507 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2508 // uncommon-trap paths work. Adjust stack after this call.
2509 // If failure_control is supplied and not null, it is filled in with
2510 // the control edge for the cast failure. Otherwise, an appropriate
2511 // uncommon trap or exception is thrown.
2512 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2513 Node* *failure_control) {
2514 kill_dead_locals(); // Benefit all the uncommon traps
2515 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2516 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2518 // Fast cutout: Check the case that the cast is vacuously true.
2519 // This detects the common cases where the test will short-circuit
2520 // away completely. We do this before we perform the null check,
2521 // because if the test is going to turn into zero code, we don't
2522 // want a residual null check left around. (Causes a slowdown,
2523 // for example, in some objArray manipulations, such as a[i]=a[j].)
2524 if (tk->singleton()) {
2525 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2526 if (objtp != NULL && objtp->klass() != NULL) {
2527 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2528 case SSC_always_true:
2529 return obj;
2530 case SSC_always_false:
2531 // It needs a null check because a null will *pass* the cast check.
2532 // A non-null value will always produce an exception.
2533 return do_null_assert(obj, T_OBJECT);
2534 }
2535 }
2536 }
2538 ciProfileData* data = NULL;
2539 if (failure_control == NULL) { // use MDO in regular case only
2540 assert(java_bc() == Bytecodes::_aastore ||
2541 java_bc() == Bytecodes::_checkcast,
2542 "interpreter profiles type checks only for these BCs");
2543 data = method()->method_data()->bci_to_data(bci());
2544 }
2546 // Make the merge point
2547 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2548 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2549 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
2550 C->set_has_split_ifs(true); // Has chance for split-if optimization
2552 // Use null-cast information if it is available
2553 bool never_see_null = false;
2554 // If we see an unexpected null at a check-cast we record it and force a
2555 // recompile; the offending check-cast will be compiled to handle NULLs.
2556 // If we see several offending BCIs, then all checkcasts in the
2557 // method will be compiled to handle NULLs.
2558 if (UncommonNullCast // Cutout for this technique
2559 && failure_control == NULL // regular case
2560 && obj != null() // And not the -Xcomp stupid case?
2561 && !too_many_traps(Deoptimization::Reason_null_check)) {
2562 // Finally, check the "null_seen" bit from the interpreter.
2563 if (data == NULL || !data->as_BitData()->null_seen()) {
2564 never_see_null = true;
2565 }
2566 }
2568 // Null check; get casted pointer; set region slot 3
2569 Node* null_ctl = top();
2570 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2572 // If not_null_obj is dead, only null-path is taken
2573 if (stopped()) { // Doing instance-of on a NULL?
2574 set_control(null_ctl);
2575 return null();
2576 }
2577 region->init_req(_null_path, null_ctl);
2578 phi ->init_req(_null_path, null()); // Set null path value
2580 Node* cast_obj = NULL; // the casted version of the object
2582 // If the profile has seen exactly one type, narrow to that type.
2583 // (The subsequent subtype check will always fold up.)
2584 if (UseTypeProfile && TypeProfileCasts && data != NULL &&
2585 // Counter has never been decremented (due to cast failure).
2586 // ...This is a reasonable thing to expect. It is true of
2587 // all casts inserted by javac to implement generic types.
2588 data->as_CounterData()->count() >= 0 &&
2589 !too_many_traps(Deoptimization::Reason_class_check)) {
2590 // (No, this isn't a call, but it's enough like a virtual call
2591 // to use the same ciMethod accessor to get the profile info...)
2592 ciCallProfile profile = method()->call_profile_at_bci(bci());
2593 if (profile.count() >= 0 && // no cast failures here
2594 profile.has_receiver(0) &&
2595 profile.morphism() == 1) {
2596 ciKlass* exact_kls = profile.receiver(0);
2597 int ssc = static_subtype_check(tk->klass(), exact_kls);
2598 if (ssc == SSC_always_true) {
2599 // If we narrow the type to match what the type profile sees,
2600 // we can then remove the rest of the cast.
2601 // This is a win, even if the exact_kls is very specific,
2602 // because downstream operations, such as method calls,
2603 // will often benefit from the sharper type.
