Wed, 27 Apr 2011 15:40:36 -0700
7029167: add support for conditional card marks
Reviewed-by: iveresov, kvn
1 /*
2 * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc_implementation/g1/heapRegion.hpp"
29 #include "gc_interface/collectedHeap.hpp"
30 #include "memory/barrierSet.hpp"
31 #include "memory/cardTableModRefBS.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/graphKit.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/sharedRuntime.hpp"
43 //----------------------------GraphKit-----------------------------------------
44 // Main utility constructor.
45 GraphKit::GraphKit(JVMState* jvms)
46 : Phase(Phase::Parser),
47 _env(C->env()),
48 _gvn(*C->initial_gvn())
49 {
50 _exceptions = jvms->map()->next_exception();
51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
52 set_jvms(jvms);
53 }
55 // Private constructor for parser.
56 GraphKit::GraphKit()
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn())
60 {
61 _exceptions = NULL;
62 set_map(NULL);
63 debug_only(_sp = -99);
64 debug_only(set_bci(-99));
65 }
69 //---------------------------clean_stack---------------------------------------
70 // Clear away rubbish from the stack area of the JVM state.
71 // This destroys any arguments that may be waiting on the stack.
72 void GraphKit::clean_stack(int from_sp) {
73 SafePointNode* map = this->map();
74 JVMState* jvms = this->jvms();
75 int stk_size = jvms->stk_size();
76 int stkoff = jvms->stkoff();
77 Node* top = this->top();
78 for (int i = from_sp; i < stk_size; i++) {
79 if (map->in(stkoff + i) != top) {
80 map->set_req(stkoff + i, top);
81 }
82 }
83 }
86 //--------------------------------sync_jvms-----------------------------------
87 // Make sure our current jvms agrees with our parse state.
88 JVMState* GraphKit::sync_jvms() const {
89 JVMState* jvms = this->jvms();
90 jvms->set_bci(bci()); // Record the new bci in the JVMState
91 jvms->set_sp(sp()); // Record the new sp in the JVMState
92 assert(jvms_in_sync(), "jvms is now in sync");
93 return jvms;
94 }
96 #ifdef ASSERT
97 bool GraphKit::jvms_in_sync() const {
98 Parse* parse = is_Parse();
99 if (parse == NULL) {
100 if (bci() != jvms()->bci()) return false;
101 if (sp() != (int)jvms()->sp()) return false;
102 return true;
103 }
104 if (jvms()->method() != parse->method()) return false;
105 if (jvms()->bci() != parse->bci()) return false;
106 int jvms_sp = jvms()->sp();
107 if (jvms_sp != parse->sp()) return false;
108 int jvms_depth = jvms()->depth();
109 if (jvms_depth != parse->depth()) return false;
110 return true;
111 }
113 // Local helper checks for special internal merge points
114 // used to accumulate and merge exception states.
115 // They are marked by the region's in(0) edge being the map itself.
116 // Such merge points must never "escape" into the parser at large,
117 // until they have been handed to gvn.transform.
118 static bool is_hidden_merge(Node* reg) {
119 if (reg == NULL) return false;
120 if (reg->is_Phi()) {
121 reg = reg->in(0);
122 if (reg == NULL) return false;
123 }
124 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
125 }
127 void GraphKit::verify_map() const {
128 if (map() == NULL) return; // null map is OK
129 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
130 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
131 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
132 }
134 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
135 assert(ex_map->next_exception() == NULL, "not already part of a chain");
136 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
137 }
138 #endif
140 //---------------------------stop_and_kill_map---------------------------------
141 // Set _map to NULL, signalling a stop to further bytecode execution.
142 // First smash the current map's control to a constant, to mark it dead.
143 void GraphKit::stop_and_kill_map() {
144 SafePointNode* dead_map = stop();
145 if (dead_map != NULL) {
146 dead_map->disconnect_inputs(NULL); // Mark the map as killed.
147 assert(dead_map->is_killed(), "must be so marked");
148 }
149 }
152 //--------------------------------stopped--------------------------------------
153 // Tell if _map is NULL, or control is top.
154 bool GraphKit::stopped() {
155 if (map() == NULL) return true;
156 else if (control() == top()) return true;
157 else return false;
158 }
161 //-----------------------------has_ex_handler----------------------------------
162 // Tell if this method or any caller method has exception handlers.
163 bool GraphKit::has_ex_handler() {
164 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
165 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
166 return true;
167 }
168 }
169 return false;
170 }
172 //------------------------------save_ex_oop------------------------------------
173 // Save an exception without blowing stack contents or other JVM state.
174 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
175 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
176 ex_map->add_req(ex_oop);
177 debug_only(verify_exception_state(ex_map));
178 }
180 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
181 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
182 Node* ex_oop = ex_map->in(ex_map->req()-1);
183 if (clear_it) ex_map->del_req(ex_map->req()-1);
184 return ex_oop;
185 }
187 //-----------------------------saved_ex_oop------------------------------------
188 // Recover a saved exception from its map.
189 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
190 return common_saved_ex_oop(ex_map, false);
191 }
193 //--------------------------clear_saved_ex_oop---------------------------------
194 // Erase a previously saved exception from its map.
195 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
196 return common_saved_ex_oop(ex_map, true);
197 }
199 #ifdef ASSERT
200 //---------------------------has_saved_ex_oop----------------------------------
201 // Erase a previously saved exception from its map.
202 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
203 return ex_map->req() == ex_map->jvms()->endoff()+1;
204 }
205 #endif
207 //-------------------------make_exception_state--------------------------------
208 // Turn the current JVM state into an exception state, appending the ex_oop.
209 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
210 sync_jvms();
211 SafePointNode* ex_map = stop(); // do not manipulate this map any more
212 set_saved_ex_oop(ex_map, ex_oop);
213 return ex_map;
214 }
217 //--------------------------add_exception_state--------------------------------
218 // Add an exception to my list of exceptions.
219 void GraphKit::add_exception_state(SafePointNode* ex_map) {
220 if (ex_map == NULL || ex_map->control() == top()) {
221 return;
222 }
223 #ifdef ASSERT
224 verify_exception_state(ex_map);
225 if (has_exceptions()) {
226 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
227 }
228 #endif
230 // If there is already an exception of exactly this type, merge with it.
231 // In particular, null-checks and other low-level exceptions common up here.
232 Node* ex_oop = saved_ex_oop(ex_map);
233 const Type* ex_type = _gvn.type(ex_oop);
234 if (ex_oop == top()) {
235 // No action needed.
236 return;
237 }
238 assert(ex_type->isa_instptr(), "exception must be an instance");
239 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
240 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
241 // We check sp also because call bytecodes can generate exceptions
242 // both before and after arguments are popped!
243 if (ex_type2 == ex_type
244 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
245 combine_exception_states(ex_map, e2);
246 return;
247 }
248 }
250 // No pre-existing exception of the same type. Chain it on the list.
251 push_exception_state(ex_map);
252 }
254 //-----------------------add_exception_states_from-----------------------------
255 void GraphKit::add_exception_states_from(JVMState* jvms) {
256 SafePointNode* ex_map = jvms->map()->next_exception();
257 if (ex_map != NULL) {
258 jvms->map()->set_next_exception(NULL);
259 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
260 next_map = ex_map->next_exception();
261 ex_map->set_next_exception(NULL);
262 add_exception_state(ex_map);
263 }
264 }
265 }
267 //-----------------------transfer_exceptions_into_jvms-------------------------
268 JVMState* GraphKit::transfer_exceptions_into_jvms() {
269 if (map() == NULL) {
270 // We need a JVMS to carry the exceptions, but the map has gone away.
271 // Create a scratch JVMS, cloned from any of the exception states...
272 if (has_exceptions()) {
273 _map = _exceptions;
274 _map = clone_map();
275 _map->set_next_exception(NULL);
276 clear_saved_ex_oop(_map);
277 debug_only(verify_map());
278 } else {
279 // ...or created from scratch
280 JVMState* jvms = new (C) JVMState(_method, NULL);
281 jvms->set_bci(_bci);
282 jvms->set_sp(_sp);
283 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms));
284 set_jvms(jvms);
285 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
286 set_all_memory(top());
287 while (map()->req() < jvms->endoff()) map()->add_req(top());
288 }
289 // (This is a kludge, in case you didn't notice.)
290 set_control(top());
291 }
292 JVMState* jvms = sync_jvms();
293 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
294 jvms->map()->set_next_exception(_exceptions);
295 _exceptions = NULL; // done with this set of exceptions
296 return jvms;
297 }
299 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
300 assert(is_hidden_merge(dstphi), "must be a special merge node");
301 assert(is_hidden_merge(srcphi), "must be a special merge node");
302 uint limit = srcphi->req();
303 for (uint i = PhiNode::Input; i < limit; i++) {
304 dstphi->add_req(srcphi->in(i));
305 }
306 }
307 static inline void add_one_req(Node* dstphi, Node* src) {
308 assert(is_hidden_merge(dstphi), "must be a special merge node");
309 assert(!is_hidden_merge(src), "must not be a special merge node");
310 dstphi->add_req(src);
311 }
313 //-----------------------combine_exception_states------------------------------
314 // This helper function combines exception states by building phis on a
315 // specially marked state-merging region. These regions and phis are
316 // untransformed, and can build up gradually. The region is marked by
317 // having a control input of its exception map, rather than NULL. Such
318 // regions do not appear except in this function, and in use_exception_state.
319 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
320 if (failing()) return; // dying anyway...
321 JVMState* ex_jvms = ex_map->_jvms;
322 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
323 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
324 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
325 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
326 assert(ex_map->req() == phi_map->req(), "matching maps");
327 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
328 Node* hidden_merge_mark = root();
329 Node* region = phi_map->control();
330 MergeMemNode* phi_mem = phi_map->merged_memory();
331 MergeMemNode* ex_mem = ex_map->merged_memory();
332 if (region->in(0) != hidden_merge_mark) {
333 // The control input is not (yet) a specially-marked region in phi_map.
334 // Make it so, and build some phis.
335 region = new (C, 2) RegionNode(2);
336 _gvn.set_type(region, Type::CONTROL);
337 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
338 region->init_req(1, phi_map->control());
339 phi_map->set_control(region);
340 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
341 record_for_igvn(io_phi);
342 _gvn.set_type(io_phi, Type::ABIO);
343 phi_map->set_i_o(io_phi);
344 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
345 Node* m = mms.memory();
346 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
347 record_for_igvn(m_phi);
348 _gvn.set_type(m_phi, Type::MEMORY);
349 mms.set_memory(m_phi);
350 }
351 }
353 // Either or both of phi_map and ex_map might already be converted into phis.
354 Node* ex_control = ex_map->control();
355 // if there is special marking on ex_map also, we add multiple edges from src
356 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
357 // how wide was the destination phi_map, originally?
358 uint orig_width = region->req();
360 if (add_multiple) {
361 add_n_reqs(region, ex_control);
362 add_n_reqs(phi_map->i_o(), ex_map->i_o());
363 } else {
364 // ex_map has no merges, so we just add single edges everywhere
365 add_one_req(region, ex_control);
366 add_one_req(phi_map->i_o(), ex_map->i_o());
367 }
368 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
369 if (mms.is_empty()) {
370 // get a copy of the base memory, and patch some inputs into it
371 const TypePtr* adr_type = mms.adr_type(C);
372 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
373 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
374 mms.set_memory(phi);
375 // Prepare to append interesting stuff onto the newly sliced phi:
376 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
377 }
378 // Append stuff from ex_map:
379 if (add_multiple) {
380 add_n_reqs(mms.memory(), mms.memory2());
381 } else {
382 add_one_req(mms.memory(), mms.memory2());
383 }
384 }
385 uint limit = ex_map->req();
386 for (uint i = TypeFunc::Parms; i < limit; i++) {
387 // Skip everything in the JVMS after tos. (The ex_oop follows.)
388 if (i == tos) i = ex_jvms->monoff();
389 Node* src = ex_map->in(i);
390 Node* dst = phi_map->in(i);
391 if (src != dst) {
392 PhiNode* phi;
393 if (dst->in(0) != region) {
394 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
395 record_for_igvn(phi);
396 _gvn.set_type(phi, phi->type());
397 phi_map->set_req(i, dst);
398 // Prepare to append interesting stuff onto the new phi:
399 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
400 } else {
401 assert(dst->is_Phi(), "nobody else uses a hidden region");
402 phi = (PhiNode*)dst;
403 }
404 if (add_multiple && src->in(0) == ex_control) {
405 // Both are phis.
406 add_n_reqs(dst, src);
407 } else {
408 while (dst->req() < region->req()) add_one_req(dst, src);
409 }
410 const Type* srctype = _gvn.type(src);
411 if (phi->type() != srctype) {
412 const Type* dsttype = phi->type()->meet(srctype);
413 if (phi->type() != dsttype) {
414 phi->set_type(dsttype);
415 _gvn.set_type(phi, dsttype);
416 }
417 }
418 }
419 }
420 }
422 //--------------------------use_exception_state--------------------------------
423 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
424 if (failing()) { stop(); return top(); }
425 Node* region = phi_map->control();
426 Node* hidden_merge_mark = root();
427 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
428 Node* ex_oop = clear_saved_ex_oop(phi_map);
429 if (region->in(0) == hidden_merge_mark) {
430 // Special marking for internal ex-states. Process the phis now.
431 region->set_req(0, region); // now it's an ordinary region
432 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
433 // Note: Setting the jvms also sets the bci and sp.
434 set_control(_gvn.transform(region));
435 uint tos = jvms()->stkoff() + sp();
436 for (uint i = 1; i < tos; i++) {
437 Node* x = phi_map->in(i);
438 if (x->in(0) == region) {
439 assert(x->is_Phi(), "expected a special phi");
440 phi_map->set_req(i, _gvn.transform(x));
441 }
442 }
443 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
444 Node* x = mms.memory();
445 if (x->in(0) == region) {
446 assert(x->is_Phi(), "nobody else uses a hidden region");
447 mms.set_memory(_gvn.transform(x));
448 }
449 }
450 if (ex_oop->in(0) == region) {
451 assert(ex_oop->is_Phi(), "expected a special phi");
452 ex_oop = _gvn.transform(ex_oop);
453 }
454 } else {
455 set_jvms(phi_map->jvms());
456 }
458 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
459 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
460 return ex_oop;
461 }
463 //---------------------------------java_bc-------------------------------------
464 Bytecodes::Code GraphKit::java_bc() const {
465 ciMethod* method = this->method();
466 int bci = this->bci();
467 if (method != NULL && bci != InvocationEntryBci)
468 return method->java_code_at_bci(bci);
469 else
470 return Bytecodes::_illegal;
471 }
473 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
474 bool must_throw) {
475 // if the exception capability is set, then we will generate code
476 // to check the JavaThread.should_post_on_exceptions flag to see
477 // if we actually need to report exception events (for this
478 // thread). If we don't need to report exception events, we will
479 // take the normal fast path provided by add_exception_events. If
480 // exception event reporting is enabled for this thread, we will
481 // take the uncommon_trap in the BuildCutout below.
