1 /*
2 * Copyright (c) 2001, 2013, 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 */
24
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"
42
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 }
54
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 }
66
67
68
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 }
84
85
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 }
95
96 //--------------------------------sync_jvms_for_reexecute---------------------
97 // Make sure our current jvms agrees with our parse state. This version
98 // uses the reexecute_sp for reexecuting bytecodes.
99 JVMState* GraphKit::sync_jvms_for_reexecute() {
100 JVMState* jvms = this->jvms();
101 jvms->set_bci(bci()); // Record the new bci in the JVMState
102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
103 return jvms;
104 }
105
106 #ifdef ASSERT
107 bool GraphKit::jvms_in_sync() const {
108 Parse* parse = is_Parse();
109 if (parse == NULL) {
110 if (bci() != jvms()->bci()) return false;
111 if (sp() != (int)jvms()->sp()) return false;
112 return true;
113 }
114 if (jvms()->method() != parse->method()) return false;
115 if (jvms()->bci() != parse->bci()) return false;
116 int jvms_sp = jvms()->sp();
117 if (jvms_sp != parse->sp()) return false;
118 int jvms_depth = jvms()->depth();
119 if (jvms_depth != parse->depth()) return false;
120 return true;
121 }
122
123 // Local helper checks for special internal merge points
124 // used to accumulate and merge exception states.
125 // They are marked by the region's in(0) edge being the map itself.
126 // Such merge points must never "escape" into the parser at large,
127 // until they have been handed to gvn.transform.
128 static bool is_hidden_merge(Node* reg) {
129 if (reg == NULL) return false;
130 if (reg->is_Phi()) {
131 reg = reg->in(0);
132 if (reg == NULL) return false;
133 }
134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
135 }
136
137 void GraphKit::verify_map() const {
138 if (map() == NULL) return; // null map is OK
139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
142 }
143
144 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
145 assert(ex_map->next_exception() == NULL, "not already part of a chain");
146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
147 }
148 #endif
149
150 //---------------------------stop_and_kill_map---------------------------------
151 // Set _map to NULL, signalling a stop to further bytecode execution.
152 // First smash the current map's control to a constant, to mark it dead.
153 void GraphKit::stop_and_kill_map() {
154 SafePointNode* dead_map = stop();
155 if (dead_map != NULL) {
156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
157 assert(dead_map->is_killed(), "must be so marked");
158 }
159 }
160
161
162 //--------------------------------stopped--------------------------------------
163 // Tell if _map is NULL, or control is top.
164 bool GraphKit::stopped() {
165 if (map() == NULL) return true;
166 else if (control() == top()) return true;
167 else return false;
168 }
169
170
171 //-----------------------------has_ex_handler----------------------------------
172 // Tell if this method or any caller method has exception handlers.
173 bool GraphKit::has_ex_handler() {
174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
176 return true;
177 }
178 }
179 return false;
180 }
181
182 //------------------------------save_ex_oop------------------------------------
183 // Save an exception without blowing stack contents or other JVM state.
184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
186 ex_map->add_req(ex_oop);
187 debug_only(verify_exception_state(ex_map));
188 }
189
190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
192 Node* ex_oop = ex_map->in(ex_map->req()-1);
193 if (clear_it) ex_map->del_req(ex_map->req()-1);
194 return ex_oop;
195 }
196
197 //-----------------------------saved_ex_oop------------------------------------
198 // Recover a saved exception from its map.
199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
200 return common_saved_ex_oop(ex_map, false);
201 }
202
203 //--------------------------clear_saved_ex_oop---------------------------------
204 // Erase a previously saved exception from its map.
205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, true);
207 }
208
209 #ifdef ASSERT
210 //---------------------------has_saved_ex_oop----------------------------------
211 // Erase a previously saved exception from its map.
212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
213 return ex_map->req() == ex_map->jvms()->endoff()+1;
214 }
215 #endif
216
217 //-------------------------make_exception_state--------------------------------
218 // Turn the current JVM state into an exception state, appending the ex_oop.
219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
220 sync_jvms();
221 SafePointNode* ex_map = stop(); // do not manipulate this map any more
222 set_saved_ex_oop(ex_map, ex_oop);
223 return ex_map;
224 }
225
226
227 //--------------------------add_exception_state--------------------------------
228 // Add an exception to my list of exceptions.