2604 Node* exact_obj = not_null_obj; // will get updated in place...
2605 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2606 &exact_obj);
2607 { PreserveJVMState pjvms(this);
2608 set_control(slow_ctl);
2609 uncommon_trap(Deoptimization::Reason_class_check,
2610 Deoptimization::Action_maybe_recompile);
2611 }
2612 if (failure_control != NULL) // failure is now impossible
2613 (*failure_control) = top();
2614 replace_in_map(not_null_obj, exact_obj);
2615 // adjust the type of the phi to the exact klass:
2616 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR));
2617 cast_obj = exact_obj;
2618 }
2619 // assert(cast_obj != NULL)... except maybe the profile lied to us.
2620 }
2621 }
2623 if (cast_obj == NULL) {
2624 // Load the object's klass
2625 Node* obj_klass = load_object_klass(not_null_obj);
2627 // Generate the subtype check
2628 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2630 // Plug in success path into the merge
2631 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2632 not_null_obj, toop));
2633 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2634 if (failure_control == NULL) {
2635 if (not_subtype_ctrl != top()) { // If failure is possible
2636 PreserveJVMState pjvms(this);
2637 set_control(not_subtype_ctrl);
2638 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2639 }
2640 } else {
2641 (*failure_control) = not_subtype_ctrl;
2642 }
2643 }
2645 region->init_req(_obj_path, control());
2646 phi ->init_req(_obj_path, cast_obj);
2648 // A merge of NULL or Casted-NotNull obj
2649 Node* res = _gvn.transform(phi);
2651 // Note I do NOT always 'replace_in_map(obj,result)' here.
2652 // if( tk->klass()->can_be_primary_super() )
2653 // This means that if I successfully store an Object into an array-of-String
2654 // I 'forget' that the Object is really now known to be a String. I have to
2655 // do this because we don't have true union types for interfaces - if I store
2656 // a Baz into an array-of-Interface and then tell the optimizer it's an
2657 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2658 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2659 // replace_in_map( obj, res );
2661 // Return final merged results
2662 set_control( _gvn.transform(region) );
2663 record_for_igvn(region);
2664 return res;
2665 }
2667 //------------------------------next_monitor-----------------------------------
2668 // What number should be given to the next monitor?
2669 int GraphKit::next_monitor() {
2670 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2671 int next = current + C->sync_stack_slots();
2672 // Keep the toplevel high water mark current:
2673 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2674 return current;
2675 }
2677 //------------------------------insert_mem_bar---------------------------------
2678 // Memory barrier to avoid floating things around
2679 // The membar serves as a pinch point between both control and all memory slices.
2680 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2681 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2682 mb->init_req(TypeFunc::Control, control());
2683 mb->init_req(TypeFunc::Memory, reset_memory());
2684 Node* membar = _gvn.transform(mb);
2685 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2686 set_all_memory_call(membar);
2687 return membar;
2688 }
2690 //-------------------------insert_mem_bar_volatile----------------------------
2691 // Memory barrier to avoid floating things around
2692 // The membar serves as a pinch point between both control and memory(alias_idx).
2693 // If you want to make a pinch point on all memory slices, do not use this
2694 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2695 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2696 // When Parse::do_put_xxx updates a volatile field, it appends a series
2697 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2698 // The first membar is on the same memory slice as the field store opcode.
2699 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2700 // All the other membars (for other volatile slices, including AliasIdxBot,
2701 // which stands for all unknown volatile slices) are control-dependent
2702 // on the first membar. This prevents later volatile loads or stores
2703 // from sliding up past the just-emitted store.
2705 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2706 mb->set_req(TypeFunc::Control,control());
2707 if (alias_idx == Compile::AliasIdxBot) {
2708 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2709 } else {
2710 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2711 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2712 }
2713 Node* membar = _gvn.transform(mb);
2714 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2715 if (alias_idx == Compile::AliasIdxBot) {
2716 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2717 } else {
2718 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2719 }
2720 return membar;
2721 }
2723 //------------------------------shared_lock------------------------------------
2724 // Emit locking code.
2725 FastLockNode* GraphKit::shared_lock(Node* obj) {
2726 // bci is either a monitorenter bc or InvocationEntryBci
2727 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2728 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2730 if( !GenerateSynchronizationCode )
2731 return NULL; // Not locking things?
2732 if (stopped()) // Dead monitor?
2733 return NULL;
2735 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2737 // Box the stack location
2738 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2739 Node* mem = reset_memory();
2741 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2742 if (PrintPreciseBiasedLockingStatistics) {
2743 // Create the counters for this fast lock.