483 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
484 Node* jthread = _gvn.transform(new (C, 1) ThreadLocalNode());
485 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
486 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false);
488 // Test the should_post_on_exceptions_flag vs. 0
489 Node* chk = _gvn.transform( new (C, 3) CmpINode(should_post_flag, intcon(0)) );
490 Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, BoolTest::eq) );
492 // Branch to slow_path if should_post_on_exceptions_flag was true
493 { BuildCutout unless(this, tst, PROB_MAX);
494 // Do not try anything fancy if we're notifying the VM on every throw.
495 // Cf. case Bytecodes::_athrow in parse2.cpp.
496 uncommon_trap(reason, Deoptimization::Action_none,
497 (ciKlass*)NULL, (char*)NULL, must_throw);
498 }
500 }
502 //------------------------------builtin_throw----------------------------------
503 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
504 bool must_throw = true;
506 if (env()->jvmti_can_post_on_exceptions()) {
507 // check if we must post exception events, take uncommon trap if so
508 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
509 // here if should_post_on_exceptions is false
510 // continue on with the normal codegen
511 }
513 // If this particular condition has not yet happened at this
514 // bytecode, then use the uncommon trap mechanism, and allow for
515 // a future recompilation if several traps occur here.
516 // If the throw is hot, try to use a more complicated inline mechanism
517 // which keeps execution inside the compiled code.
518 bool treat_throw_as_hot = false;
519 ciMethodData* md = method()->method_data();
521 if (ProfileTraps) {
522 if (too_many_traps(reason)) {
523 treat_throw_as_hot = true;
524 }
525 // (If there is no MDO at all, assume it is early in
526 // execution, and that any deopts are part of the
527 // startup transient, and don't need to be remembered.)
529 // Also, if there is a local exception handler, treat all throws
530 // as hot if there has been at least one in this method.
531 if (C->trap_count(reason) != 0
532 && method()->method_data()->trap_count(reason) != 0
533 && has_ex_handler()) {
534 treat_throw_as_hot = true;
535 }
536 }
538 // If this throw happens frequently, an uncommon trap might cause
539 // a performance pothole. If there is a local exception handler,
540 // and if this particular bytecode appears to be deoptimizing often,
541 // let us handle the throw inline, with a preconstructed instance.
542 // Note: If the deopt count has blown up, the uncommon trap
543 // runtime is going to flush this nmethod, not matter what.
544 if (treat_throw_as_hot
545 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
546 // If the throw is local, we use a pre-existing instance and
547 // punt on the backtrace. This would lead to a missing backtrace
548 // (a repeat of 4292742) if the backtrace object is ever asked
549 // for its backtrace.
550 // Fixing this remaining case of 4292742 requires some flavor of
551 // escape analysis. Leave that for the future.
552 ciInstance* ex_obj = NULL;
553 switch (reason) {
554 case Deoptimization::Reason_null_check:
555 ex_obj = env()->NullPointerException_instance();
556 break;
557 case Deoptimization::Reason_div0_check:
558 ex_obj = env()->ArithmeticException_instance();
559 break;
560 case Deoptimization::Reason_range_check:
561 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
562 break;
563 case Deoptimization::Reason_class_check:
564 if (java_bc() == Bytecodes::_aastore) {
565 ex_obj = env()->ArrayStoreException_instance();
566 } else {
567 ex_obj = env()->ClassCastException_instance();
568 }
569 break;
570 }
571 if (failing()) { stop(); return; } // exception allocation might fail
572 if (ex_obj != NULL) {
573 // Cheat with a preallocated exception object.
574 if (C->log() != NULL)
575 C->log()->elem("hot_throw preallocated='1' reason='%s'",
576 Deoptimization::trap_reason_name(reason));
577 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
578 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
580 // Clear the detail message of the preallocated exception object.
581 // Weblogic sometimes mutates the detail message of exceptions
582 // using reflection.
583 int offset = java_lang_Throwable::get_detailMessage_offset();
584 const TypePtr* adr_typ = ex_con->add_offset(offset);
586 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
587 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
588 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT);
590 add_exception_state(make_exception_state(ex_node));
591 return;
592 }
593 }
595 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
596 // It won't be much cheaper than bailing to the interp., since we'll
597 // have to pass up all the debug-info, and the runtime will have to
598 // create the stack trace.
600 // Usual case: Bail to interpreter.
601 // Reserve the right to recompile if we haven't seen anything yet.
603 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
604 if (treat_throw_as_hot
605 && (method()->method_data()->trap_recompiled_at(bci())
606 || C->too_many_traps(reason))) {
607 // We cannot afford to take more traps here. Suffer in the interpreter.
608 if (C->log() != NULL)
609 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
610 Deoptimization::trap_reason_name(reason),
611 C->trap_count(reason));
612 action = Deoptimization::Action_none;
613 }
615 // "must_throw" prunes the JVM state to include only the stack, if there
616 // are no local exception handlers. This should cut down on register
617 // allocation time and code size, by drastically reducing the number
618 // of in-edges on the call to the uncommon trap.
620 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
621 }
624 //----------------------------PreserveJVMState---------------------------------
625 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
626 debug_only(kit->verify_map());
627 _kit = kit;
628 _map = kit->map(); // preserve the map
629 _sp = kit->sp();
630 kit->set_map(clone_map ? kit->clone_map() : NULL);
631 #ifdef ASSERT
632 _bci = kit->bci();
633 Parse* parser = kit->is_Parse();
634 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
635 _block = block;
636 #endif
637 }
638 PreserveJVMState::~PreserveJVMState() {
639 GraphKit* kit = _kit;
640 #ifdef ASSERT
641 assert(kit->bci() == _bci, "bci must not shift");
642 Parse* parser = kit->is_Parse();
643 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
644 assert(block == _block, "block must not shift");
645 #endif
646 kit->set_map(_map);
647 kit->set_sp(_sp);
648 }
651 //-----------------------------BuildCutout-------------------------------------
652 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
653 : PreserveJVMState(kit)
654 {
655 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
656 SafePointNode* outer_map = _map; // preserved map is caller's
657 SafePointNode* inner_map = kit->map();
658 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
659 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
660 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
661 }
662 BuildCutout::~BuildCutout() {
663 GraphKit* kit = _kit;
664 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
665 }
667 //---------------------------PreserveReexecuteState----------------------------
668 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
669 assert(!kit->stopped(), "must call stopped() before");
670 _kit = kit;
671 _sp = kit->sp();
672 _reexecute = kit->jvms()->_reexecute;
673 }
674 PreserveReexecuteState::~PreserveReexecuteState() {
675 if (_kit->stopped()) return;
676 _kit->jvms()->_reexecute = _reexecute;
677 _kit->set_sp(_sp);
678 }
680 //------------------------------clone_map--------------------------------------
681 // Implementation of PreserveJVMState
682 //
683 // Only clone_map(...) here. If this function is only used in the
684 // PreserveJVMState class we may want to get rid of this extra
685 // function eventually and do it all there.
687 SafePointNode* GraphKit::clone_map() {
688 if (map() == NULL) return NULL;
690 // Clone the memory edge first
691 Node* mem = MergeMemNode::make(C, map()->memory());
692 gvn().set_type_bottom(mem);
694 SafePointNode *clonemap = (SafePointNode*)map()->clone();
695 JVMState* jvms = this->jvms();
696 JVMState* clonejvms = jvms->clone_shallow(C);
697 clonemap->set_memory(mem);
698 clonemap->set_jvms(clonejvms);
699 clonejvms->set_map(clonemap);
700 record_for_igvn(clonemap);
701 gvn().set_type_bottom(clonemap);
702 return clonemap;
703 }
706 //-----------------------------set_map_clone-----------------------------------
707 void GraphKit::set_map_clone(SafePointNode* m) {
708 _map = m;
709 _map = clone_map();
710 _map->set_next_exception(NULL);
711 debug_only(verify_map());
712 }
715 //----------------------------kill_dead_locals---------------------------------
716 // Detect any locals which are known to be dead, and force them to top.
717 void GraphKit::kill_dead_locals() {
718 // Consult the liveness information for the locals. If any
719 // of them are unused, then they can be replaced by top(). This
720 // should help register allocation time and cut down on the size
721 // of the deoptimization information.
723 // This call is made from many of the bytecode handling
724 // subroutines called from the Big Switch in do_one_bytecode.
725 // Every bytecode which might include a slow path is responsible
726 // for killing its dead locals. The more consistent we
727 // are about killing deads, the fewer useless phis will be
728 // constructed for them at various merge points.
730 // bci can be -1 (InvocationEntryBci). We return the entry
731 // liveness for the method.
733 if (method() == NULL || method()->code_size() == 0) {
734 // We are building a graph for a call to a native method.
735 // All locals are live.
736 return;
737 }
739 ResourceMark rm;
741 // Consult the liveness information for the locals. If any
742 // of them are unused, then they can be replaced by top(). This
743 // should help register allocation time and cut down on the size
744 // of the deoptimization information.
745 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
747 int len = (int)live_locals.size();
748 assert(len <= jvms()->loc_size(), "too many live locals");
749 for (int local = 0; local < len; local++) {
750 if (!live_locals.at(local)) {
751 set_local(local, top());
752 }
753 }
754 }
756 #ifdef ASSERT
757 //-------------------------dead_locals_are_killed------------------------------
758 // Return true if all dead locals are set to top in the map.
759 // Used to assert "clean" debug info at various points.
760 bool GraphKit::dead_locals_are_killed() {
761 if (method() == NULL || method()->code_size() == 0) {
762 // No locals need to be dead, so all is as it should be.
763 return true;
764 }
766 // Make sure somebody called kill_dead_locals upstream.
767 ResourceMark rm;
768 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
769 if (jvms->loc_size() == 0) continue; // no locals to consult
770 SafePointNode* map = jvms->map();
771 ciMethod* method = jvms->method();
772 int bci = jvms->bci();
773 if (jvms == this->jvms()) {
774 bci = this->bci(); // it might not yet be synched
775 }
776 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
777 int len = (int)live_locals.size();
778 if (!live_locals.is_valid() || len == 0)
779 // This method is trivial, or is poisoned by a breakpoint.
780 return true;
781 assert(len == jvms->loc_size(), "live map consistent with locals map");
782 for (int local = 0; local < len; local++) {
783 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
784 if (PrintMiscellaneous && (Verbose || WizardMode)) {
785 tty->print_cr("Zombie local %d: ", local);
786 jvms->dump();
787 }
788 return false;
789 }
790 }
791 }
792 return true;
793 }
795 #endif //ASSERT
797 // Helper function for enforcing certain bytecodes to reexecute if
798 // deoptimization happens
799 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
800 ciMethod* cur_method = jvms->method();
801 int cur_bci = jvms->bci();
802 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
803 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
804 return Interpreter::bytecode_should_reexecute(code) ||
805 is_anewarray && code == Bytecodes::_multianewarray;
806 // Reexecute _multianewarray bytecode which was replaced with
807 // sequence of [a]newarray. See Parse::do_multianewarray().
808 //
809 // Note: interpreter should not have it set since this optimization
810 // is limited by dimensions and guarded by flag so in some cases
811 // multianewarray() runtime calls will be generated and
812 // the bytecode should not be reexecutes (stack will not be reset).
813 } else
814 return false;
815 }
817 // Helper function for adding JVMState and debug information to node
818 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
819 // Add the safepoint edges to the call (or other safepoint).
821 // Make sure dead locals are set to top. This
822 // should help register allocation time and cut down on the size
823 // of the deoptimization information.
824 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
826 // Walk the inline list to fill in the correct set of JVMState's
827 // Also fill in the associated edges for each JVMState.
829 JVMState* youngest_jvms = sync_jvms();
831 // If we are guaranteed to throw, we can prune everything but the
832 // input to the current bytecode.
833 bool can_prune_locals = false;
834 uint stack_slots_not_pruned = 0;
835 int inputs = 0, depth = 0;
836 if (must_throw) {
837 assert(method() == youngest_jvms->method(), "sanity");
838 if (compute_stack_effects(inputs, depth)) {
839 can_prune_locals = true;
840 stack_slots_not_pruned = inputs;
841 }
842 }
844 if (env()->jvmti_can_access_local_variables()) {
845 // At any safepoint, this method can get breakpointed, which would
846 // then require an immediate deoptimization.
847 can_prune_locals = false; // do not prune locals
848 stack_slots_not_pruned = 0;
849 }
851 // do not scribble on the input jvms
852 JVMState* out_jvms = youngest_jvms->clone_deep(C);
853 call->set_jvms(out_jvms); // Start jvms list for call node
855 // For a known set of bytecodes, the interpreter should reexecute them if
856 // deoptimization happens. We set the reexecute state for them here
857 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
858 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
859 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
860 }
862 // Presize the call:
863 debug_only(uint non_debug_edges = call->req());
864 call->add_req_batch(top(), youngest_jvms->debug_depth());
865 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
867 // Set up edges so that the call looks like this:
868 // Call [state:] ctl io mem fptr retadr
869 // [parms:] parm0 ... parmN
870 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
871 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
872 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
873 // Note that caller debug info precedes callee debug info.
875 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
876 uint debug_ptr = call->req();
878 // Loop over the map input edges associated with jvms, add them
879 // to the call node, & reset all offsets to match call node array.
880 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
881 uint debug_end = debug_ptr;
882 uint debug_start = debug_ptr - in_jvms->debug_size();
883 debug_ptr = debug_start; // back up the ptr
885 uint p = debug_start; // walks forward in [debug_start, debug_end)
886 uint j, k, l;
887 SafePointNode* in_map = in_jvms->map();
888 out_jvms->set_map(call);
890 if (can_prune_locals) {
891 assert(in_jvms->method() == out_jvms->method(), "sanity");
892 // If the current throw can reach an exception handler in this JVMS,
893 // then we must keep everything live that can reach that handler.
894 // As a quick and dirty approximation, we look for any handlers at all.
895 if (in_jvms->method()->has_exception_handlers()) {
896 can_prune_locals = false;
897 }
898 }
900 // Add the Locals
901 k = in_jvms->locoff();
902 l = in_jvms->loc_size();
903 out_jvms->set_locoff(p);
904 if (!can_prune_locals) {
905 for (j = 0; j < l; j++)
906 call->set_req(p++, in_map->in(k+j));
907 } else {
908 p += l; // already set to top above by add_req_batch
909 }
911 // Add the Expression Stack
912 k = in_jvms->stkoff();
913 l = in_jvms->sp();
914 out_jvms->set_stkoff(p);
915 if (!can_prune_locals) {
916 for (j = 0; j < l; j++)
917 call->set_req(p++, in_map->in(k+j));
918 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
919 // Divide stack into {S0,...,S1}, where S0 is set to top.