229 void GraphKit::add_exception_state(SafePointNode* ex_map) {
230 if (ex_map == NULL || ex_map->control() == top()) {
231 return;
232 }
233 #ifdef ASSERT
234 verify_exception_state(ex_map);
235 if (has_exceptions()) {
236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
237 }
238 #endif
239
240 // If there is already an exception of exactly this type, merge with it.
241 // In particular, null-checks and other low-level exceptions common up here.
242 Node* ex_oop = saved_ex_oop(ex_map);
243 const Type* ex_type = _gvn.type(ex_oop);
244 if (ex_oop == top()) {
245 // No action needed.
246 return;
247 }
248 assert(ex_type->isa_instptr(), "exception must be an instance");
249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
251 // We check sp also because call bytecodes can generate exceptions
252 // both before and after arguments are popped!
253 if (ex_type2 == ex_type
254 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
255 combine_exception_states(ex_map, e2);
256 return;
257 }
258 }
259
260 // No pre-existing exception of the same type. Chain it on the list.
261 push_exception_state(ex_map);
262 }
263
264 //-----------------------add_exception_states_from-----------------------------
265 void GraphKit::add_exception_states_from(JVMState* jvms) {
266 SafePointNode* ex_map = jvms->map()->next_exception();
267 if (ex_map != NULL) {
268 jvms->map()->set_next_exception(NULL);
269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
270 next_map = ex_map->next_exception();
271 ex_map->set_next_exception(NULL);
272 add_exception_state(ex_map);
273 }
274 }
275 }
276
277 //-----------------------transfer_exceptions_into_jvms-------------------------
278 JVMState* GraphKit::transfer_exceptions_into_jvms() {
279 if (map() == NULL) {
280 // We need a JVMS to carry the exceptions, but the map has gone away.
281 // Create a scratch JVMS, cloned from any of the exception states...
282 if (has_exceptions()) {
283 _map = _exceptions;
284 _map = clone_map();
285 _map->set_next_exception(NULL);
286 clear_saved_ex_oop(_map);
287 debug_only(verify_map());
288 } else {
289 // ...or created from scratch
290 JVMState* jvms = new (C) JVMState(_method, NULL);
291 jvms->set_bci(_bci);
292 jvms->set_sp(_sp);
293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
294 set_jvms(jvms);
295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
296 set_all_memory(top());
297 while (map()->req() < jvms->endoff()) map()->add_req(top());
298 }
299 // (This is a kludge, in case you didn't notice.)
300 set_control(top());
301 }
302 JVMState* jvms = sync_jvms();
303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
304 jvms->map()->set_next_exception(_exceptions);
305 _exceptions = NULL; // done with this set of exceptions
306 return jvms;
307 }
308
309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
310 assert(is_hidden_merge(dstphi), "must be a special merge node");
311 assert(is_hidden_merge(srcphi), "must be a special merge node");
312 uint limit = srcphi->req();
313 for (uint i = PhiNode::Input; i < limit; i++) {
314 dstphi->add_req(srcphi->in(i));
315 }
316 }
317 static inline void add_one_req(Node* dstphi, Node* src) {
318 assert(is_hidden_merge(dstphi), "must be a special merge node");
319 assert(!is_hidden_merge(src), "must not be a special merge node");
320 dstphi->add_req(src);
321 }
322
323 //-----------------------combine_exception_states------------------------------
324 // This helper function combines exception states by building phis on a
325 // specially marked state-merging region. These regions and phis are
326 // untransformed, and can build up gradually. The region is marked by
327 // having a control input of its exception map, rather than NULL. Such
328 // regions do not appear except in this function, and in use_exception_state.
329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
330 if (failing()) return; // dying anyway...
331 JVMState* ex_jvms = ex_map->_jvms;
332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
336 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
337 assert(ex_map->req() == phi_map->req(), "matching maps");
338 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
339 Node* hidden_merge_mark = root();
340 Node* region = phi_map->control();
341 MergeMemNode* phi_mem = phi_map->merged_memory();
342 MergeMemNode* ex_mem = ex_map->merged_memory();
343 if (region->in(0) != hidden_merge_mark) {
344 // The control input is not (yet) a specially-marked region in phi_map.