2744 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2745 }
2746 // Add monitor to debug info for the slow path. If we block inside the
2747 // slow path and de-opt, we need the monitor hanging around
2748 map()->push_monitor( flock );
2750 const TypeFunc *tf = LockNode::lock_type();
2751 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2753 lock->init_req( TypeFunc::Control, control() );
2754 lock->init_req( TypeFunc::Memory , mem );
2755 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2756 lock->init_req( TypeFunc::FramePtr, frameptr() );
2757 lock->init_req( TypeFunc::ReturnAdr, top() );
2759 lock->init_req(TypeFunc::Parms + 0, obj);
2760 lock->init_req(TypeFunc::Parms + 1, box);
2761 lock->init_req(TypeFunc::Parms + 2, flock);
2762 add_safepoint_edges(lock);
2764 lock = _gvn.transform( lock )->as_Lock();
2766 // lock has no side-effects, sets few values
2767 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2769 insert_mem_bar(Op_MemBarAcquire);
2771 // Add this to the worklist so that the lock can be eliminated
2772 record_for_igvn(lock);
2774 #ifndef PRODUCT
2775 if (PrintLockStatistics) {
2776 // Update the counter for this lock. Don't bother using an atomic
2777 // operation since we don't require absolute accuracy.
2778 lock->create_lock_counter(map()->jvms());
2779 int adr_type = Compile::AliasIdxRaw;
2780 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr()));
2781 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
2782 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
2783 store_to_memory(control(), counter_addr, incr, T_INT, adr_type);
2784 }
2785 #endif
2787 return flock;
2788 }
2791 //------------------------------shared_unlock----------------------------------
2792 // Emit unlocking code.
2793 void GraphKit::shared_unlock(Node* box, Node* obj) {
2794 // bci is either a monitorenter bc or InvocationEntryBci
2795 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2796 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2798 if( !GenerateSynchronizationCode )
2799 return;
2800 if (stopped()) { // Dead monitor?
2801 map()->pop_monitor(); // Kill monitor from debug info
2802 return;
2803 }
2805 // Memory barrier to avoid floating things down past the locked region
2806 insert_mem_bar(Op_MemBarRelease);
2808 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2809 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2810 uint raw_idx = Compile::AliasIdxRaw;
2811 unlock->init_req( TypeFunc::Control, control() );
2812 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2813 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2814 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2815 unlock->init_req( TypeFunc::ReturnAdr, top() );
2817 unlock->init_req(TypeFunc::Parms + 0, obj);
2818 unlock->init_req(TypeFunc::Parms + 1, box);
2819 unlock = _gvn.transform(unlock)->as_Unlock();
2821 Node* mem = reset_memory();
2823 // unlock has no side-effects, sets few values
2824 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2826 // Kill monitor from debug info
2827 map()->pop_monitor( );
2828 }
2830 //-------------------------------get_layout_helper-----------------------------
2831 // If the given klass is a constant or known to be an array,
2832 // fetch the constant layout helper value into constant_value
2833 // and return (Node*)NULL. Otherwise, load the non-constant
2834 // layout helper value, and return the node which represents it.
2835 // This two-faced routine is useful because allocation sites
2836 // almost always feature constant types.
2837 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2838 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2839 if (!StressReflectiveCode && inst_klass != NULL) {
2840 ciKlass* klass = inst_klass->klass();
2841 bool xklass = inst_klass->klass_is_exact();
2842 if (xklass || klass->is_array_klass()) {
2843 jint lhelper = klass->layout_helper();
2844 if (lhelper != Klass::_lh_neutral_value) {
2845 constant_value = lhelper;
2846 return (Node*) NULL;
2847 }
2848 }
2849 }
2850 constant_value = Klass::_lh_neutral_value; // put in a known value
2851 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
2852 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2853 }
2855 // We just put in an allocate/initialize with a big raw-memory effect.
2856 // Hook selected additional alias categories on the initialization.