920 uint s1 = stack_slots_not_pruned;
921 stack_slots_not_pruned = 0; // for next iteration
922 if (s1 > l) s1 = l;
923 uint s0 = l - s1;
924 p += s0; // skip the tops preinstalled by add_req_batch
925 for (j = s0; j < l; j++)
926 call->set_req(p++, in_map->in(k+j));
927 } else {
928 p += l; // already set to top above by add_req_batch
929 }
931 // Add the Monitors
932 k = in_jvms->monoff();
933 l = in_jvms->mon_size();
934 out_jvms->set_monoff(p);
935 for (j = 0; j < l; j++)
936 call->set_req(p++, in_map->in(k+j));
938 // Copy any scalar object fields.
939 k = in_jvms->scloff();
940 l = in_jvms->scl_size();
941 out_jvms->set_scloff(p);
942 for (j = 0; j < l; j++)
943 call->set_req(p++, in_map->in(k+j));
945 // Finish the new jvms.
946 out_jvms->set_endoff(p);
948 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
949 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
950 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
951 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
952 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
953 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
955 // Update the two tail pointers in parallel.
956 out_jvms = out_jvms->caller();
957 in_jvms = in_jvms->caller();
958 }
960 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
962 // Test the correctness of JVMState::debug_xxx accessors:
963 assert(call->jvms()->debug_start() == non_debug_edges, "");
964 assert(call->jvms()->debug_end() == call->req(), "");
965 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
966 }
968 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
969 Bytecodes::Code code = java_bc();
970 if (code == Bytecodes::_wide) {
971 code = method()->java_code_at_bci(bci() + 1);
972 }
974 BasicType rtype = T_ILLEGAL;
975 int rsize = 0;
977 if (code != Bytecodes::_illegal) {
978 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
979 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
980 if (rtype < T_CONFLICT)
981 rsize = type2size[rtype];
982 }
984 switch (code) {
985 case Bytecodes::_illegal:
986 return false;
988 case Bytecodes::_ldc:
989 case Bytecodes::_ldc_w:
990 case Bytecodes::_ldc2_w:
991 inputs = 0;
992 break;
994 case Bytecodes::_dup: inputs = 1; break;
995 case Bytecodes::_dup_x1: inputs = 2; break;
996 case Bytecodes::_dup_x2: inputs = 3; break;
997 case Bytecodes::_dup2: inputs = 2; break;
998 case Bytecodes::_dup2_x1: inputs = 3; break;
999 case Bytecodes::_dup2_x2: inputs = 4; break;
1000 case Bytecodes::_swap: inputs = 2; break;
1001 case Bytecodes::_arraylength: inputs = 1; break;
1003 case Bytecodes::_getstatic:
1004 case Bytecodes::_putstatic:
1005 case Bytecodes::_getfield:
1006 case Bytecodes::_putfield:
1007 {
1008 bool is_get = (depth >= 0), is_static = (depth & 1);
1009 bool ignore;
1010 ciBytecodeStream iter(method());
1011 iter.reset_to_bci(bci());
1012 iter.next();
1013 ciField* field = iter.get_field(ignore);
1014 int size = field->type()->size();
1015 inputs = (is_static ? 0 : 1);
1016 if (is_get) {
1017 depth = size - inputs;
1018 } else {
1019 inputs += size; // putxxx pops the value from the stack
1020 depth = - inputs;
1021 }
1022 }
1023 break;
1025 case Bytecodes::_invokevirtual:
1026 case Bytecodes::_invokespecial:
1027 case Bytecodes::_invokestatic:
1028 case Bytecodes::_invokedynamic:
1029 case Bytecodes::_invokeinterface:
1030 {
1031 bool ignore;
1032 ciBytecodeStream iter(method());
1033 iter.reset_to_bci(bci());
1034 iter.next();
1035 ciMethod* method = iter.get_method(ignore);
1036 // (Do not use ciMethod::arg_size(), because
1037 // it might be an unloaded method, which doesn't
1038 // know whether it is static or not.)
1039 inputs = method->invoke_arg_size(code);
1040 int size = method->return_type()->size();
1041 depth = size - inputs;
1042 }
1043 break;
1045 case Bytecodes::_multianewarray:
1046 {
1047 ciBytecodeStream iter(method());
1048 iter.reset_to_bci(bci());
1049 iter.next();
1050 inputs = iter.get_dimensions();
1051 assert(rsize == 1, "");
1052 depth = rsize - inputs;
1053 }
1054 break;
1056 case Bytecodes::_ireturn:
1057 case Bytecodes::_lreturn:
1058 case Bytecodes::_freturn:
1059 case Bytecodes::_dreturn:
1060 case Bytecodes::_areturn:
1061 assert(rsize = -depth, "");
1062 inputs = rsize;
1063 break;
1065 case Bytecodes::_jsr:
1066 case Bytecodes::_jsr_w:
1067 inputs = 0;
1068 depth = 1; // S.B. depth=1, not zero
1069 break;
1071 default:
1072 // bytecode produces a typed result
1073 inputs = rsize - depth;
1074 assert(inputs >= 0, "");
1075 break;
1076 }
1078 #ifdef ASSERT
1079 // spot check
1080 int outputs = depth + inputs;
1081 assert(outputs >= 0, "sanity");
1082 switch (code) {
1083 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1084 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1085 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1086 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1087 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1088 }
1089 #endif //ASSERT
1091 return true;
1092 }
1096 //------------------------------basic_plus_adr---------------------------------
1097 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1098 // short-circuit a common case
1099 if (offset == intcon(0)) return ptr;
1100 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1101 }
1103 Node* GraphKit::ConvI2L(Node* offset) {
1104 // short-circuit a common case
1105 jint offset_con = find_int_con(offset, Type::OffsetBot);
1106 if (offset_con != Type::OffsetBot) {
1107 return longcon((long) offset_con);
1108 }
1109 return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1110 }
1111 Node* GraphKit::ConvL2I(Node* offset) {
1112 // short-circuit a common case
1113 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1114 if (offset_con != (jlong)Type::OffsetBot) {
1115 return intcon((int) offset_con);
1116 }
1117 return _gvn.transform( new (C, 2) ConvL2INode(offset));
1118 }
1120 //-------------------------load_object_klass-----------------------------------
1121 Node* GraphKit::load_object_klass(Node* obj) {
1122 // Special-case a fresh allocation to avoid building nodes:
1123 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1124 if (akls != NULL) return akls;
1125 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1126 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1127 }
1129 //-------------------------load_array_length-----------------------------------
1130 Node* GraphKit::load_array_length(Node* array) {
1131 // Special-case a fresh allocation to avoid building nodes:
1132 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1133 Node *alen;
1134 if (alloc == NULL) {
1135 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1136 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1137 } else {
1138 alen = alloc->Ideal_length();
1139 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1140 if (ccast != alen) {
1141 alen = _gvn.transform(ccast);
1142 }
1143 }
1144 return alen;
1145 }
1147 //------------------------------do_null_check----------------------------------
1148 // Helper function to do a NULL pointer check. Returned value is
1149 // the incoming address with NULL casted away. You are allowed to use the
1150 // not-null value only if you are control dependent on the test.
1151 extern int explicit_null_checks_inserted,
1152 explicit_null_checks_elided;
1153 Node* GraphKit::null_check_common(Node* value, BasicType type,
1154 // optional arguments for variations:
1155 bool assert_null,
1156 Node* *null_control) {
1157 assert(!assert_null || null_control == NULL, "not both at once");
1158 if (stopped()) return top();
1159 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1160 // For some performance testing, we may wish to suppress null checking.
1161 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1162 return value;
1163 }
1164 explicit_null_checks_inserted++;
1166 // Construct NULL check
1167 Node *chk = NULL;
1168 switch(type) {
1169 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1170 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1171 case T_ARRAY : // fall through
1172 type = T_OBJECT; // simplify further tests
1173 case T_OBJECT : {
1174 const Type *t = _gvn.type( value );
1176 const TypeOopPtr* tp = t->isa_oopptr();
1177 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1178 // Only for do_null_check, not any of its siblings:
1179 && !assert_null && null_control == NULL) {
1180 // Usually, any field access or invocation on an unloaded oop type
1181 // will simply fail to link, since the statically linked class is
1182 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1183 // the static class is loaded but the sharper oop type is not.
1184 // Rather than checking for this obscure case in lots of places,
1185 // we simply observe that a null check on an unloaded class
1186 // will always be followed by a nonsense operation, so we
1187 // can just issue the uncommon trap here.
1188 // Our access to the unloaded class will only be correct
1189 // after it has been loaded and initialized, which requires
1190 // a trip through the interpreter.
1191 #ifndef PRODUCT
1192 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1193 #endif
1194 uncommon_trap(Deoptimization::Reason_unloaded,
1195 Deoptimization::Action_reinterpret,
1196 tp->klass(), "!loaded");
1197 return top();
1198 }
1200 if (assert_null) {
1201 // See if the type is contained in NULL_PTR.
1202 // If so, then the value is already null.
1203 if (t->higher_equal(TypePtr::NULL_PTR)) {
1204 explicit_null_checks_elided++;
1205 return value; // Elided null assert quickly!
1206 }
1207 } else {
1208 // See if mixing in the NULL pointer changes type.
1209 // If so, then the NULL pointer was not allowed in the original
1210 // type. In other words, "value" was not-null.
1211 if (t->meet(TypePtr::NULL_PTR) != t) {
1212 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1213 explicit_null_checks_elided++;
1214 return value; // Elided null check quickly!
1215 }
1216 }
1217 chk = new (C, 3) CmpPNode( value, null() );
1218 break;
1219 }
1221 default : ShouldNotReachHere();
1222 }
1223 assert(chk != NULL, "sanity check");
1224 chk = _gvn.transform(chk);
1226 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1227 BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1228 Node *tst = _gvn.transform( btst );
1230 //-----------
1231 // if peephole optimizations occurred, a prior test existed.
1232 // If a prior test existed, maybe it dominates as we can avoid this test.
1233 if (tst != btst && type == T_OBJECT) {
1234 // At this point we want to scan up the CFG to see if we can
1235 // find an identical test (and so avoid this test altogether).
1236 Node *cfg = control();
1237 int depth = 0;
1238 while( depth < 16 ) { // Limit search depth for speed
1239 if( cfg->Opcode() == Op_IfTrue &&
1240 cfg->in(0)->in(1) == tst ) {
1241 // Found prior test. Use "cast_not_null" to construct an identical
1242 // CastPP (and hence hash to) as already exists for the prior test.
1243 // Return that casted value.
1244 if (assert_null) {
1245 replace_in_map(value, null());
1246 return null(); // do not issue the redundant test
1247 }
1248 Node *oldcontrol = control();
1249 set_control(cfg);
1250 Node *res = cast_not_null(value);
1251 set_control(oldcontrol);
1252 explicit_null_checks_elided++;
1253 return res;
1254 }
1255 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1256 if (cfg == NULL) break; // Quit at region nodes
1257 depth++;
1258 }
1259 }
1261 //-----------
1262 // Branch to failure if null
1263 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1264 Deoptimization::DeoptReason reason;
1265 if (assert_null)
1266 reason = Deoptimization::Reason_null_assert;
1267 else if (type == T_OBJECT)
1268 reason = Deoptimization::Reason_null_check;
1269 else
1270 reason = Deoptimization::Reason_div0_check;
1272 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1273 // ciMethodData::has_trap_at will return a conservative -1 if any
1274 // must-be-null assertion has failed. This could cause performance
1275 // problems for a method after its first do_null_assert failure.
1276 // Consider using 'Reason_class_check' instead?
1278 // To cause an implicit null check, we set the not-null probability
1279 // to the maximum (PROB_MAX). For an explicit check the probability
1280 // is set to a smaller value.
1281 if (null_control != NULL || too_many_traps(reason)) {
1282 // probability is less likely
1283 ok_prob = PROB_LIKELY_MAG(3);
1284 } else if (!assert_null &&
1285 (ImplicitNullCheckThreshold > 0) &&
1286 method() != NULL &&
1287 (method()->method_data()->trap_count(reason)
1288 >= (uint)ImplicitNullCheckThreshold)) {
1289 ok_prob = PROB_LIKELY_MAG(3);
1290 }
1292 if (null_control != NULL) {
1293 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1294 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1295 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
1296 if (null_true == top())
1297 explicit_null_checks_elided++;
1298 (*null_control) = null_true;
1299 } else {
1300 BuildCutout unless(this, tst, ok_prob);
1301 // Check for optimizer eliding test at parse time
1302 if (stopped()) {
1303 // Failure not possible; do not bother making uncommon trap.
1304 explicit_null_checks_elided++;
1305 } else if (assert_null) {
1306 uncommon_trap(reason,
1307 Deoptimization::Action_make_not_entrant,
1308 NULL, "assert_null");
1309 } else {
1310 replace_in_map(value, zerocon(type));
1311 builtin_throw(reason);
1312 }
1313 }
1315 // Must throw exception, fall-thru not possible?
1316 if (stopped()) {
1317 return top(); // No result
1318 }
1320 if (assert_null) {
1321 // Cast obj to null on this path.
1322 replace_in_map(value, zerocon(type));
1323 return zerocon(type);
1324 }
1326 // Cast obj to not-null on this path, if there is no null_control.
1327 // (If there is a null_control, a non-null value may come back to haunt us.)
1328 if (type == T_OBJECT) {
1329 Node* cast = cast_not_null(value, false);
1330 if (null_control == NULL || (*null_control) == top())
1331 replace_in_map(value, cast);
1332 value = cast;
1333 }
1335 return value;
1336 }
1339 //------------------------------cast_not_null----------------------------------
1340 // Cast obj to not-null on this path
1341 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1342 const Type *t = _gvn.type(obj);
1343 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1344 // Object is already not-null?
1345 if( t == t_not_null ) return obj;
1347 Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1348 cast->init_req(0, control());
1349 cast = _gvn.transform( cast );
1351 // Scan for instances of 'obj' in the current JVM mapping.
1352 // These instances are known to be not-null after the test.
1353 if (do_replace_in_map)
1354 replace_in_map(obj, cast);
1356 return cast; // Return casted value
1357 }
1360 //--------------------------replace_in_map-------------------------------------
1361 void GraphKit::replace_in_map(Node* old, Node* neww) {
1362 this->map()->replace_edge(old, neww);
1364 // Note: This operation potentially replaces any edge
1365 // on the map. This includes locals, stack, and monitors
1366 // of the current (innermost) JVM state.
1368 // We can consider replacing in caller maps.
1369 // The idea would be that an inlined function's null checks
1370 // can be shared with the entire inlining tree.
1371 // The expense of doing this is that the PreserveJVMState class
1372 // would have to preserve caller states too, with a deep copy.
1373 }
1377 //=============================================================================
1378 //--------------------------------memory---------------------------------------
1379 Node* GraphKit::memory(uint alias_idx) {
1380 MergeMemNode* mem = merged_memory();
1381 Node* p = mem->memory_at(alias_idx);
1382 _gvn.set_type(p, Type::MEMORY); // must be mapped
1383 return p;
1384 }
1386 //-----------------------------reset_memory------------------------------------
1387 Node* GraphKit::reset_memory() {
1388 Node* mem = map()->memory();
1389 // do not use this node for any more parsing!