345 // Make it so, and build some phis.
346 region = new (C) RegionNode(2);
347 _gvn.set_type(region, Type::CONTROL);
348 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
349 region->init_req(1, phi_map->control());
350 phi_map->set_control(region);
351 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
352 record_for_igvn(io_phi);
353 _gvn.set_type(io_phi, Type::ABIO);
354 phi_map->set_i_o(io_phi);
355 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
356 Node* m = mms.memory();
357 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
358 record_for_igvn(m_phi);
359 _gvn.set_type(m_phi, Type::MEMORY);
360 mms.set_memory(m_phi);
361 }
362 }
363
364 // Either or both of phi_map and ex_map might already be converted into phis.
365 Node* ex_control = ex_map->control();
366 // if there is special marking on ex_map also, we add multiple edges from src
367 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
368 // how wide was the destination phi_map, originally?
369 uint orig_width = region->req();
370
371 if (add_multiple) {
372 add_n_reqs(region, ex_control);
373 add_n_reqs(phi_map->i_o(), ex_map->i_o());
374 } else {
375 // ex_map has no merges, so we just add single edges everywhere
376 add_one_req(region, ex_control);
377 add_one_req(phi_map->i_o(), ex_map->i_o());
378 }
379 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
380 if (mms.is_empty()) {
381 // get a copy of the base memory, and patch some inputs into it
382 const TypePtr* adr_type = mms.adr_type(C);
383 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
384 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
385 mms.set_memory(phi);
386 // Prepare to append interesting stuff onto the newly sliced phi:
387 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
388 }
389 // Append stuff from ex_map:
390 if (add_multiple) {
391 add_n_reqs(mms.memory(), mms.memory2());
392 } else {
393 add_one_req(mms.memory(), mms.memory2());
394 }
395 }
396 uint limit = ex_map->req();
397 for (uint i = TypeFunc::Parms; i < limit; i++) {
398 // Skip everything in the JVMS after tos. (The ex_oop follows.)
399 if (i == tos) i = ex_jvms->monoff();
400 Node* src = ex_map->in(i);
401 Node* dst = phi_map->in(i);
402 if (src != dst) {
403 PhiNode* phi;
404 if (dst->in(0) != region) {
405 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
406 record_for_igvn(phi);
407 _gvn.set_type(phi, phi->type());
408 phi_map->set_req(i, dst);
409 // Prepare to append interesting stuff onto the new phi:
410 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
411 } else {
412 assert(dst->is_Phi(), "nobody else uses a hidden region");
413 phi = dst->as_Phi();
414 }
415 if (add_multiple && src->in(0) == ex_control) {
416 // Both are phis.
417 add_n_reqs(dst, src);
418 } else {
419 while (dst->req() < region->req()) add_one_req(dst, src);
420 }
421 const Type* srctype = _gvn.type(src);
422 if (phi->type() != srctype) {
423 const Type* dsttype = phi->type()->meet_speculative(srctype);
424 if (phi->type() != dsttype) {
425 phi->set_type(dsttype);
426 _gvn.set_type(phi, dsttype);
427 }
428 }
429 }
430 }
431 }
432
433 //--------------------------use_exception_state--------------------------------
434 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
435 if (failing()) { stop(); return top(); }
436 Node* region = phi_map->control();
437 Node* hidden_merge_mark = root();
438 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
439 Node* ex_oop = clear_saved_ex_oop(phi_map);
440 if (region->in(0) == hidden_merge_mark) {
441 // Special marking for internal ex-states. Process the phis now.
442 region->set_req(0, region); // now it's an ordinary region
443 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
444 // Note: Setting the jvms also sets the bci and sp.