2857 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2858 MergeMemNode* init_in_merge,
2859 Node* init_out_raw) {
2860 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2861 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2863 Node* prevmem = kit.memory(alias_idx);
2864 init_in_merge->set_memory_at(alias_idx, prevmem);
2865 kit.set_memory(init_out_raw, alias_idx);
2866 }
2868 //---------------------------set_output_for_allocation-------------------------
2869 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2870 const TypeOopPtr* oop_type,
2871 bool raw_mem_only) {
2872 int rawidx = Compile::AliasIdxRaw;
2873 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2874 add_safepoint_edges(alloc);
2875 Node* allocx = _gvn.transform(alloc);
2876 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2877 // create memory projection for i_o
2878 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2879 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2881 // create a memory projection as for the normal control path
2882 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2883 set_memory(malloc, rawidx);
2885 // a normal slow-call doesn't change i_o, but an allocation does
2886 // we create a separate i_o projection for the normal control path
2887 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2888 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2890 // put in an initialization barrier
2891 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2892 rawoop)->as_Initialize();
2893 assert(alloc->initialization() == init, "2-way macro link must work");
2894 assert(init ->allocation() == alloc, "2-way macro link must work");
2895 if (ReduceFieldZeroing && !raw_mem_only) {
2896 // Extract memory strands which may participate in the new object's
2897 // initialization, and source them from the new InitializeNode.
2898 // This will allow us to observe initializations when they occur,
2899 // and link them properly (as a group) to the InitializeNode.
2900 assert(init->in(InitializeNode::Memory) == malloc, "");
2901 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
2902 init->set_req(InitializeNode::Memory, minit_in);
2903 record_for_igvn(minit_in); // fold it up later, if possible
2904 Node* minit_out = memory(rawidx);
2905 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
2906 if (oop_type->isa_aryptr()) {
2907 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
2908 int elemidx = C->get_alias_index(telemref);
2909 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
2910 } else if (oop_type->isa_instptr()) {
2911 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
2912 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
2913 ciField* field = ik->nonstatic_field_at(i);
2914 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
2915 continue; // do not bother to track really large numbers of fields
2916 // Find (or create) the alias category for this field:
2917 int fieldidx = C->alias_type(field)->index();
2918 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
2919 }
2920 }
2921 }
2923 // Cast raw oop to the real thing...
2924 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
2925 javaoop = _gvn.transform(javaoop);
2926 C->set_recent_alloc(control(), javaoop);
2927 assert(just_allocated_object(control()) == javaoop, "just allocated");
2929 #ifdef ASSERT
2930 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
2931 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
2932 "Ideal_allocation works");
2933 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
2934 "Ideal_allocation works");
2935 if (alloc->is_AllocateArray()) {
2936 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
2937 "Ideal_allocation works");
2938 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
2939 "Ideal_allocation works");
2940 } else {
2941 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
2942 }
2943 }
2944 #endif //ASSERT
2946 return javaoop;
2947 }
2949 //---------------------------new_instance--------------------------------------
2950 // This routine takes a klass_node which may be constant (for a static type)
2951 // or may be non-constant (for reflective code). It will work equally well
2952 // for either, and the graph will fold nicely if the optimizer later reduces
2953 // the type to a constant.
2954 // The optional arguments are for specialized use by intrinsics:
2955 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
2956 // - If 'raw_mem_only', do not cast the result to an oop.
2957 // - If 'return_size_val', report the the total object size to the caller.
2958 Node* GraphKit::new_instance(Node* klass_node,
2959 Node* extra_slow_test,
2960 bool raw_mem_only, // affect only raw memory
2961 Node* *return_size_val) {
2962 // Compute size in doublewords
2963 // The size is always an integral number of doublewords, represented
2964 // as a positive bytewise size stored in the klass's layout_helper.
2965 // The layout_helper also encodes (in a low bit) the need for a slow path.
2966 jint layout_con = Klass::_lh_neutral_value;
2967 Node* layout_val = get_layout_helper(klass_node, layout_con);
2968 int layout_is_con = (layout_val == NULL);
2970 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
2971 // Generate the initial go-slow test. It's either ALWAYS (return a
2972 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
2973 // case) a computed value derived from the layout_helper.
2974 Node* initial_slow_test = NULL;
2975 if (layout_is_con) {
2976 assert(!StressReflectiveCode, "stress mode does not use these paths");
2977 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
2978 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
2980 } else { // reflective case
2981 // This reflective path is used by Unsafe.allocateInstance.
2982 // (It may be stress-tested by specifying StressReflectiveCode.)
2983 // Basically, we want to get into the VM is there's an illegal argument.
2984 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
2985 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
2986 if (extra_slow_test != intcon(0)) {
2987 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
2988 }
2989 // (Macro-expander will further convert this to a Bool, if necessary.)
2990 }
2992 // Find the size in bytes. This is easy; it's the layout_helper.
2993 // The size value must be valid even if the slow path is taken.
2994 Node* size = NULL;
2995 if (layout_is_con) {
2996 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
2997 } else { // reflective case
2998 // This reflective path is used by clone and Unsafe.allocateInstance.