1390 debug_only( map()->set_memory((Node*)NULL) );
1391 return _gvn.transform( mem );
1392 }
1394 //------------------------------set_all_memory---------------------------------
1395 void GraphKit::set_all_memory(Node* newmem) {
1396 Node* mergemem = MergeMemNode::make(C, newmem);
1397 gvn().set_type_bottom(mergemem);
1398 map()->set_memory(mergemem);
1399 }
1401 //------------------------------set_all_memory_call----------------------------
1402 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1403 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1404 set_all_memory(newmem);
1405 }
1407 //=============================================================================
1408 //
1409 // parser factory methods for MemNodes
1410 //
1411 // These are layered on top of the factory methods in LoadNode and StoreNode,
1412 // and integrate with the parser's memory state and _gvn engine.
1413 //
1415 // factory methods in "int adr_idx"
1416 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1417 int adr_idx,
1418 bool require_atomic_access) {
1419 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1420 const TypePtr* adr_type = NULL; // debug-mode-only argument
1421 debug_only(adr_type = C->get_adr_type(adr_idx));
1422 Node* mem = memory(adr_idx);
1423 Node* ld;
1424 if (require_atomic_access && bt == T_LONG) {
1425 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1426 } else {
1427 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1428 }
1429 return _gvn.transform(ld);
1430 }
1432 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1433 int adr_idx,
1434 bool require_atomic_access) {
1435 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1436 const TypePtr* adr_type = NULL;
1437 debug_only(adr_type = C->get_adr_type(adr_idx));
1438 Node *mem = memory(adr_idx);
1439 Node* st;
1440 if (require_atomic_access && bt == T_LONG) {
1441 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1442 } else {
1443 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1444 }
1445 st = _gvn.transform(st);
1446 set_memory(st, adr_idx);
1447 // Back-to-back stores can only remove intermediate store with DU info
1448 // so push on worklist for optimizer.
1449 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1450 record_for_igvn(st);
1452 return st;
1453 }
1456 void GraphKit::pre_barrier(Node* ctl,
1457 Node* obj,
1458 Node* adr,
1459 uint adr_idx,
1460 Node* val,
1461 const TypeOopPtr* val_type,
1462 BasicType bt) {
1463 BarrierSet* bs = Universe::heap()->barrier_set();
1464 set_control(ctl);
1465 switch (bs->kind()) {
1466 case BarrierSet::G1SATBCT:
1467 case BarrierSet::G1SATBCTLogging:
1468 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt);
1469 break;
1471 case BarrierSet::CardTableModRef:
1472 case BarrierSet::CardTableExtension:
1473 case BarrierSet::ModRef:
1474 break;
1476 case BarrierSet::Other:
1477 default :
1478 ShouldNotReachHere();
1480 }
1481 }
1483 void GraphKit::post_barrier(Node* ctl,
1484 Node* store,
1485 Node* obj,
1486 Node* adr,
1487 uint adr_idx,
1488 Node* val,
1489 BasicType bt,
1490 bool use_precise) {
1491 BarrierSet* bs = Universe::heap()->barrier_set();
1492 set_control(ctl);
1493 switch (bs->kind()) {
1494 case BarrierSet::G1SATBCT:
1495 case BarrierSet::G1SATBCTLogging:
1496 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1497 break;
1499 case BarrierSet::CardTableModRef:
1500 case BarrierSet::CardTableExtension:
1501 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1502 break;
1504 case BarrierSet::ModRef:
1505 break;
1507 case BarrierSet::Other:
1508 default :
1509 ShouldNotReachHere();
1511 }
1512 }
1514 Node* GraphKit::store_oop(Node* ctl,
1515 Node* obj,
1516 Node* adr,
1517 const TypePtr* adr_type,
1518 Node* val,
1519 const TypeOopPtr* val_type,
1520 BasicType bt,
1521 bool use_precise) {
1523 set_control(ctl);
1524 if (stopped()) return top(); // Dead path ?
1526 assert(bt == T_OBJECT, "sanity");
1527 assert(val != NULL, "not dead path");
1528 uint adr_idx = C->get_alias_index(adr_type);
1529 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1531 pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt);
1532 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1533 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1534 return store;
1535 }
1537 // Could be an array or object we don't know at compile time (unsafe ref.)
1538 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1539 Node* obj, // containing obj
1540 Node* adr, // actual adress to store val at
1541 const TypePtr* adr_type,
1542 Node* val,
1543 BasicType bt) {
1544 Compile::AliasType* at = C->alias_type(adr_type);
1545 const TypeOopPtr* val_type = NULL;
1546 if (adr_type->isa_instptr()) {
1547 if (at->field() != NULL) {
1548 // known field. This code is a copy of the do_put_xxx logic.
1549 ciField* field = at->field();
1550 if (!field->type()->is_loaded()) {
1551 val_type = TypeInstPtr::BOTTOM;
1552 } else {
1553 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1554 }
1555 }
1556 } else if (adr_type->isa_aryptr()) {
1557 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1558 }
1559 if (val_type == NULL) {
1560 val_type = TypeInstPtr::BOTTOM;
1561 }
1562 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1563 }
1566 //-------------------------array_element_address-------------------------
1567 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1568 const TypeInt* sizetype) {
1569 uint shift = exact_log2(type2aelembytes(elembt));
1570 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1572 // short-circuit a common case (saves lots of confusing waste motion)
1573 jint idx_con = find_int_con(idx, -1);
1574 if (idx_con >= 0) {
1575 intptr_t offset = header + ((intptr_t)idx_con << shift);
1576 return basic_plus_adr(ary, offset);
1577 }
1579 // must be correct type for alignment purposes
1580 Node* base = basic_plus_adr(ary, header);
1581 #ifdef _LP64
1582 // The scaled index operand to AddP must be a clean 64-bit value.
1583 // Java allows a 32-bit int to be incremented to a negative
1584 // value, which appears in a 64-bit register as a large
1585 // positive number. Using that large positive number as an
1586 // operand in pointer arithmetic has bad consequences.
1587 // On the other hand, 32-bit overflow is rare, and the possibility
1588 // can often be excluded, if we annotate the ConvI2L node with
1589 // a type assertion that its value is known to be a small positive
1590 // number. (The prior range check has ensured this.)
1591 // This assertion is used by ConvI2LNode::Ideal.
1592 int index_max = max_jint - 1; // array size is max_jint, index is one less
1593 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1594 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1595 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1596 #endif
1597 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1598 return basic_plus_adr(ary, base, scale);
1599 }
1601 //-------------------------load_array_element-------------------------
1602 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1603 const Type* elemtype = arytype->elem();
1604 BasicType elembt = elemtype->array_element_basic_type();
1605 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1606 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1607 return ld;
1608 }
1610 //-------------------------set_arguments_for_java_call-------------------------
1611 // Arguments (pre-popped from the stack) are taken from the JVMS.
1612 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1613 // Add the call arguments:
1614 uint nargs = call->method()->arg_size();
1615 for (uint i = 0; i < nargs; i++) {
1616 Node* arg = argument(i);
1617 call->init_req(i + TypeFunc::Parms, arg);
1618 }
1619 }
1621 //---------------------------set_edges_for_java_call---------------------------
1622 // Connect a newly created call into the current JVMS.
1623 // A return value node (if any) is returned from set_edges_for_java_call.
1624 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1626 // Add the predefined inputs:
1627 call->init_req( TypeFunc::Control, control() );
1628 call->init_req( TypeFunc::I_O , i_o() );
1629 call->init_req( TypeFunc::Memory , reset_memory() );
1630 call->init_req( TypeFunc::FramePtr, frameptr() );
1631 call->init_req( TypeFunc::ReturnAdr, top() );
1633 add_safepoint_edges(call, must_throw);
1635 Node* xcall = _gvn.transform(call);
1637 if (xcall == top()) {
1638 set_control(top());
1639 return;
1640 }
1641 assert(xcall == call, "call identity is stable");
1643 // Re-use the current map to produce the result.
1645 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1646 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1647 set_all_memory_call(xcall, separate_io_proj);
1649 //return xcall; // no need, caller already has it
1650 }
1652 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1653 if (stopped()) return top(); // maybe the call folded up?
1655 // Capture the return value, if any.
1656 Node* ret;
1657 if (call->method() == NULL ||
1658 call->method()->return_type()->basic_type() == T_VOID)
1659 ret = top();
1660 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1662 // Note: Since any out-of-line call can produce an exception,
1663 // we always insert an I_O projection from the call into the result.
1665 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1667 if (separate_io_proj) {
1668 // The caller requested separate projections be used by the fall
1669 // through and exceptional paths, so replace the projections for
1670 // the fall through path.
1671 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) ));
1672 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ));
1673 }
1674 return ret;
1675 }
1677 //--------------------set_predefined_input_for_runtime_call--------------------
1678 // Reading and setting the memory state is way conservative here.
1679 // The real problem is that I am not doing real Type analysis on memory,
1680 // so I cannot distinguish card mark stores from other stores. Across a GC
1681 // point the Store Barrier and the card mark memory has to agree. I cannot
1682 // have a card mark store and its barrier split across the GC point from
1683 // either above or below. Here I get that to happen by reading ALL of memory.
1684 // A better answer would be to separate out card marks from other memory.
1685 // For now, return the input memory state, so that it can be reused
1686 // after the call, if this call has restricted memory effects.
1687 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1688 // Set fixed predefined input arguments
1689 Node* memory = reset_memory();
1690 call->init_req( TypeFunc::Control, control() );
1691 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1692 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1693 call->init_req( TypeFunc::FramePtr, frameptr() );
1694 call->init_req( TypeFunc::ReturnAdr, top() );
1695 return memory;
1696 }
1698 //-------------------set_predefined_output_for_runtime_call--------------------
1699 // Set control and memory (not i_o) from the call.
1700 // If keep_mem is not NULL, use it for the output state,
1701 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1702 // If hook_mem is NULL, this call produces no memory effects at all.
1703 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1704 // then only that memory slice is taken from the call.
1705 // In the last case, we must put an appropriate memory barrier before
1706 // the call, so as to create the correct anti-dependencies on loads
1707 // preceding the call.
1708 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1709 Node* keep_mem,
1710 const TypePtr* hook_mem) {
1711 // no i/o
1712 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1713 if (keep_mem) {
1714 // First clone the existing memory state
1715 set_all_memory(keep_mem);
1716 if (hook_mem != NULL) {
1717 // Make memory for the call
1718 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1719 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1720 // We also use hook_mem to extract specific effects from arraycopy stubs.
1721 set_memory(mem, hook_mem);
1722 }
1723 // ...else the call has NO memory effects.
1725 // Make sure the call advertises its memory effects precisely.
1726 // This lets us build accurate anti-dependences in gcm.cpp.
1727 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1728 "call node must be constructed correctly");
1729 } else {
1730 assert(hook_mem == NULL, "");
1731 // This is not a "slow path" call; all memory comes from the call.
1732 set_all_memory_call(call);
1733 }
1734 }
1737 // Replace the call with the current state of the kit.
1738 void GraphKit::replace_call(CallNode* call, Node* result) {
1739 JVMState* ejvms = NULL;
1740 if (has_exceptions()) {
1741 ejvms = transfer_exceptions_into_jvms();
1742 }
1744 SafePointNode* final_state = stop();
1746 // Find all the needed outputs of this call
1747 CallProjections callprojs;
1748 call->extract_projections(&callprojs, true);
1750 // Replace all the old call edges with the edges from the inlining result
1751 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control));
1752 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory));
1753 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O));
1754 Node* final_mem = final_state->in(TypeFunc::Memory);
1756 // Replace the result with the new result if it exists and is used
1757 if (callprojs.resproj != NULL && result != NULL) {
1758 C->gvn_replace_by(callprojs.resproj, result);
1759 }
1761 if (ejvms == NULL) {
1762 // No exception edges to simply kill off those paths
1763 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1764 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1765 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1767 // Replace the old exception object with top
1768 if (callprojs.exobj != NULL) {
1769 C->gvn_replace_by(callprojs.exobj, C->top());
1770 }
1771 } else {
1772 GraphKit ekit(ejvms);
1774 // Load my combined exception state into the kit, with all phis transformed:
1775 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1777 Node* ex_oop = ekit.use_exception_state(ex_map);
1779 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1780 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1781 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1783 // Replace the old exception object with the newly created one
1784 if (callprojs.exobj != NULL) {
1785 C->gvn_replace_by(callprojs.exobj, ex_oop);
1786 }
1787 }
1789 // Disconnect the call from the graph
1790 call->disconnect_inputs(NULL);
1791 C->gvn_replace_by(call, C->top());
1793 // Clean up any MergeMems that feed other MergeMems since the
1794 // optimizer doesn't like that.
1795 if (final_mem->is_MergeMem()) {
1796 Node_List wl;
1797 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1798 Node* m = i.get();
1799 if (m->is_MergeMem() && !wl.contains(m)) {
1800 wl.push(m);
1801 }
1802 }
1803 while (wl.size() > 0) {
1804 _gvn.transform(wl.pop());
1805 }
1806 }
1807 }
1810 //------------------------------increment_counter------------------------------
1811 // for statistics: increment a VM counter by 1
1813 void GraphKit::increment_counter(address counter_addr) {
1814 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1815 increment_counter(adr1);
1816 }
1818 void GraphKit::increment_counter(Node* counter_addr) {
1819 int adr_type = Compile::AliasIdxRaw;
1820 Node* ctrl = control();
1821 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type);
1822 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1823 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type );
1824 }
1827 //------------------------------uncommon_trap----------------------------------
1828 // Bail out to the interpreter in mid-method. Implemented by calling the
1829 // uncommon_trap blob. This helper function inserts a runtime call with the
1830 // right debug info.
1831 void GraphKit::uncommon_trap(int trap_request,
1832 ciKlass* klass, const char* comment,
1833 bool must_throw,
1834 bool keep_exact_action) {
1835 if (failing()) stop();
1836 if (stopped()) return; // trap reachable?
1838 // Note: If ProfileTraps is true, and if a deopt. actually
1839 // occurs here, the runtime will make sure an MDO exists. There is
1840 // no need to call method()->ensure_method_data() at this point.
1842 #ifdef ASSERT
1843 if (!must_throw) {
1844 // Make sure the stack has at least enough depth to execute
1845 // the current bytecode.
1846 int inputs, ignore;
1847 if (compute_stack_effects(inputs, ignore)) {
1848 assert(sp() >= inputs, "must have enough JVMS stack to execute");
1849 // It is a frequent error in library_call.cpp to issue an
1850 // uncommon trap with the _sp value already popped.