445 set_control(_gvn.transform(region));
446 uint tos = jvms()->stkoff() + sp();
447 for (uint i = 1; i < tos; i++) {
448 Node* x = phi_map->in(i);
449 if (x->in(0) == region) {
450 assert(x->is_Phi(), "expected a special phi");
451 phi_map->set_req(i, _gvn.transform(x));
452 }
453 }
454 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
455 Node* x = mms.memory();
456 if (x->in(0) == region) {
457 assert(x->is_Phi(), "nobody else uses a hidden region");
458 mms.set_memory(_gvn.transform(x));
459 }
460 }
461 if (ex_oop->in(0) == region) {
462 assert(ex_oop->is_Phi(), "expected a special phi");
463 ex_oop = _gvn.transform(ex_oop);
464 }
465 } else {
466 set_jvms(phi_map->jvms());
467 }
468
469 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
470 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
471 return ex_oop;
472 }
473
474 //---------------------------------java_bc-------------------------------------
475 Bytecodes::Code GraphKit::java_bc() const {
476 ciMethod* method = this->method();
477 int bci = this->bci();
478 if (method != NULL && bci != InvocationEntryBci)
479 return method->java_code_at_bci(bci);
480 else
481 return Bytecodes::_illegal;
482 }
483
484 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
485 bool must_throw) {
486 // if the exception capability is set, then we will generate code
487 // to check the JavaThread.should_post_on_exceptions flag to see
488 // if we actually need to report exception events (for this
489 // thread). If we don't need to report exception events, we will
490 // take the normal fast path provided by add_exception_events. If
491 // exception event reporting is enabled for this thread, we will
492 // take the uncommon_trap in the BuildCutout below.
493
494 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
495 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
496 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
497 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
498
499 // Test the should_post_on_exceptions_flag vs. 0
500 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
501 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
502
503 // Branch to slow_path if should_post_on_exceptions_flag was true
504 { BuildCutout unless(this, tst, PROB_MAX);
505 // Do not try anything fancy if we're notifying the VM on every throw.
506 // Cf. case Bytecodes::_athrow in parse2.cpp.
507 uncommon_trap(reason, Deoptimization::Action_none,
508 (ciKlass*)NULL, (char*)NULL, must_throw);
509 }
510
511 }
512
513 //------------------------------builtin_throw----------------------------------
514 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
515 bool must_throw = true;
516
517 if (env()->jvmti_can_post_on_exceptions()) {
518 // check if we must post exception events, take uncommon trap if so
519 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
520 // here if should_post_on_exceptions is false
521 // continue on with the normal codegen
522 }
523
524 // If this particular condition has not yet happened at this
525 // bytecode, then use the uncommon trap mechanism, and allow for
526 // a future recompilation if several traps occur here.
527 // If the throw is hot, try to use a more complicated inline mechanism
528 // which keeps execution inside the compiled code.
529 bool treat_throw_as_hot = false;
530 ciMethodData* md = method()->method_data();
531
532 if (ProfileTraps) {
533 if (too_many_traps(reason)) {
534 treat_throw_as_hot = true;
535 }
536 // (If there is no MDO at all, assume it is early in
537 // execution, and that any deopts are part of the
538 // startup transient, and don't need to be remembered.)
539
540 // Also, if there is a local exception handler, treat all throws
541 // as hot if there has been at least one in this method.
542 if (C->trap_count(reason) != 0
543 && method()->method_data()->trap_count(reason) != 0
544 && has_ex_handler()) {
545 treat_throw_as_hot = true;
546 }
547 }
548
549 // If this throw happens frequently, an uncommon trap might cause
550 // a performance pothole. If there is a local exception handler,
551 // and if this particular bytecode appears to be deoptimizing often,
552 // let us handle the throw inline, with a preconstructed instance.
553 // Note: If the deopt count has blown up, the uncommon trap
554 // runtime is going to flush this nmethod, not matter what.
555 if (treat_throw_as_hot
556 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
557 // If the throw is local, we use a pre-existing instance and
558 // punt on the backtrace. This would lead to a missing backtrace
559 // (a repeat of 4292742) if the backtrace object is ever asked
560 // for its backtrace.
561 // Fixing this remaining case of 4292742 requires some flavor of
562 // escape analysis. Leave that for the future.