2999 size = ConvI2X(layout_val);
3001 // Clear the low bits to extract layout_helper_size_in_bytes:
3002 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3003 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3004 size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
3005 }
3006 if (return_size_val != NULL) {
3007 (*return_size_val) = size;
3008 }
3010 // This is a precise notnull oop of the klass.
3011 // (Actually, it need not be precise if this is a reflective allocation.)
3012 // It's what we cast the result to.
3013 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3014 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3015 const TypeOopPtr* oop_type = tklass->as_instance_type();
3017 // Now generate allocation code
3019 // The entire memory state is needed for slow path of the allocation
3020 // since GC and deoptimization can happened.
3021 Node *mem = reset_memory();
3022 set_all_memory(mem); // Create new memory state
3024 AllocateNode* alloc
3025 = new (C, AllocateNode::ParmLimit)
3026 AllocateNode(C, AllocateNode::alloc_type(),
3027 control(), mem, i_o(),
3028 size, klass_node,
3029 initial_slow_test);
3031 return set_output_for_allocation(alloc, oop_type, raw_mem_only);
3032 }
3034 //-------------------------------new_array-------------------------------------
3035 // helper for both newarray and anewarray
3036 // The 'length' parameter is (obviously) the length of the array.
3037 // See comments on new_instance for the meaning of the other arguments.
3038 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3039 Node* length, // number of array elements
3040 int nargs, // number of arguments to push back for uncommon trap
3041 bool raw_mem_only, // affect only raw memory
3042 Node* *return_size_val) {
3043 jint layout_con = Klass::_lh_neutral_value;
3044 Node* layout_val = get_layout_helper(klass_node, layout_con);
3045 int layout_is_con = (layout_val == NULL);
3047 if (!layout_is_con && !StressReflectiveCode &&
3048 !too_many_traps(Deoptimization::Reason_class_check)) {
3049 // This is a reflective array creation site.
3050 // Optimistically assume that it is a subtype of Object[],
3051 // so that we can fold up all the address arithmetic.
3052 layout_con = Klass::array_layout_helper(T_OBJECT);
3053 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
3054 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
3055 { BuildCutout unless(this, bol_lh, PROB_MAX);
3056 _sp += nargs;
3057 uncommon_trap(Deoptimization::Reason_class_check,
3058 Deoptimization::Action_maybe_recompile);
3059 }
3060 layout_val = NULL;
3061 layout_is_con = true;
3062 }
3064 // Generate the initial go-slow test. Make sure we do not overflow
3065 // if length is huge (near 2Gig) or negative! We do not need
3066 // exact double-words here, just a close approximation of needed
3067 // double-words. We can't add any offset or rounding bits, lest we
3068 // take a size -1 of bytes and make it positive. Use an unsigned
3069 // compare, so negative sizes look hugely positive.
3070 int fast_size_limit = FastAllocateSizeLimit;
3071 if (layout_is_con) {
3072 assert(!StressReflectiveCode, "stress mode does not use these paths");
3073 // Increase the size limit if we have exact knowledge of array type.
3074 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3075 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3076 }
3078 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3079 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3080 if (initial_slow_test->is_Bool()) {
3081 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3082 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3083 }
3085 // --- Size Computation ---
3086 // array_size = round_to_heap(array_header + (length << elem_shift));
3087 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3088 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3089 // The rounding mask is strength-reduced, if possible.
3090 int round_mask = MinObjAlignmentInBytes - 1;
3091 Node* header_size = NULL;
3092 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3093 // (T_BYTE has the weakest alignment and size restrictions...)
3094 if (layout_is_con) {
3095 int hsize = Klass::layout_helper_header_size(layout_con);
3096 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3097 BasicType etype = Klass::layout_helper_element_type(layout_con);
3098 if ((round_mask & ~right_n_bits(eshift)) == 0)
3099 round_mask = 0; // strength-reduce it if it goes away completely
3100 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3101 assert(header_size_min <= hsize, "generic minimum is smallest");
3102 header_size_min = hsize;
3103 header_size = intcon(hsize + round_mask);
3104 } else {
3105 Node* hss = intcon(Klass::_lh_header_size_shift);
3106 Node* hsm = intcon(Klass::_lh_header_size_mask);
3107 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3108 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3109 Node* mask = intcon(round_mask);
3110 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3111 }
3113 Node* elem_shift = NULL;
3114 if (layout_is_con) {
3115 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3116 if (eshift != 0)
3117 elem_shift = intcon(eshift);
3118 } else {
3119 // There is no need to mask or shift this value.