1851 }
1852 }
1853 #endif
1855 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1856 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1858 switch (action) {
1859 case Deoptimization::Action_maybe_recompile:
1860 case Deoptimization::Action_reinterpret:
1861 // Temporary fix for 6529811 to allow virtual calls to be sure they
1862 // get the chance to go from mono->bi->mega
1863 if (!keep_exact_action &&
1864 Deoptimization::trap_request_index(trap_request) < 0 &&
1865 too_many_recompiles(reason)) {
1866 // This BCI is causing too many recompilations.
1867 action = Deoptimization::Action_none;
1868 trap_request = Deoptimization::make_trap_request(reason, action);
1869 } else {
1870 C->set_trap_can_recompile(true);
1871 }
1872 break;
1873 case Deoptimization::Action_make_not_entrant:
1874 C->set_trap_can_recompile(true);
1875 break;
1876 #ifdef ASSERT
1877 case Deoptimization::Action_none:
1878 case Deoptimization::Action_make_not_compilable:
1879 break;
1880 default:
1881 assert(false, "bad action");
1882 #endif
1883 }
1885 if (TraceOptoParse) {
1886 char buf[100];
1887 tty->print_cr("Uncommon trap %s at bci:%d",
1888 Deoptimization::format_trap_request(buf, sizeof(buf),
1889 trap_request), bci());
1890 }
1892 CompileLog* log = C->log();
1893 if (log != NULL) {
1894 int kid = (klass == NULL)? -1: log->identify(klass);
1895 log->begin_elem("uncommon_trap bci='%d'", bci());
1896 char buf[100];
1897 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1898 trap_request));
1899 if (kid >= 0) log->print(" klass='%d'", kid);
1900 if (comment != NULL) log->print(" comment='%s'", comment);
1901 log->end_elem();
1902 }
1904 // Make sure any guarding test views this path as very unlikely
1905 Node *i0 = control()->in(0);
1906 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1907 IfNode *iff = i0->as_If();
1908 float f = iff->_prob; // Get prob
1909 if (control()->Opcode() == Op_IfTrue) {
1910 if (f > PROB_UNLIKELY_MAG(4))
1911 iff->_prob = PROB_MIN;
1912 } else {
1913 if (f < PROB_LIKELY_MAG(4))
1914 iff->_prob = PROB_MAX;
1915 }
1916 }
1918 // Clear out dead values from the debug info.
1919 kill_dead_locals();
1921 // Now insert the uncommon trap subroutine call
1922 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
1923 const TypePtr* no_memory_effects = NULL;
1924 // Pass the index of the class to be loaded
1925 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1926 (must_throw ? RC_MUST_THROW : 0),
1927 OptoRuntime::uncommon_trap_Type(),
1928 call_addr, "uncommon_trap", no_memory_effects,
1929 intcon(trap_request));
1930 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1931 "must extract request correctly from the graph");
1932 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1934 call->set_req(TypeFunc::ReturnAdr, returnadr());
1935 // The debug info is the only real input to this call.
1937 // Halt-and-catch fire here. The above call should never return!
1938 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1939 _gvn.set_type_bottom(halt);
1940 root()->add_req(halt);
1942 stop_and_kill_map();
1943 }
1946 //--------------------------just_allocated_object------------------------------
1947 // Report the object that was just allocated.
1948 // It must be the case that there are no intervening safepoints.
1949 // We use this to determine if an object is so "fresh" that
1950 // it does not require card marks.
1951 Node* GraphKit::just_allocated_object(Node* current_control) {
1952 if (C->recent_alloc_ctl() == current_control)
1953 return C->recent_alloc_obj();
1954 return NULL;
1955 }
1958 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1959 // (Note: TypeFunc::make has a cache that makes this fast.)
1960 const TypeFunc* tf = TypeFunc::make(dest_method);
1961 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
1962 for (int j = 0; j < nargs; j++) {
1963 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1964 if( targ->basic_type() == T_DOUBLE ) {
1965 // If any parameters are doubles, they must be rounded before
1966 // the call, dstore_rounding does gvn.transform
1967 Node *arg = argument(j);
1968 arg = dstore_rounding(arg);
1969 set_argument(j, arg);
1970 }
1971 }
1972 }
1974 void GraphKit::round_double_result(ciMethod* dest_method) {
1975 // A non-strict method may return a double value which has an extended
1976 // exponent, but this must not be visible in a caller which is 'strict'
1977 // If a strict caller invokes a non-strict callee, round a double result
1979 BasicType result_type = dest_method->return_type()->basic_type();
1980 assert( method() != NULL, "must have caller context");
1981 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
1982 // Destination method's return value is on top of stack
1983 // dstore_rounding() does gvn.transform
1984 Node *result = pop_pair();
1985 result = dstore_rounding(result);
1986 push_pair(result);
1987 }
1988 }
1990 // rounding for strict float precision conformance
1991 Node* GraphKit::precision_rounding(Node* n) {
1992 return UseStrictFP && _method->flags().is_strict()
1993 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
1994 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
1995 : n;
1996 }
1998 // rounding for strict double precision conformance
1999 Node* GraphKit::dprecision_rounding(Node *n) {
2000 return UseStrictFP && _method->flags().is_strict()
2001 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2002 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
2003 : n;
2004 }
2006 // rounding for non-strict double stores
2007 Node* GraphKit::dstore_rounding(Node* n) {
2008 return Matcher::strict_fp_requires_explicit_rounding
2009 && UseSSE <= 1
2010 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
2011 : n;
2012 }
2014 //=============================================================================
2015 // Generate a fast path/slow path idiom. Graph looks like:
2016 // [foo] indicates that 'foo' is a parameter
2017 //
2018 // [in] NULL
2019 // \ /
2020 // CmpP
2021 // Bool ne
2022 // If
2023 // / \
2024 // True False-<2>
2025 // / |
2026 // / cast_not_null
2027 // Load | | ^
2028 // [fast_test] | |
2029 // gvn to opt_test | |
2030 // / \ | <1>
2031 // True False |
2032 // | \\ |
2033 // [slow_call] \[fast_result]
2034 // Ctl Val \ \
2035 // | \ \
2036 // Catch <1> \ \
2037 // / \ ^ \ \
2038 // Ex No_Ex | \ \
2039 // | \ \ | \ <2> \
2040 // ... \ [slow_res] | | \ [null_result]
2041 // \ \--+--+--- | |
2042 // \ | / \ | /
2043 // --------Region Phi
2044 //
2045 //=============================================================================
2046 // Code is structured as a series of driver functions all called 'do_XXX' that
2047 // call a set of helper functions. Helper functions first, then drivers.
2049 //------------------------------null_check_oop---------------------------------
2050 // Null check oop. Set null-path control into Region in slot 3.
2051 // Make a cast-not-nullness use the other not-null control. Return cast.
2052 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2053 bool never_see_null) {
2054 // Initial NULL check taken path
2055 (*null_control) = top();
2056 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2058 // Generate uncommon_trap:
2059 if (never_see_null && (*null_control) != top()) {
2060 // If we see an unexpected null at a check-cast we record it and force a
2061 // recompile; the offending check-cast will be compiled to handle NULLs.
2062 // If we see more than one offending BCI, then all checkcasts in the
2063 // method will be compiled to handle NULLs.
2064 PreserveJVMState pjvms(this);
2065 set_control(*null_control);
2066 replace_in_map(value, null());
2067 uncommon_trap(Deoptimization::Reason_null_check,
2068 Deoptimization::Action_make_not_entrant);
2069 (*null_control) = top(); // NULL path is dead
2070 }
2072 // Cast away null-ness on the result
2073 return cast;
2074 }
2076 //------------------------------opt_iff----------------------------------------
2077 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2078 // Return slow-path control.
2079 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2080 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2082 // Fast path taken; set region slot 2
2083 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2084 region->init_req(2,fast_taken); // Capture fast-control
2086 // Fast path not-taken, i.e. slow path
2087 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2088 return slow_taken;
2089 }
2091 //-----------------------------make_runtime_call-------------------------------
2092 Node* GraphKit::make_runtime_call(int flags,
2093 const TypeFunc* call_type, address call_addr,
2094 const char* call_name,
2095 const TypePtr* adr_type,
2096 // The following parms are all optional.
2097 // The first NULL ends the list.
2098 Node* parm0, Node* parm1,
2099 Node* parm2, Node* parm3,
2100 Node* parm4, Node* parm5,
2101 Node* parm6, Node* parm7) {
2102 // Slow-path call
2103 int size = call_type->domain()->cnt();
2104 bool is_leaf = !(flags & RC_NO_LEAF);
2105 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2106 if (call_name == NULL) {
2107 assert(!is_leaf, "must supply name for leaf");
2108 call_name = OptoRuntime::stub_name(call_addr);
2109 }
2110 CallNode* call;
2111 if (!is_leaf) {
2112 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2113 bci(), adr_type);
2114 } else if (flags & RC_NO_FP) {
2115 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2116 } else {
2117 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2118 }
2120 // The following is similar to set_edges_for_java_call,
2121 // except that the memory effects of the call are restricted to AliasIdxRaw.
2123 // Slow path call has no side-effects, uses few values
2124 bool wide_in = !(flags & RC_NARROW_MEM);
2125 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2127 Node* prev_mem = NULL;
2128 if (wide_in) {
2129 prev_mem = set_predefined_input_for_runtime_call(call);
2130 } else {
2131 assert(!wide_out, "narrow in => narrow out");
2132 Node* narrow_mem = memory(adr_type);
2133 prev_mem = reset_memory();
2134 map()->set_memory(narrow_mem);
2135 set_predefined_input_for_runtime_call(call);
2136 }
2138 // Hook each parm in order. Stop looking at the first NULL.
2139 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2140 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2141 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2142 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2143 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2144 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2145 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2146 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2147 /* close each nested if ===> */ } } } } } } } }
2148 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2150 if (!is_leaf) {
2151 // Non-leaves can block and take safepoints:
2152 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2153 }
2154 // Non-leaves can throw exceptions:
2155 if (has_io) {
2156 call->set_req(TypeFunc::I_O, i_o());
2157 }
2159 if (flags & RC_UNCOMMON) {
2160 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2161 // (An "if" probability corresponds roughly to an unconditional count.
2162 // Sort of.)
2163 call->set_cnt(PROB_UNLIKELY_MAG(4));
2164 }
2166 Node* c = _gvn.transform(call);
2167 assert(c == call, "cannot disappear");
2169 if (wide_out) {
2170 // Slow path call has full side-effects.
2171 set_predefined_output_for_runtime_call(call);
2172 } else {
2173 // Slow path call has few side-effects, and/or sets few values.
2174 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2175 }
2177 if (has_io) {
2178 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2179 }
2180 return call;
2182 }
2184 //------------------------------merge_memory-----------------------------------
2185 // Merge memory from one path into the current memory state.
2186 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2187 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2188 Node* old_slice = mms.force_memory();
2189 Node* new_slice = mms.memory2();
2190 if (old_slice != new_slice) {
2191 PhiNode* phi;
2192 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2193 phi = new_slice->as_Phi();
2194 #ifdef ASSERT
2195 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2196 old_slice = old_slice->in(new_path);
2197 // Caller is responsible for ensuring that any pre-existing
2198 // phis are already aware of old memory.
2199 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2200 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2201 #endif
2202 mms.set_memory(phi);
2203 } else {
2204 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2205 _gvn.set_type(phi, Type::MEMORY);
2206 phi->set_req(new_path, new_slice);
2207 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2208 }
2209 }
2210 }
2211 }
2213 //------------------------------make_slow_call_ex------------------------------
2214 // Make the exception handler hookups for the slow call
2215 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2216 if (stopped()) return;
2218 // Make a catch node with just two handlers: fall-through and catch-all
2219 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2220 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2221 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2222 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2224 { PreserveJVMState pjvms(this);
2225 set_control(excp);
2226 set_i_o(i_o);
2228 if (excp != top()) {
2229 // Create an exception state also.
2230 // Use an exact type if the caller has specified a specific exception.
2231 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2232 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
2233 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2234 }
2235 }
2237 // Get the no-exception control from the CatchNode.
2238 set_control(norm);
2239 }
2242 //-------------------------------gen_subtype_check-----------------------------
2243 // Generate a subtyping check. Takes as input the subtype and supertype.
2244 // Returns 2 values: sets the default control() to the true path and returns
2245 // the false path. Only reads invariant memory; sets no (visible) memory.
2246 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2247 // but that's not exposed to the optimizer. This call also doesn't take in an
2248 // Object; if you wish to check an Object you need to load the Object's class
2249 // prior to coming here.
2250 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2251 // Fast check for identical types, perhaps identical constants.
2252 // The types can even be identical non-constants, in cases
2253 // involving Array.newInstance, Object.clone, etc.
2254 if (subklass == superklass)
2255 return top(); // false path is dead; no test needed.
2257 if (_gvn.type(superklass)->singleton()) {
2258 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2259 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2261 // In the common case of an exact superklass, try to fold up the
2262 // test before generating code. You may ask, why not just generate
2263 // the code and then let it fold up? The answer is that the generated
2264 // code will necessarily include null checks, which do not always
2265 // completely fold away. If they are also needless, then they turn
2266 // into a performance loss. Example:
2267 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2268 // Here, the type of 'fa' is often exact, so the store check
2269 // of fa[1]=x will fold up, without testing the nullness of x.
2270 switch (static_subtype_check(superk, subk)) {
2271 case SSC_always_false:
2272 {
2273 Node* always_fail = control();
2274 set_control(top());
2275 return always_fail;
2276 }
2277 case SSC_always_true:
2278 return top();
2279 case SSC_easy_test:
2280 {
2281 // Just do a direct pointer compare and be done.
2282 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2283 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2284 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2285 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2286 return _gvn.transform( new(C, 1) IfFalseNode(iff) );
2287 }
2288 case SSC_full_test:
2289 break;
2290 default:
2291 ShouldNotReachHere();
2292 }
2293 }
2295 // %%% Possible further optimization: Even if the superklass is not exact,
2296 // if the subklass is the unique subtype of the superklass, the check
2297 // will always succeed. We could leave a dependency behind to ensure this.
2299 // First load the super-klass's check-offset
2300 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
2301 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2302 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
2303 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2305 // Load from the sub-klass's super-class display list, or a 1-word cache of
2306 // the secondary superclass list, or a failing value with a sentinel offset
2307 // if the super-klass is an interface or exceptionally deep in the Java
2308 // hierarchy and we have to scan the secondary superclass list the hard way.
2309 // Worst-case type is a little odd: NULL is allowed as a result (usually
2310 // klass loads can never produce a NULL).
2311 Node *chk_off_X = ConvI2X(chk_off);
2312 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2313 // For some types like interfaces the following loadKlass is from a 1-word
2314 // cache which is mutable so can't use immutable memory. Other
2315 // types load from the super-class display table which is immutable.
2316 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2317 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2319 // Compile speed common case: ARE a subtype and we canNOT fail
2320 if( superklass == nkls )
2321 return top(); // false path is dead; no test needed.