563 ciInstance* ex_obj = NULL;
564 switch (reason) {
565 case Deoptimization::Reason_null_check:
566 ex_obj = env()->NullPointerException_instance();
567 break;
568 case Deoptimization::Reason_div0_check:
569 ex_obj = env()->ArithmeticException_instance();
570 break;
571 case Deoptimization::Reason_range_check:
572 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
573 break;
574 case Deoptimization::Reason_class_check:
575 if (java_bc() == Bytecodes::_aastore) {
576 ex_obj = env()->ArrayStoreException_instance();
577 } else {
578 ex_obj = env()->ClassCastException_instance();
579 }
580 break;
581 }
582 if (failing()) { stop(); return; } // exception allocation might fail
583 if (ex_obj != NULL) {
584 // Cheat with a preallocated exception object.
585 if (C->log() != NULL)
586 C->log()->elem("hot_throw preallocated='1' reason='%s'",
587 Deoptimization::trap_reason_name(reason));
588 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
589 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
590
591 // Clear the detail message of the preallocated exception object.
592 // Weblogic sometimes mutates the detail message of exceptions
593 // using reflection.
594 int offset = java_lang_Throwable::get_detailMessage_offset();
595 const TypePtr* adr_typ = ex_con->add_offset(offset);
596
597 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
598 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
599 // Conservatively release stores of object references.
600 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
601
602 add_exception_state(make_exception_state(ex_node));
603 return;
604 }
605 }
606
607 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
608 // It won't be much cheaper than bailing to the interp., since we'll
609 // have to pass up all the debug-info, and the runtime will have to
610 // create the stack trace.
611
612 // Usual case: Bail to interpreter.
613 // Reserve the right to recompile if we haven't seen anything yet.
614
615 assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
616 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
617 if (treat_throw_as_hot
618 && (method()->method_data()->trap_recompiled_at(bci(), NULL)
619 || C->too_many_traps(reason))) {
620 // We cannot afford to take more traps here. Suffer in the interpreter.
621 if (C->log() != NULL)
622 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
623 Deoptimization::trap_reason_name(reason),
624 C->trap_count(reason));
625 action = Deoptimization::Action_none;
626 }
627
628 // "must_throw" prunes the JVM state to include only the stack, if there
629 // are no local exception handlers. This should cut down on register
630 // allocation time and code size, by drastically reducing the number
631 // of in-edges on the call to the uncommon trap.
632
633 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
634 }
635
636
637 //----------------------------PreserveJVMState---------------------------------
638 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
639 debug_only(kit->verify_map());
640 _kit = kit;
641 _map = kit->map(); // preserve the map
642 _sp = kit->sp();
643 kit->set_map(clone_map ? kit->clone_map() : NULL);
644 Compile::current()->inc_preserve_jvm_state();
645 #ifdef ASSERT
646 _bci = kit->bci();
647 Parse* parser = kit->is_Parse();
648 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
649 _block = block;
650 #endif
651 }
652 PreserveJVMState::~PreserveJVMState() {
653 GraphKit* kit = _kit;
654 #ifdef ASSERT
655 assert(kit->bci() == _bci, "bci must not shift");
656 Parse* parser = kit->is_Parse();
657 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
658 assert(block == _block, "block must not shift");
659 #endif
660 kit->set_map(_map);
661 kit->set_sp(_sp);
662 Compile::current()->dec_preserve_jvm_state();
663 }
664
665
666 //-----------------------------BuildCutout-------------------------------------
667 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
668 : PreserveJVMState(kit)
669 {
670 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
671 SafePointNode* outer_map = _map; // preserved map is caller's
672 SafePointNode* inner_map = kit->map();
673 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
674 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
675 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
676 }
677 BuildCutout::~BuildCutout() {
678 GraphKit* kit = _kit;
679 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
680 }
681
682 //---------------------------PreserveReexecuteState----------------------------
683 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
684 assert(!kit->stopped(), "must call stopped() before");
685 _kit = kit;
686 _sp = kit->sp();
687 _reexecute = kit->jvms()->_reexecute;
688 }
689 PreserveReexecuteState::~PreserveReexecuteState() {
690 if (_kit->stopped()) return;
691 _kit->jvms()->_reexecute = _reexecute;
692 _kit->set_sp(_sp);
693 }
694
695 //------------------------------clone_map--------------------------------------
696 // Implementation of PreserveJVMState
697 //
698 // Only clone_map(...) here. If this function is only used in the
699 // PreserveJVMState class we may want to get rid of this extra
700 // function eventually and do it all there.