3120 // The semantics of LShiftINode include an implicit mask to 0x1F.
3121 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3122 elem_shift = layout_val;
3123 }
3125 // Transition to native address size for all offset calculations:
3126 Node* lengthx = ConvI2X(length);
3127 Node* headerx = ConvI2X(header_size);
3128 #ifdef _LP64
3129 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3130 if (tllen != NULL && tllen->_lo < 0) {
3131 // Add a manual constraint to a positive range. Cf. array_element_address.
3132 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3133 if (size_max > tllen->_hi) size_max = tllen->_hi;
3134 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3135 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3136 }
3137 }
3138 #endif
3140 // Combine header size (plus rounding) and body size. Then round down.
3141 // This computation cannot overflow, because it is used only in two
3142 // places, one where the length is sharply limited, and the other
3143 // after a successful allocation.
3144 Node* abody = lengthx;
3145 if (elem_shift != NULL)
3146 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3147 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3148 if (round_mask != 0) {
3149 Node* mask = MakeConX(~round_mask);
3150 size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3151 }
3152 // else if round_mask == 0, the size computation is self-rounding
3154 if (return_size_val != NULL) {
3155 // This is the size
3156 (*return_size_val) = size;
3157 }
3159 // Now generate allocation code
3161 // The entire memory state is needed for slow path of the allocation
3162 // since GC and deoptimization can happened.
3163 Node *mem = reset_memory();
3164 set_all_memory(mem); // Create new memory state
3166 // Create the AllocateArrayNode and its result projections
3167 AllocateArrayNode* alloc
3168 = new (C, AllocateArrayNode::ParmLimit)
3169 AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3170 control(), mem, i_o(),
3171 size, klass_node,
3172 initial_slow_test,
3173 length);
3175 // Cast to correct type. Note that the klass_node may be constant or not,
3176 // and in the latter case the actual array type will be inexact also.
3177 // (This happens via a non-constant argument to inline_native_newArray.)
3178 // In any case, the value of klass_node provides the desired array type.
3179 const TypeInt* length_type = _gvn.find_int_type(length);
3180 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3181 if (ary_type->isa_aryptr() && length_type != NULL) {
3182 // Try to get a better type than POS for the size
3183 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3184 }
3186 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
3188 // Cast length on remaining path to be as narrow as possible
3189 if (map()->find_edge(length) >= 0) {
3190 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3191 if (ccast != length) {
3192 _gvn.set_type_bottom(ccast);
3193 record_for_igvn(ccast);
3194 replace_in_map(length, ccast);
3195 }
3196 }
3198 return javaoop;
3199 }
3201 // The following "Ideal_foo" functions are placed here because they recognize
3202 // the graph shapes created by the functions immediately above.
3204 //---------------------------Ideal_allocation----------------------------------
3205 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3206 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3207 if (ptr == NULL) { // reduce dumb test in callers
3208 return NULL;
3209 }
3210 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3211 ptr = ptr->in(1);
3212 if (ptr == NULL) return NULL;
3213 }
3214 if (ptr->is_Proj()) {
3215 Node* allo = ptr->in(0);
3216 if (allo != NULL && allo->is_Allocate()) {
3217 return allo->as_Allocate();
3218 }
3219 }
3220 // Report failure to match.
3221 return NULL;
3222 }
3224 // Fancy version which also strips off an offset (and reports it to caller).
3225 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3226 intptr_t& offset) {
3227 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3228 if (base == NULL) return NULL;
3229 return Ideal_allocation(base, phase);
3230 }
3232 // Trace Initialize <- Proj[Parm] <- Allocate
3233 AllocateNode* InitializeNode::allocation() {
3234 Node* rawoop = in(InitializeNode::RawAddress);
3235 if (rawoop->is_Proj()) {
3236 Node* alloc = rawoop->in(0);
3237 if (alloc->is_Allocate()) {
3238 return alloc->as_Allocate();
3239 }
3240 }
3241 return NULL;
3242 }
3244 // Trace Allocate -> Proj[Parm] -> Initialize
3245 InitializeNode* AllocateNode::initialization() {
3246 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3247 if (rawoop == NULL) return NULL;
3248 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3249 Node* init = rawoop->fast_out(i);
3250 if (init->is_Initialize()) {
3251 assert(init->as_Initialize()->allocation() == this, "2-way link");
3252 return init->as_Initialize();
3253 }
3254 }
3255 return NULL;
3256 }
3258 //----------------------------- store barriers ----------------------------
3259 #define __ ideal.