2323 // See if we get an immediate positive hit. Happens roughly 83% of the
2324 // time. Test to see if the value loaded just previously from the subklass
2325 // is exactly the superklass.
2326 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2327 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2328 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2329 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2330 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2332 // Compile speed common case: Check for being deterministic right now. If
2333 // chk_off is a constant and not equal to cacheoff then we are NOT a
2334 // subklass. In this case we need exactly the 1 test above and we can
2335 // return those results immediately.
2336 if (!might_be_cache) {
2337 Node* not_subtype_ctrl = control();
2338 set_control(iftrue1); // We need exactly the 1 test above
2339 return not_subtype_ctrl;
2340 }
2342 // Gather the various success & failures here
2343 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2344 record_for_igvn(r_ok_subtype);
2345 RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2346 record_for_igvn(r_not_subtype);
2348 r_ok_subtype->init_req(1, iftrue1);
2350 // Check for immediate negative hit. Happens roughly 11% of the time (which
2351 // is roughly 63% of the remaining cases). Test to see if the loaded
2352 // check-offset points into the subklass display list or the 1-element
2353 // cache. If it points to the display (and NOT the cache) and the display
2354 // missed then it's not a subtype.
2355 Node *cacheoff = _gvn.intcon(cacheoff_con);
2356 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2357 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2358 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2359 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2360 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2362 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2363 // No performance impact (too rare) but allows sharing of secondary arrays
2364 // which has some footprint reduction.
2365 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2366 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2367 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2368 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2369 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2371 // -- Roads not taken here: --
2372 // We could also have chosen to perform the self-check at the beginning
2373 // of this code sequence, as the assembler does. This would not pay off
2374 // the same way, since the optimizer, unlike the assembler, can perform
2375 // static type analysis to fold away many successful self-checks.
2376 // Non-foldable self checks work better here in second position, because
2377 // the initial primary superclass check subsumes a self-check for most
2378 // types. An exception would be a secondary type like array-of-interface,
2379 // which does not appear in its own primary supertype display.
2380 // Finally, we could have chosen to move the self-check into the
2381 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2382 // dependent manner. But it is worthwhile to have the check here,
2383 // where it can be perhaps be optimized. The cost in code space is
2384 // small (register compare, branch).
2386 // Now do a linear scan of the secondary super-klass array. Again, no real
2387 // performance impact (too rare) but it's gotta be done.
2388 // Since the code is rarely used, there is no penalty for moving it
2389 // out of line, and it can only improve I-cache density.
2390 // The decision to inline or out-of-line this final check is platform
2391 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2392 Node* psc = _gvn.transform(
2393 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2395 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2396 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2397 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2398 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2399 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2401 // Return false path; set default control to true path.
2402 set_control( _gvn.transform(r_ok_subtype) );
2403 return _gvn.transform(r_not_subtype);
2404 }
2406 //----------------------------static_subtype_check-----------------------------
2407 // Shortcut important common cases when superklass is exact:
2408 // (0) superklass is java.lang.Object (can occur in reflective code)
2409 // (1) subklass is already limited to a subtype of superklass => always ok
2410 // (2) subklass does not overlap with superklass => always fail
2411 // (3) superklass has NO subtypes and we can check with a simple compare.
2412 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2413 if (StressReflectiveCode) {
2414 return SSC_full_test; // Let caller generate the general case.
2415 }
2417 if (superk == env()->Object_klass()) {
2418 return SSC_always_true; // (0) this test cannot fail
2419 }
2421 ciType* superelem = superk;
2422 if (superelem->is_array_klass())
2423 superelem = superelem->as_array_klass()->base_element_type();
2425 if (!subk->is_interface()) { // cannot trust static interface types yet
2426 if (subk->is_subtype_of(superk)) {
2427 return SSC_always_true; // (1) false path dead; no dynamic test needed
2428 }
2429 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2430 !superk->is_subtype_of(subk)) {
2431 return SSC_always_false;
2432 }
2433 }
2435 // If casting to an instance klass, it must have no subtypes
2436 if (superk->is_interface()) {
2437 // Cannot trust interfaces yet.
2438 // %%% S.B. superk->nof_implementors() == 1
2439 } else if (superelem->is_instance_klass()) {
2440 ciInstanceKlass* ik = superelem->as_instance_klass();
2441 if (!ik->has_subklass() && !ik->is_interface()) {
2442 if (!ik->is_final()) {
2443 // Add a dependency if there is a chance of a later subclass.
2444 C->dependencies()->assert_leaf_type(ik);
2445 }
2446 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2447 }
2448 } else {
2449 // A primitive array type has no subtypes.
2450 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2451 }
2453 return SSC_full_test;
2454 }
2456 // Profile-driven exact type check:
2457 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2458 float prob,
2459 Node* *casted_receiver) {
2460 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2461 Node* recv_klass = load_object_klass(receiver);
2462 Node* want_klass = makecon(tklass);
2463 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2464 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2465 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2466 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2467 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2469 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2470 assert(recv_xtype->klass_is_exact(), "");
2472 // Subsume downstream occurrences of receiver with a cast to
2473 // recv_xtype, since now we know what the type will be.
2474 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2475 (*casted_receiver) = _gvn.transform(cast);
2476 // (User must make the replace_in_map call.)
2478 return fail;
2479 }
2482 //------------------------------seems_never_null-------------------------------
2483 // Use null_seen information if it is available from the profile.
2484 // If we see an unexpected null at a type check we record it and force a
2485 // recompile; the offending check will be recompiled to handle NULLs.
2486 // If we see several offending BCIs, then all checks in the
2487 // method will be recompiled.
2488 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2489 if (UncommonNullCast // Cutout for this technique
2490 && obj != null() // And not the -Xcomp stupid case?
2491 && !too_many_traps(Deoptimization::Reason_null_check)
2492 ) {
2493 if (data == NULL)
2494 // Edge case: no mature data. Be optimistic here.
2495 return true;
2496 // If the profile has not seen a null, assume it won't happen.
2497 assert(java_bc() == Bytecodes::_checkcast ||
2498 java_bc() == Bytecodes::_instanceof ||
2499 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2500 return !data->as_BitData()->null_seen();
2501 }
2502 return false;
2503 }
2505 //------------------------maybe_cast_profiled_receiver-------------------------
2506 // If the profile has seen exactly one type, narrow to exactly that type.
2507 // Subsequent type checks will always fold up.
2508 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2509 ciProfileData* data,
2510 ciKlass* require_klass) {
2511 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2512 if (data == NULL) return NULL;
2514 // Make sure we haven't already deoptimized from this tactic.
2515 if (too_many_traps(Deoptimization::Reason_class_check))
2516 return NULL;
2518 // (No, this isn't a call, but it's enough like a virtual call
2519 // to use the same ciMethod accessor to get the profile info...)
2520 ciCallProfile profile = method()->call_profile_at_bci(bci());
2521 if (profile.count() >= 0 && // no cast failures here
2522 profile.has_receiver(0) &&
2523 profile.morphism() == 1) {
2524 ciKlass* exact_kls = profile.receiver(0);
2525 if (require_klass == NULL ||
2526 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2527 // If we narrow the type to match what the type profile sees,
2528 // we can then remove the rest of the cast.
2529 // This is a win, even if the exact_kls is very specific,
2530 // because downstream operations, such as method calls,
2531 // will often benefit from the sharper type.
2532 Node* exact_obj = not_null_obj; // will get updated in place...
2533 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2534 &exact_obj);
2535 { PreserveJVMState pjvms(this);
2536 set_control(slow_ctl);
2537 uncommon_trap(Deoptimization::Reason_class_check,
2538 Deoptimization::Action_maybe_recompile);
2539 }
2540 replace_in_map(not_null_obj, exact_obj);
2541 return exact_obj;
2542 }
2543 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2544 }
2546 return NULL;
2547 }
2550 //-------------------------------gen_instanceof--------------------------------
2551 // Generate an instance-of idiom. Used by both the instance-of bytecode
2552 // and the reflective instance-of call.
2553 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) {
2554 kill_dead_locals(); // Benefit all the uncommon traps
2555 assert( !stopped(), "dead parse path should be checked in callers" );
2556 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2557 "must check for not-null not-dead klass in callers");
2559 // Make the merge point
2560 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2561 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2562 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2563 C->set_has_split_ifs(true); // Has chance for split-if optimization
2565 ciProfileData* data = NULL;
2566 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2567 data = method()->method_data()->bci_to_data(bci());
2568 }
2569 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2570 && seems_never_null(obj, data));
2572 // Null check; get casted pointer; set region slot 3
2573 Node* null_ctl = top();
2574 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2576 // If not_null_obj is dead, only null-path is taken
2577 if (stopped()) { // Doing instance-of on a NULL?
2578 set_control(null_ctl);
2579 return intcon(0);
2580 }
2581 region->init_req(_null_path, null_ctl);
2582 phi ->init_req(_null_path, intcon(0)); // Set null path value
2583 if (null_ctl == top()) {
2584 // Do this eagerly, so that pattern matches like is_diamond_phi
2585 // will work even during parsing.
2586 assert(_null_path == PATH_LIMIT-1, "delete last");
2587 region->del_req(_null_path);
2588 phi ->del_req(_null_path);
2589 }
2591 if (ProfileDynamicTypes && data != NULL) {
2592 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL);
2593 if (stopped()) { // Profile disagrees with this path.
2594 set_control(null_ctl); // Null is the only remaining possibility.
2595 return intcon(0);
2596 }
2597 if (cast_obj != NULL)
2598 not_null_obj = cast_obj;
2599 }
2601 // Load the object's klass
2602 Node* obj_klass = load_object_klass(not_null_obj);
2604 // Generate the subtype check
2605 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2607 // Plug in the success path to the general merge in slot 1.
2608 region->init_req(_obj_path, control());
2609 phi ->init_req(_obj_path, intcon(1));
2611 // Plug in the failing path to the general merge in slot 2.
2612 region->init_req(_fail_path, not_subtype_ctrl);
2613 phi ->init_req(_fail_path, intcon(0));
2615 // Return final merged results
2616 set_control( _gvn.transform(region) );
2617 record_for_igvn(region);
2618 return _gvn.transform(phi);
2619 }
2621 //-------------------------------gen_checkcast---------------------------------
2622 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2623 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2624 // uncommon-trap paths work. Adjust stack after this call.
2625 // If failure_control is supplied and not null, it is filled in with
2626 // the control edge for the cast failure. Otherwise, an appropriate
2627 // uncommon trap or exception is thrown.
2628 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2629 Node* *failure_control) {
2630 kill_dead_locals(); // Benefit all the uncommon traps
2631 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2632 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2634 // Fast cutout: Check the case that the cast is vacuously true.
2635 // This detects the common cases where the test will short-circuit
2636 // away completely. We do this before we perform the null check,
2637 // because if the test is going to turn into zero code, we don't
2638 // want a residual null check left around. (Causes a slowdown,
2639 // for example, in some objArray manipulations, such as a[i]=a[j].)
2640 if (tk->singleton()) {
2641 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2642 if (objtp != NULL && objtp->klass() != NULL) {
2643 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2644 case SSC_always_true:
2645 return obj;
2646 case SSC_always_false:
2647 // It needs a null check because a null will *pass* the cast check.
2648 // A non-null value will always produce an exception.
2649 return do_null_assert(obj, T_OBJECT);
2650 }
2651 }
2652 }
2654 ciProfileData* data = NULL;
2655 if (failure_control == NULL) { // use MDO in regular case only
2656 assert(java_bc() == Bytecodes::_aastore ||
2657 java_bc() == Bytecodes::_checkcast,
2658 "interpreter profiles type checks only for these BCs");
2659 data = method()->method_data()->bci_to_data(bci());
2660 }
2662 // Make the merge point
2663 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2664 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2665 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
2666 C->set_has_split_ifs(true); // Has chance for split-if optimization
2668 // Use null-cast information if it is available
2669 bool never_see_null = ((failure_control == NULL) // regular case only
2670 && seems_never_null(obj, data));
2672 // Null check; get casted pointer; set region slot 3
2673 Node* null_ctl = top();
2674 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2676 // If not_null_obj is dead, only null-path is taken
2677 if (stopped()) { // Doing instance-of on a NULL?
2678 set_control(null_ctl);
2679 return null();
2680 }
2681 region->init_req(_null_path, null_ctl);
2682 phi ->init_req(_null_path, null()); // Set null path value
2683 if (null_ctl == top()) {
2684 // Do this eagerly, so that pattern matches like is_diamond_phi
2685 // will work even during parsing.
2686 assert(_null_path == PATH_LIMIT-1, "delete last");
2687 region->del_req(_null_path);
2688 phi ->del_req(_null_path);
2689 }
2691 Node* cast_obj = NULL;
2692 if (data != NULL &&
2693 // Counter has never been decremented (due to cast failure).
2694 // ...This is a reasonable thing to expect. It is true of
2695 // all casts inserted by javac to implement generic types.
2696 data->as_CounterData()->count() >= 0) {
2697 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass());
2698 if (cast_obj != NULL) {
2699 if (failure_control != NULL) // failure is now impossible
2700 (*failure_control) = top();
2701 // adjust the type of the phi to the exact klass:
2702 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR));
2703 }
2704 }
2706 if (cast_obj == NULL) {
2707 // Load the object's klass
2708 Node* obj_klass = load_object_klass(not_null_obj);
2710 // Generate the subtype check
2711 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2713 // Plug in success path into the merge
2714 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2715 not_null_obj, toop));
2716 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2717 if (failure_control == NULL) {
2718 if (not_subtype_ctrl != top()) { // If failure is possible
2719 PreserveJVMState pjvms(this);
2720 set_control(not_subtype_ctrl);
2721 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2722 }
2723 } else {
2724 (*failure_control) = not_subtype_ctrl;
2725 }
2726 }
2728 region->init_req(_obj_path, control());
2729 phi ->init_req(_obj_path, cast_obj);
2731 // A merge of NULL or Casted-NotNull obj
2732 Node* res = _gvn.transform(phi);
2734 // Note I do NOT always 'replace_in_map(obj,result)' here.
2735 // if( tk->klass()->can_be_primary_super() )
2736 // This means that if I successfully store an Object into an array-of-String
2737 // I 'forget' that the Object is really now known to be a String. I have to
2738 // do this because we don't have true union types for interfaces - if I store
2739 // a Baz into an array-of-Interface and then tell the optimizer it's an
2740 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2741 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2742 // replace_in_map( obj, res );
2744 // Return final merged results
2745 set_control( _gvn.transform(region) );
2746 record_for_igvn(region);
2747 return res;
2748 }
2750 //------------------------------next_monitor-----------------------------------
2751 // What number should be given to the next monitor?
2752 int GraphKit::next_monitor() {
2753 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2754 int next = current + C->sync_stack_slots();
2755 // Keep the toplevel high water mark current:
2756 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2757 return current;
2758 }
2760 //------------------------------insert_mem_bar---------------------------------
2761 // Memory barrier to avoid floating things around
2762 // The membar serves as a pinch point between both control and all memory slices.