701
702 SafePointNode* GraphKit::clone_map() {
703 if (map() == NULL) return NULL;
704
705 // Clone the memory edge first
706 Node* mem = MergeMemNode::make(C, map()->memory());
707 gvn().set_type_bottom(mem);
708
709 SafePointNode *clonemap = (SafePointNode*)map()->clone();
710 JVMState* jvms = this->jvms();
711 JVMState* clonejvms = jvms->clone_shallow(C);
712 clonemap->set_memory(mem);
713 clonemap->set_jvms(clonejvms);
714 clonejvms->set_map(clonemap);
715 record_for_igvn(clonemap);
716 gvn().set_type_bottom(clonemap);
717 return clonemap;
718 }
719
720
721 //-----------------------------set_map_clone-----------------------------------
722 void GraphKit::set_map_clone(SafePointNode* m) {
723 _map = m;
724 _map = clone_map();
725 _map->set_next_exception(NULL);
726 debug_only(verify_map());
727 }
728
729
730 //----------------------------kill_dead_locals---------------------------------
731 // Detect any locals which are known to be dead, and force them to top.
732 void GraphKit::kill_dead_locals() {
733 // Consult the liveness information for the locals. If any
734 // of them are unused, then they can be replaced by top(). This
735 // should help register allocation time and cut down on the size
736 // of the deoptimization information.
737
738 // This call is made from many of the bytecode handling
739 // subroutines called from the Big Switch in do_one_bytecode.
740 // Every bytecode which might include a slow path is responsible
741 // for killing its dead locals. The more consistent we
742 // are about killing deads, the fewer useless phis will be
743 // constructed for them at various merge points.
744
745 // bci can be -1 (InvocationEntryBci). We return the entry
746 // liveness for the method.
747
748 if (method() == NULL || method()->code_size() == 0) {
749 // We are building a graph for a call to a native method.
750 // All locals are live.
751 return;
752 }
753
754 ResourceMark rm;
755
756 // Consult the liveness information for the locals. If any
757 // of them are unused, then they can be replaced by top(). This
758 // should help register allocation time and cut down on the size
759 // of the deoptimization information.
760 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
761
762 int len = (int)live_locals.size();
763 assert(len <= jvms()->loc_size(), "too many live locals");
764 for (int local = 0; local < len; local++) {
765 if (!live_locals.at(local)) {
766 set_local(local, top());
767 }
768 }
769 }
770
771 #ifdef ASSERT
772 //-------------------------dead_locals_are_killed------------------------------
773 // Return true if all dead locals are set to top in the map.
774 // Used to assert "clean" debug info at various points.
775 bool GraphKit::dead_locals_are_killed() {
776 if (method() == NULL || method()->code_size() == 0) {
777 // No locals need to be dead, so all is as it should be.
778 return true;
779 }
780
781 // Make sure somebody called kill_dead_locals upstream.
782 ResourceMark rm;
783 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
784 if (jvms->loc_size() == 0) continue; // no locals to consult
785 SafePointNode* map = jvms->map();
786 ciMethod* method = jvms->method();
787 int bci = jvms->bci();
788 if (jvms == this->jvms()) {
789 bci = this->bci(); // it might not yet be synched
790 }
791 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
792 int len = (int)live_locals.size();
793 if (!live_locals.is_valid() || len == 0)
794 // This method is trivial, or is poisoned by a breakpoint.
795 return true;
796 assert(len == jvms->loc_size(), "live map consistent with locals map");
797 for (int local = 0; local < len; local++) {
798 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
799 if (PrintMiscellaneous && (Verbose || WizardMode)) {
800 tty->print_cr("Zombie local %d: ", local);
801 jvms->dump();
802 }
803 return false;
804 }
805 }
806 }
807 return true;
808 }
809
810 #endif //ASSERT
811
812 // Helper function for enforcing certain bytecodes to reexecute if
813 // deoptimization happens
814 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
815 ciMethod* cur_method = jvms->method();
816 int cur_bci = jvms->bci();
817 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
818 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
819 return Interpreter::bytecode_should_reexecute(code) ||
820 is_anewarray && code == Bytecodes::_multianewarray;
821 // Reexecute _multianewarray bytecode which was replaced with
822 // sequence of [a]newarray. See Parse::do_multianewarray().