3261 void GraphKit::sync_kit(IdealKit& ideal) {
3262 // Final sync IdealKit and graphKit.
3263 __ drain_delay_transform();
3264 set_all_memory(__ merged_memory());
3265 set_control(__ ctrl());
3266 }
3268 // vanilla/CMS post barrier
3269 // Insert a write-barrier store. This is to let generational GC work; we have
3270 // to flag all oop-stores before the next GC point.
3271 void GraphKit::write_barrier_post(Node* oop_store,
3272 Node* obj,
3273 Node* adr,
3274 uint adr_idx,
3275 Node* val,
3276 bool use_precise) {
3277 // No store check needed if we're storing a NULL or an old object
3278 // (latter case is probably a string constant). The concurrent
3279 // mark sweep garbage collector, however, needs to have all nonNull
3280 // oop updates flagged via card-marks.
3281 if (val != NULL && val->is_Con()) {
3282 // must be either an oop or NULL
3283 const Type* t = val->bottom_type();
3284 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3285 // stores of null never (?) need barriers
3286 return;
3287 ciObject* con = t->is_oopptr()->const_oop();
3288 if (con != NULL
3289 && con->is_perm()
3290 && Universe::heap()->can_elide_permanent_oop_store_barriers())
3291 // no store barrier needed, because no old-to-new ref created
3292 return;
3293 }
3295 if (use_ReduceInitialCardMarks()
3296 && obj == just_allocated_object(control())) {
3297 // We can skip marks on a freshly-allocated object in Eden.
3298 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3299 // That routine informs GC to take appropriate compensating steps,
3300 // upon a slow-path allocation, so as to make this card-mark
3301 // elision safe.
3302 return;
3303 }
3305 if (!use_precise) {
3306 // All card marks for a (non-array) instance are in one place:
3307 adr = obj;
3308 }
3309 // (Else it's an array (or unknown), and we want more precise card marks.)
3310 assert(adr != NULL, "");
3312 IdealKit ideal(gvn(), control(), merged_memory(), true);
3314 // Convert the pointer to an int prior to doing math on it
3315 Node* cast = __ CastPX(__ ctrl(), adr);
3317 // Divide by card size
3318 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3319 "Only one we handle so far.");
3320 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3322 // Combine card table base and card offset
3323 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3325 // Get the alias_index for raw card-mark memory
3326 int adr_type = Compile::AliasIdxRaw;
3327 // Smash zero into card
3328 Node* zero = __ ConI(0);
3329 BasicType bt = T_BYTE;
3330 if( !UseConcMarkSweepGC ) {
3331 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3332 } else {
3333 // Specialized path for CM store barrier
3334 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3335 }
3337 // Final sync IdealKit and GraphKit.
3338 sync_kit(ideal);
3339 }
3341 // G1 pre/post barriers
3342 void GraphKit::g1_write_barrier_pre(Node* obj,
3343 Node* adr,
3344 uint alias_idx,
3345 Node* val,
3346 const TypeOopPtr* val_type,
3347 BasicType bt) {
3348 IdealKit ideal(gvn(), control(), merged_memory(), true);
3350 Node* tls = __ thread(); // ThreadLocalStorage
3352 Node* no_ctrl = NULL;
3353 Node* no_base = __ top();
3354 Node* zero = __ ConI(0);
3356 float likely = PROB_LIKELY(0.999);
3357 float unlikely = PROB_UNLIKELY(0.999);
3359 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3360 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3362 // Offsets into the thread
3363 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3364 PtrQueue::byte_offset_of_active());
3365 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3366 PtrQueue::byte_offset_of_index());
3367 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3368 PtrQueue::byte_offset_of_buf());
3369 // Now the actual pointers into the thread
3371 // set_control( ctl);
3373 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3374 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3375 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3377 // Now some of the values
3379 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3381 // if (!marking)
3382 __ if_then(marking, BoolTest::ne, zero); {
3383 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3385 const Type* t1 = adr->bottom_type();
3386 const Type* t2 = val->bottom_type();
3388 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3389 // if (orig != NULL)
3390 __ if_then(orig, BoolTest::ne, null()); {
3391 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3393 // load original value
3394 // alias_idx correct??