2763 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2764 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2765 mb->init_req(TypeFunc::Control, control());
2766 mb->init_req(TypeFunc::Memory, reset_memory());
2767 Node* membar = _gvn.transform(mb);
2768 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2769 set_all_memory_call(membar);
2770 return membar;
2771 }
2773 //-------------------------insert_mem_bar_volatile----------------------------
2774 // Memory barrier to avoid floating things around
2775 // The membar serves as a pinch point between both control and memory(alias_idx).
2776 // If you want to make a pinch point on all memory slices, do not use this
2777 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2778 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2779 // When Parse::do_put_xxx updates a volatile field, it appends a series
2780 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2781 // The first membar is on the same memory slice as the field store opcode.
2782 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2783 // All the other membars (for other volatile slices, including AliasIdxBot,
2784 // which stands for all unknown volatile slices) are control-dependent
2785 // on the first membar. This prevents later volatile loads or stores
2786 // from sliding up past the just-emitted store.
2788 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2789 mb->set_req(TypeFunc::Control,control());
2790 if (alias_idx == Compile::AliasIdxBot) {
2791 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2792 } else {
2793 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2794 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2795 }
2796 Node* membar = _gvn.transform(mb);
2797 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2798 if (alias_idx == Compile::AliasIdxBot) {
2799 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2800 } else {
2801 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2802 }
2803 return membar;
2804 }
2806 //------------------------------shared_lock------------------------------------
2807 // Emit locking code.
2808 FastLockNode* GraphKit::shared_lock(Node* obj) {
2809 // bci is either a monitorenter bc or InvocationEntryBci
2810 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2811 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2813 if( !GenerateSynchronizationCode )
2814 return NULL; // Not locking things?
2815 if (stopped()) // Dead monitor?
2816 return NULL;
2818 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2820 // Box the stack location
2821 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2822 Node* mem = reset_memory();
2824 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2825 if (PrintPreciseBiasedLockingStatistics) {
2826 // Create the counters for this fast lock.
2827 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2828 }
2829 // Add monitor to debug info for the slow path. If we block inside the
2830 // slow path and de-opt, we need the monitor hanging around
2831 map()->push_monitor( flock );
2833 const TypeFunc *tf = LockNode::lock_type();
2834 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2836 lock->init_req( TypeFunc::Control, control() );
2837 lock->init_req( TypeFunc::Memory , mem );
2838 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2839 lock->init_req( TypeFunc::FramePtr, frameptr() );
2840 lock->init_req( TypeFunc::ReturnAdr, top() );
2842 lock->init_req(TypeFunc::Parms + 0, obj);
2843 lock->init_req(TypeFunc::Parms + 1, box);
2844 lock->init_req(TypeFunc::Parms + 2, flock);
2845 add_safepoint_edges(lock);
2847 lock = _gvn.transform( lock )->as_Lock();
2849 // lock has no side-effects, sets few values
2850 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2852 insert_mem_bar(Op_MemBarAcquire);
2854 // Add this to the worklist so that the lock can be eliminated
2855 record_for_igvn(lock);
2857 #ifndef PRODUCT
2858 if (PrintLockStatistics) {
2859 // Update the counter for this lock. Don't bother using an atomic
2860 // operation since we don't require absolute accuracy.
2861 lock->create_lock_counter(map()->jvms());
2862 increment_counter(lock->counter()->addr());
2863 }
2864 #endif
2866 return flock;
2867 }
2870 //------------------------------shared_unlock----------------------------------
2871 // Emit unlocking code.
2872 void GraphKit::shared_unlock(Node* box, Node* obj) {
2873 // bci is either a monitorenter bc or InvocationEntryBci
2874 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2875 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2877 if( !GenerateSynchronizationCode )
2878 return;
2879 if (stopped()) { // Dead monitor?
2880 map()->pop_monitor(); // Kill monitor from debug info
2881 return;
2882 }
2884 // Memory barrier to avoid floating things down past the locked region
2885 insert_mem_bar(Op_MemBarRelease);
2887 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2888 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2889 uint raw_idx = Compile::AliasIdxRaw;
2890 unlock->init_req( TypeFunc::Control, control() );
2891 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2892 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2893 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2894 unlock->init_req( TypeFunc::ReturnAdr, top() );
2896 unlock->init_req(TypeFunc::Parms + 0, obj);
2897 unlock->init_req(TypeFunc::Parms + 1, box);
2898 unlock = _gvn.transform(unlock)->as_Unlock();
2900 Node* mem = reset_memory();
2902 // unlock has no side-effects, sets few values
2903 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2905 // Kill monitor from debug info
2906 map()->pop_monitor( );
2907 }
2909 //-------------------------------get_layout_helper-----------------------------
2910 // If the given klass is a constant or known to be an array,
2911 // fetch the constant layout helper value into constant_value
2912 // and return (Node*)NULL. Otherwise, load the non-constant
2913 // layout helper value, and return the node which represents it.
2914 // This two-faced routine is useful because allocation sites
2915 // almost always feature constant types.
2916 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2917 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2918 if (!StressReflectiveCode && inst_klass != NULL) {
2919 ciKlass* klass = inst_klass->klass();
2920 bool xklass = inst_klass->klass_is_exact();
2921 if (xklass || klass->is_array_klass()) {
2922 jint lhelper = klass->layout_helper();
2923 if (lhelper != Klass::_lh_neutral_value) {
2924 constant_value = lhelper;
2925 return (Node*) NULL;
2926 }
2927 }
2928 }
2929 constant_value = Klass::_lh_neutral_value; // put in a known value
2930 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
2931 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2932 }
2934 // We just put in an allocate/initialize with a big raw-memory effect.
2935 // Hook selected additional alias categories on the initialization.
2936 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2937 MergeMemNode* init_in_merge,
2938 Node* init_out_raw) {
2939 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2940 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2942 Node* prevmem = kit.memory(alias_idx);
2943 init_in_merge->set_memory_at(alias_idx, prevmem);
2944 kit.set_memory(init_out_raw, alias_idx);
2945 }
2947 //---------------------------set_output_for_allocation-------------------------
2948 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2949 const TypeOopPtr* oop_type) {
2950 int rawidx = Compile::AliasIdxRaw;
2951 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2952 add_safepoint_edges(alloc);
2953 Node* allocx = _gvn.transform(alloc);
2954 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2955 // create memory projection for i_o
2956 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2957 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2959 // create a memory projection as for the normal control path
2960 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2961 set_memory(malloc, rawidx);
2963 // a normal slow-call doesn't change i_o, but an allocation does
2964 // we create a separate i_o projection for the normal control path
2965 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2966 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2968 // put in an initialization barrier
2969 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2970 rawoop)->as_Initialize();
2971 assert(alloc->initialization() == init, "2-way macro link must work");
2972 assert(init ->allocation() == alloc, "2-way macro link must work");
2973 {
2974 // Extract memory strands which may participate in the new object's
2975 // initialization, and source them from the new InitializeNode.
2976 // This will allow us to observe initializations when they occur,
2977 // and link them properly (as a group) to the InitializeNode.
2978 assert(init->in(InitializeNode::Memory) == malloc, "");
2979 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
2980 init->set_req(InitializeNode::Memory, minit_in);
2981 record_for_igvn(minit_in); // fold it up later, if possible
2982 Node* minit_out = memory(rawidx);
2983 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
2984 if (oop_type->isa_aryptr()) {
2985 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
2986 int elemidx = C->get_alias_index(telemref);
2987 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
2988 } else if (oop_type->isa_instptr()) {
2989 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
2990 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
2991 ciField* field = ik->nonstatic_field_at(i);
2992 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
2993 continue; // do not bother to track really large numbers of fields
2994 // Find (or create) the alias category for this field:
2995 int fieldidx = C->alias_type(field)->index();
2996 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
2997 }
2998 }
2999 }
3001 // Cast raw oop to the real thing...
3002 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
3003 javaoop = _gvn.transform(javaoop);
3004 C->set_recent_alloc(control(), javaoop);
3005 assert(just_allocated_object(control()) == javaoop, "just allocated");
3007 #ifdef ASSERT
3008 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3009 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3010 "Ideal_allocation works");
3011 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3012 "Ideal_allocation works");
3013 if (alloc->is_AllocateArray()) {
3014 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3015 "Ideal_allocation works");
3016 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3017 "Ideal_allocation works");
3018 } else {
3019 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3020 }
3021 }
3022 #endif //ASSERT
3024 return javaoop;
3025 }
3027 //---------------------------new_instance--------------------------------------
3028 // This routine takes a klass_node which may be constant (for a static type)
3029 // or may be non-constant (for reflective code). It will work equally well
3030 // for either, and the graph will fold nicely if the optimizer later reduces
3031 // the type to a constant.
3032 // The optional arguments are for specialized use by intrinsics:
3033 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3034 // - If 'return_size_val', report the the total object size to the caller.
3035 Node* GraphKit::new_instance(Node* klass_node,
3036 Node* extra_slow_test,
3037 Node* *return_size_val) {
3038 // Compute size in doublewords
3039 // The size is always an integral number of doublewords, represented
3040 // as a positive bytewise size stored in the klass's layout_helper.
3041 // The layout_helper also encodes (in a low bit) the need for a slow path.
3042 jint layout_con = Klass::_lh_neutral_value;
3043 Node* layout_val = get_layout_helper(klass_node, layout_con);
3044 int layout_is_con = (layout_val == NULL);
3046 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3047 // Generate the initial go-slow test. It's either ALWAYS (return a
3048 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3049 // case) a computed value derived from the layout_helper.
3050 Node* initial_slow_test = NULL;
3051 if (layout_is_con) {
3052 assert(!StressReflectiveCode, "stress mode does not use these paths");
3053 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3054 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3056 } else { // reflective case
3057 // This reflective path is used by Unsafe.allocateInstance.
3058 // (It may be stress-tested by specifying StressReflectiveCode.)
3059 // Basically, we want to get into the VM is there's an illegal argument.
3060 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3061 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
3062 if (extra_slow_test != intcon(0)) {
3063 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
3064 }
3065 // (Macro-expander will further convert this to a Bool, if necessary.)
3066 }
3068 // Find the size in bytes. This is easy; it's the layout_helper.
3069 // The size value must be valid even if the slow path is taken.
3070 Node* size = NULL;
3071 if (layout_is_con) {
3072 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3073 } else { // reflective case
3074 // This reflective path is used by clone and Unsafe.allocateInstance.
3075 size = ConvI2X(layout_val);
3077 // Clear the low bits to extract layout_helper_size_in_bytes:
3078 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3079 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3080 size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
3081 }
3082 if (return_size_val != NULL) {
3083 (*return_size_val) = size;
3084 }
3086 // This is a precise notnull oop of the klass.
3087 // (Actually, it need not be precise if this is a reflective allocation.)
3088 // It's what we cast the result to.
3089 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3090 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3091 const TypeOopPtr* oop_type = tklass->as_instance_type();
3093 // Now generate allocation code
3095 // The entire memory state is needed for slow path of the allocation
3096 // since GC and deoptimization can happened.
3097 Node *mem = reset_memory();
3098 set_all_memory(mem); // Create new memory state
3100 AllocateNode* alloc
3101 = new (C, AllocateNode::ParmLimit)
3102 AllocateNode(C, AllocateNode::alloc_type(),
3103 control(), mem, i_o(),
3104 size, klass_node,
3105 initial_slow_test);
3107 return set_output_for_allocation(alloc, oop_type);
3108 }
3110 //-------------------------------new_array-------------------------------------
3111 // helper for both newarray and anewarray
3112 // The 'length' parameter is (obviously) the length of the array.
3113 // See comments on new_instance for the meaning of the other arguments.
3114 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3115 Node* length, // number of array elements
3116 int nargs, // number of arguments to push back for uncommon trap
3117 Node* *return_size_val) {
3118 jint layout_con = Klass::_lh_neutral_value;
3119 Node* layout_val = get_layout_helper(klass_node, layout_con);
3120 int layout_is_con = (layout_val == NULL);
3122 if (!layout_is_con && !StressReflectiveCode &&
3123 !too_many_traps(Deoptimization::Reason_class_check)) {
3124 // This is a reflective array creation site.
3125 // Optimistically assume that it is a subtype of Object[],
3126 // so that we can fold up all the address arithmetic.
3127 layout_con = Klass::array_layout_helper(T_OBJECT);
3128 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
3129 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
3130 { BuildCutout unless(this, bol_lh, PROB_MAX);
3131 _sp += nargs;
3132 uncommon_trap(Deoptimization::Reason_class_check,
3133 Deoptimization::Action_maybe_recompile);
3134 }
3135 layout_val = NULL;
3136 layout_is_con = true;
3137 }
3139 // Generate the initial go-slow test. Make sure we do not overflow
3140 // if length is huge (near 2Gig) or negative! We do not need
3141 // exact double-words here, just a close approximation of needed
3142 // double-words. We can't add any offset or rounding bits, lest we
3143 // take a size -1 of bytes and make it positive. Use an unsigned
3144 // compare, so negative sizes look hugely positive.
3145 int fast_size_limit = FastAllocateSizeLimit;
3146 if (layout_is_con) {
3147 assert(!StressReflectiveCode, "stress mode does not use these paths");
3148 // Increase the size limit if we have exact knowledge of array type.
3149 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3150 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3151 }
3153 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3154 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3155 if (initial_slow_test->is_Bool()) {
3156 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3157 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3158 }
3160 // --- Size Computation ---
3161 // array_size = round_to_heap(array_header + (length << elem_shift));
3162 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3163 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3164 // The rounding mask is strength-reduced, if possible.
3165 int round_mask = MinObjAlignmentInBytes - 1;
3166 Node* header_size = NULL;
3167 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3168 // (T_BYTE has the weakest alignment and size restrictions...)
3169 if (layout_is_con) {
3170 int hsize = Klass::layout_helper_header_size(layout_con);
3171 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3172 BasicType etype = Klass::layout_helper_element_type(layout_con);
3173 if ((round_mask & ~right_n_bits(eshift)) == 0)
3174 round_mask = 0; // strength-reduce it if it goes away completely
3175 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3176 assert(header_size_min <= hsize, "generic minimum is smallest");
3177 header_size_min = hsize;
3178 header_size = intcon(hsize + round_mask);
3179 } else {
3180 Node* hss = intcon(Klass::_lh_header_size_shift);
3181 Node* hsm = intcon(Klass::_lh_header_size_mask);
3182 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3183 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3184 Node* mask = intcon(round_mask);
3185 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3186 }
3188 Node* elem_shift = NULL;
3189 if (layout_is_con) {
3190 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3191 if (eshift != 0)
3192 elem_shift = intcon(eshift);
3193 } else {
3194 // There is no need to mask or shift this value.
3195 // The semantics of LShiftINode include an implicit mask to 0x1F.