3396 // is the queue for this thread full?
3397 __ if_then(index, BoolTest::ne, zero, likely); {
3399 // decrement the index
3400 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3401 Node* next_indexX = next_index;
3402 #ifdef _LP64
3403 // We could refine the type for what it's worth
3404 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3405 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3406 #endif
3408 // Now get the buffer location we will log the original value into and store it
3409 Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3410 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw);
3412 // update the index
3413 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3415 } __ else_(); {
3417 // logging buffer is full, call the runtime
3418 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3419 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls);
3420 } __ end_if(); // (!index)
3421 } __ end_if(); // (orig != NULL)
3422 } __ end_if(); // (!marking)
3424 // Final sync IdealKit and GraphKit.
3425 sync_kit(ideal);
3426 }
3428 //
3429 // Update the card table and add card address to the queue
3430 //
3431 void GraphKit::g1_mark_card(IdealKit& ideal,
3432 Node* card_adr,
3433 Node* oop_store,
3434 uint oop_alias_idx,
3435 Node* index,
3436 Node* index_adr,
3437 Node* buffer,
3438 const TypeFunc* tf) {
3440 Node* zero = __ ConI(0);
3441 Node* no_base = __ top();
3442 BasicType card_bt = T_BYTE;
3443 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3444 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3446 // Now do the queue work
3447 __ if_then(index, BoolTest::ne, zero); {
3449 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3450 Node* next_indexX = next_index;
3451 #ifdef _LP64
3452 // We could refine the type for what it's worth
3453 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3454 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3455 #endif // _LP64
3456 Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3458 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3459 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3461 } __ else_(); {
3462 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3463 } __ end_if();
3465 }
3467 void GraphKit::g1_write_barrier_post(Node* oop_store,
3468 Node* obj,
3469 Node* adr,
3470 uint alias_idx,
3471 Node* val,
3472 BasicType bt,
3473 bool use_precise) {
3474 // If we are writing a NULL then we need no post barrier
3476 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3477 // Must be NULL
3478 const Type* t = val->bottom_type();
3479 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3480 // No post barrier if writing NULLx
3481 return;
3482 }
3484 if (!use_precise) {
3485 // All card marks for a (non-array) instance are in one place:
3486 adr = obj;
3487 }
3488 // (Else it's an array (or unknown), and we want more precise card marks.)
3489 assert(adr != NULL, "");
3491 IdealKit ideal(gvn(), control(), merged_memory(), true);
3493 Node* tls = __ thread(); // ThreadLocalStorage
3495 Node* no_ctrl = NULL;
3496 Node* no_base = __ top();
3497 float likely = PROB_LIKELY(0.999);
3498 float unlikely = PROB_UNLIKELY(0.999);
3499 Node* zero = __ ConI(0);
3500 Node* zeroX = __ ConX(0);
3502 // Get the alias_index for raw card-mark memory
3503 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3505 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3507 // Offsets into the thread
3508 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3509 PtrQueue::byte_offset_of_index());
3510 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3511 PtrQueue::byte_offset_of_buf());
3513 // Pointers into the thread
3515 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3516 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3518 // Now some values
3520 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3521 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3524 // Convert the store obj pointer to an int prior to doing math on it
3525 // Must use ctrl to prevent "integerized oop" existing across safepoint
3526 Node* cast = __ CastPX(__ ctrl(), adr);
3528 // Divide pointer by card size
3529 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3531 // Combine card table base and card offset
3532 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3534 // If we know the value being stored does it cross regions?
3536 if (val != NULL) {
3537 // Does the store cause us to cross regions?
3539 // Should be able to do an unsigned compare of region_size instead of
3540 // and extra shift. Do we have an unsigned compare??
3541 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3542 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3544 // if (xor_res == 0) same region so skip
3545 __ if_then(xor_res, BoolTest::ne, zeroX); {
3547 // No barrier if we are storing a NULL
3548 __ if_then(val, BoolTest::ne, null(), unlikely); {
3550 // Ok must mark the card if not already dirty
3552 // load the original value of the card
3553 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3555 __ if_then(card_val, BoolTest::ne, zero); {
3556 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3557 } __ end_if();
3558 } __ end_if();
3559 } __ end_if();
3560 } else {
3561 // Object.clone() instrinsic uses this path.
3562 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3563 }
3565 // Final sync IdealKit and GraphKit.
3566 sync_kit(ideal);
3567 }
3568 #undef __