3196 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3197 elem_shift = layout_val;
3198 }
3200 // Transition to native address size for all offset calculations:
3201 Node* lengthx = ConvI2X(length);
3202 Node* headerx = ConvI2X(header_size);
3203 #ifdef _LP64
3204 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3205 if (tllen != NULL && tllen->_lo < 0) {
3206 // Add a manual constraint to a positive range. Cf. array_element_address.
3207 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3208 if (size_max > tllen->_hi) size_max = tllen->_hi;
3209 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3210 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3211 }
3212 }
3213 #endif
3215 // Combine header size (plus rounding) and body size. Then round down.
3216 // This computation cannot overflow, because it is used only in two
3217 // places, one where the length is sharply limited, and the other
3218 // after a successful allocation.
3219 Node* abody = lengthx;
3220 if (elem_shift != NULL)
3221 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3222 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3223 if (round_mask != 0) {
3224 Node* mask = MakeConX(~round_mask);
3225 size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3226 }
3227 // else if round_mask == 0, the size computation is self-rounding
3229 if (return_size_val != NULL) {
3230 // This is the size
3231 (*return_size_val) = size;
3232 }
3234 // Now generate allocation code
3236 // The entire memory state is needed for slow path of the allocation
3237 // since GC and deoptimization can happened.
3238 Node *mem = reset_memory();
3239 set_all_memory(mem); // Create new memory state
3241 // Create the AllocateArrayNode and its result projections
3242 AllocateArrayNode* alloc
3243 = new (C, AllocateArrayNode::ParmLimit)
3244 AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3245 control(), mem, i_o(),
3246 size, klass_node,
3247 initial_slow_test,
3248 length);
3250 // Cast to correct type. Note that the klass_node may be constant or not,
3251 // and in the latter case the actual array type will be inexact also.
3252 // (This happens via a non-constant argument to inline_native_newArray.)
3253 // In any case, the value of klass_node provides the desired array type.
3254 const TypeInt* length_type = _gvn.find_int_type(length);
3255 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3256 if (ary_type->isa_aryptr() && length_type != NULL) {
3257 // Try to get a better type than POS for the size
3258 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3259 }
3261 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3263 // Cast length on remaining path to be as narrow as possible
3264 if (map()->find_edge(length) >= 0) {
3265 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3266 if (ccast != length) {
3267 _gvn.set_type_bottom(ccast);
3268 record_for_igvn(ccast);
3269 replace_in_map(length, ccast);
3270 }
3271 }
3273 return javaoop;
3274 }
3276 // The following "Ideal_foo" functions are placed here because they recognize
3277 // the graph shapes created by the functions immediately above.
3279 //---------------------------Ideal_allocation----------------------------------
3280 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3281 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3282 if (ptr == NULL) { // reduce dumb test in callers
3283 return NULL;
3284 }
3285 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3286 ptr = ptr->in(1);
3287 if (ptr == NULL) return NULL;
3288 }
3289 if (ptr->is_Proj()) {
3290 Node* allo = ptr->in(0);
3291 if (allo != NULL && allo->is_Allocate()) {
3292 return allo->as_Allocate();
3293 }
3294 }
3295 // Report failure to match.
3296 return NULL;
3297 }
3299 // Fancy version which also strips off an offset (and reports it to caller).
3300 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3301 intptr_t& offset) {
3302 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3303 if (base == NULL) return NULL;
3304 return Ideal_allocation(base, phase);
3305 }
3307 // Trace Initialize <- Proj[Parm] <- Allocate
3308 AllocateNode* InitializeNode::allocation() {
3309 Node* rawoop = in(InitializeNode::RawAddress);
3310 if (rawoop->is_Proj()) {
3311 Node* alloc = rawoop->in(0);
3312 if (alloc->is_Allocate()) {
3313 return alloc->as_Allocate();
3314 }
3315 }
3316 return NULL;
3317 }
3319 // Trace Allocate -> Proj[Parm] -> Initialize
3320 InitializeNode* AllocateNode::initialization() {
3321 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3322 if (rawoop == NULL) return NULL;
3323 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3324 Node* init = rawoop->fast_out(i);
3325 if (init->is_Initialize()) {
3326 assert(init->as_Initialize()->allocation() == this, "2-way link");
3327 return init->as_Initialize();
3328 }
3329 }
3330 return NULL;
3331 }
3333 //----------------------------- loop predicates ---------------------------
3335 //------------------------------add_predicate_impl----------------------------
3336 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3337 // Too many traps seen?
3338 if (too_many_traps(reason)) {
3339 #ifdef ASSERT
3340 if (TraceLoopPredicate) {
3341 int tc = C->trap_count(reason);
3342 tty->print("too many traps=%s tcount=%d in ",
3343 Deoptimization::trap_reason_name(reason), tc);
3344 method()->print(); // which method has too many predicate traps
3345 tty->cr();
3346 }
3347 #endif
3348 // We cannot afford to take more traps here,
3349 // do not generate predicate.
3350 return;
3351 }
3353 Node *cont = _gvn.intcon(1);
3354 Node* opq = _gvn.transform(new (C, 2) Opaque1Node(C, cont));
3355 Node *bol = _gvn.transform(new (C, 2) Conv2BNode(opq));
3356 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3357 Node* iffalse = _gvn.transform(new (C, 1) IfFalseNode(iff));
3358 C->add_predicate_opaq(opq);
3359 {
3360 PreserveJVMState pjvms(this);
3361 set_control(iffalse);
3362 _sp += nargs;
3363 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3364 }
3365 Node* iftrue = _gvn.transform(new (C, 1) IfTrueNode(iff));
3366 set_control(iftrue);
3367 }
3369 //------------------------------add_predicate---------------------------------
3370 void GraphKit::add_predicate(int nargs) {
3371 if (UseLoopPredicate) {
3372 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3373 }
3374 }
3376 //----------------------------- store barriers ----------------------------
3377 #define __ ideal.
3379 void GraphKit::sync_kit(IdealKit& ideal) {
3380 set_all_memory(__ merged_memory());
3381 set_i_o(__ i_o());
3382 set_control(__ ctrl());
3383 }
3385 void GraphKit::final_sync(IdealKit& ideal) {
3386 // Final sync IdealKit and graphKit.
3387 __ drain_delay_transform();
3388 sync_kit(ideal);
3389 }
3391 // vanilla/CMS post barrier
3392 // Insert a write-barrier store. This is to let generational GC work; we have
3393 // to flag all oop-stores before the next GC point.
3394 void GraphKit::write_barrier_post(Node* oop_store,
3395 Node* obj,
3396 Node* adr,
3397 uint adr_idx,
3398 Node* val,
3399 bool use_precise) {
3400 // No store check needed if we're storing a NULL or an old object
3401 // (latter case is probably a string constant). The concurrent
3402 // mark sweep garbage collector, however, needs to have all nonNull
3403 // oop updates flagged via card-marks.
3404 if (val != NULL && val->is_Con()) {
3405 // must be either an oop or NULL
3406 const Type* t = val->bottom_type();
3407 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3408 // stores of null never (?) need barriers
3409 return;
3410 ciObject* con = t->is_oopptr()->const_oop();
3411 if (con != NULL
3412 && con->is_perm()
3413 && Universe::heap()->can_elide_permanent_oop_store_barriers())
3414 // no store barrier needed, because no old-to-new ref created
3415 return;
3416 }
3418 if (use_ReduceInitialCardMarks()
3419 && obj == just_allocated_object(control())) {
3420 // We can skip marks on a freshly-allocated object in Eden.
3421 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3422 // That routine informs GC to take appropriate compensating steps,
3423 // upon a slow-path allocation, so as to make this card-mark
3424 // elision safe.
3425 return;
3426 }
3428 if (!use_precise) {
3429 // All card marks for a (non-array) instance are in one place:
3430 adr = obj;
3431 }
3432 // (Else it's an array (or unknown), and we want more precise card marks.)
3433 assert(adr != NULL, "");
3435 IdealKit ideal(this, true);
3437 // Convert the pointer to an int prior to doing math on it
3438 Node* cast = __ CastPX(__ ctrl(), adr);
3440 // Divide by card size
3441 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3442 "Only one we handle so far.");
3443 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3445 // Combine card table base and card offset
3446 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3448 // Get the alias_index for raw card-mark memory
3449 int adr_type = Compile::AliasIdxRaw;
3450 Node* zero = __ ConI(0); // Dirty card value
3451 BasicType bt = T_BYTE;
3453 if (UseCondCardMark) {
3454 // The classic GC reference write barrier is typically implemented
3455 // as a store into the global card mark table. Unfortunately
3456 // unconditional stores can result in false sharing and excessive
3457 // coherence traffic as well as false transactional aborts.
3458 // UseCondCardMark enables MP "polite" conditional card mark
3459 // stores. In theory we could relax the load from ctrl() to
3460 // no_ctrl, but that doesn't buy much latitude.
3461 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3462 __ if_then(card_val, BoolTest::ne, zero);
3463 }
3465 // Smash zero into card
3466 if( !UseConcMarkSweepGC ) {
3467 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3468 } else {
3469 // Specialized path for CM store barrier
3470 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3471 }
3473 if (UseCondCardMark) {
3474 __ end_if();
3475 }
3477 // Final sync IdealKit and GraphKit.
3478 final_sync(ideal);
3479 }
3481 // G1 pre/post barriers
3482 void GraphKit::g1_write_barrier_pre(Node* obj,
3483 Node* adr,
3484 uint alias_idx,
3485 Node* val,
3486 const TypeOopPtr* val_type,
3487 BasicType bt) {
3488 IdealKit ideal(this, true);
3490 Node* tls = __ thread(); // ThreadLocalStorage
3492 Node* no_ctrl = NULL;
3493 Node* no_base = __ top();
3494 Node* zero = __ ConI(0);
3496 float likely = PROB_LIKELY(0.999);
3497 float unlikely = PROB_UNLIKELY(0.999);
3499 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3500 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3502 // Offsets into the thread
3503 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3504 PtrQueue::byte_offset_of_active());
3505 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3506 PtrQueue::byte_offset_of_index());
3507 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3508 PtrQueue::byte_offset_of_buf());
3509 // Now the actual pointers into the thread
3511 // set_control( ctl);
3513 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3514 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3515 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3517 // Now some of the values
3519 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3521 // if (!marking)
3522 __ if_then(marking, BoolTest::ne, zero); {
3523 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3525 const Type* t1 = adr->bottom_type();
3526 const Type* t2 = val->bottom_type();
3528 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3529 // if (orig != NULL)
3530 __ if_then(orig, BoolTest::ne, null()); {
3531 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3533 // load original value
3534 // alias_idx correct??
3536 // is the queue for this thread full?
3537 __ if_then(index, BoolTest::ne, zero, likely); {
3539 // decrement the index
3540 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3541 Node* next_indexX = next_index;
3542 #ifdef _LP64
3543 // We could refine the type for what it's worth
3544 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3545 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3546 #endif
3548 // Now get the buffer location we will log the original value into and store it
3549 Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3550 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw);
3552 // update the index
3553 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3555 } __ else_(); {
3557 // logging buffer is full, call the runtime
3558 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3559 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls);
3560 } __ end_if(); // (!index)
3561 } __ end_if(); // (orig != NULL)
3562 } __ end_if(); // (!marking)
3564 // Final sync IdealKit and GraphKit.
3565 final_sync(ideal);
3566 }
3568 //
3569 // Update the card table and add card address to the queue
3570 //
3571 void GraphKit::g1_mark_card(IdealKit& ideal,
3572 Node* card_adr,
3573 Node* oop_store,
3574 uint oop_alias_idx,
3575 Node* index,
3576 Node* index_adr,
3577 Node* buffer,
3578 const TypeFunc* tf) {
3580 Node* zero = __ ConI(0);
3581 Node* no_base = __ top();
3582 BasicType card_bt = T_BYTE;
3583 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3584 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3586 // Now do the queue work
3587 __ if_then(index, BoolTest::ne, zero); {
3589 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3590 Node* next_indexX = next_index;
3591 #ifdef _LP64
3592 // We could refine the type for what it's worth
3593 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3594 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3595 #endif // _LP64
3596 Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3598 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3599 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3601 } __ else_(); {
3602 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3603 } __ end_if();
3605 }
3607 void GraphKit::g1_write_barrier_post(Node* oop_store,
3608 Node* obj,
3609 Node* adr,
3610 uint alias_idx,
3611 Node* val,
3612 BasicType bt,
3613 bool use_precise) {
3614 // If we are writing a NULL then we need no post barrier
3616 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3617 // Must be NULL
3618 const Type* t = val->bottom_type();
3619 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3620 // No post barrier if writing NULLx
3621 return;
3622 }
3624 if (!use_precise) {
3625 // All card marks for a (non-array) instance are in one place:
3626 adr = obj;
3627 }
3628 // (Else it's an array (or unknown), and we want more precise card marks.)
3629 assert(adr != NULL, "");
3631 IdealKit ideal(this, true);
3633 Node* tls = __ thread(); // ThreadLocalStorage
3635 Node* no_base = __ top();
3636 float likely = PROB_LIKELY(0.999);
3637 float unlikely = PROB_UNLIKELY(0.999);
3638 Node* zero = __ ConI(0);
3639 Node* zeroX = __ ConX(0);
3641 // Get the alias_index for raw card-mark memory
3642 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3644 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3646 // Offsets into the thread
3647 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3648 PtrQueue::byte_offset_of_index());
3649 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3650 PtrQueue::byte_offset_of_buf());
3652 // Pointers into the thread
3654 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3655 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3657 // Now some values
3658 // Use ctrl to avoid hoisting these values past a safepoint, which could
3659 // potentially reset these fields in the JavaThread.
3660 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3661 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3663 // Convert the store obj pointer to an int prior to doing math on it
3664 // Must use ctrl to prevent "integerized oop" existing across safepoint
3665 Node* cast = __ CastPX(__ ctrl(), adr);
3667 // Divide pointer by card size
3668 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3670 // Combine card table base and card offset
3671 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3673 // If we know the value being stored does it cross regions?
3675 if (val != NULL) {
3676 // Does the store cause us to cross regions?
3678 // Should be able to do an unsigned compare of region_size instead of
3679 // and extra shift. Do we have an unsigned compare??
3680 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3681 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3683 // if (xor_res == 0) same region so skip
3684 __ if_then(xor_res, BoolTest::ne, zeroX); {
3686 // No barrier if we are storing a NULL
3687 __ if_then(val, BoolTest::ne, null(), unlikely); {
3689 // Ok must mark the card if not already dirty
3691 // load the original value of the card
3692 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3694 __ if_then(card_val, BoolTest::ne, zero); {
3695 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3696 } __ end_if();
3697 } __ end_if();
3698 } __ end_if();
3699 } else {
3700 // Object.clone() instrinsic uses this path.
3701 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3702 }
3704 // Final sync IdealKit and GraphKit.
3705 final_sync(ideal);
3706 }
3707 #undef __