823 //
824 // Note: interpreter should not have it set since this optimization
825 // is limited by dimensions and guarded by flag so in some cases
826 // multianewarray() runtime calls will be generated and
827 // the bytecode should not be reexecutes (stack will not be reset).
828 } else
829 return false;
830 }
831
832 // Helper function for adding JVMState and debug information to node
833 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
834 // Add the safepoint edges to the call (or other safepoint).
835
836 // Make sure dead locals are set to top. This
837 // should help register allocation time and cut down on the size
838 // of the deoptimization information.
839 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
840
841 // Walk the inline list to fill in the correct set of JVMState's
842 // Also fill in the associated edges for each JVMState.
843
844 // If the bytecode needs to be reexecuted we need to put
845 // the arguments back on the stack.
846 const bool should_reexecute = jvms()->should_reexecute();
847 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
848
849 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
850 // undefined if the bci is different. This is normal for Parse but it
851 // should not happen for LibraryCallKit because only one bci is processed.
852 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
853 "in LibraryCallKit the reexecute bit should not change");
854
855 // If we are guaranteed to throw, we can prune everything but the
856 // input to the current bytecode.
857 bool can_prune_locals = false;
858 uint stack_slots_not_pruned = 0;
859 int inputs = 0, depth = 0;
860 if (must_throw) {
861 assert(method() == youngest_jvms->method(), "sanity");
862 if (compute_stack_effects(inputs, depth)) {
863 can_prune_locals = true;
864 stack_slots_not_pruned = inputs;
865 }
866 }
867
868 if (env()->should_retain_local_variables()) {
869 // At any safepoint, this method can get breakpointed, which would
870 // then require an immediate deoptimization.
871 can_prune_locals = false; // do not prune locals
872 stack_slots_not_pruned = 0;
873 }
874
875 // do not scribble on the input jvms
876 JVMState* out_jvms = youngest_jvms->clone_deep(C);
877 call->set_jvms(out_jvms); // Start jvms list for call node
878
879 // For a known set of bytecodes, the interpreter should reexecute them if
880 // deoptimization happens. We set the reexecute state for them here
881 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
882 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
883 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
884 }
885
886 // Presize the call:
887 DEBUG_ONLY(uint non_debug_edges = call->req());
888 call->add_req_batch(top(), youngest_jvms->debug_depth());
889 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
890
891 // Set up edges so that the call looks like this:
892 // Call [state:] ctl io mem fptr retadr
893 // [parms:] parm0 ... parmN
894 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
895 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
896 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
897 // Note that caller debug info precedes callee debug info.
898
899 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
900 uint debug_ptr = call->req();
901
902 // Loop over the map input edges associated with jvms, add them
903 // to the call node, & reset all offsets to match call node array.
904 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
905 uint debug_end = debug_ptr;
906 uint debug_start = debug_ptr - in_jvms->debug_size();
907 debug_ptr = debug_start; // back up the ptr
908
909 uint p = debug_start; // walks forward in [debug_start, debug_end)
910 uint j, k, l;
911 SafePointNode* in_map = in_jvms->map();
912 out_jvms->set_map(call);
913
914 if (can_prune_locals) {
915 assert(in_jvms->method() == out_jvms->method(), "sanity");
916 // If the current throw can reach an exception handler in this JVMS,
917 // then we must keep everything live that can reach that handler.
918 // As a quick and dirty approximation, we look for any handlers at all.
919 if (in_jvms->method()->has_exception_handlers()) {
920 can_prune_locals = false;
921 }
922 }
923
924 // Add the Locals
925 k = in_jvms->locoff();
926 l = in_jvms->loc_size();
927 out_jvms->set_locoff(p);
928 if (!can_prune_locals) {
929 for (j = 0; j < l; j++)
930 call->set_req(p++, in_map->in(k+j));
931 } else {
932 p += l; // already set to top above by add_req_batch
933 }
934
935 // Add the Expression Stack
936 k = in_jvms->stkoff();
937 l = in_jvms->sp();
938 out_jvms->set_stkoff(p);
939 if (!can_prune_locals) {
940 for (j = 0; j < l; j++)
941 call->set_req(p++, in_map->in(k+j));
942 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
943 // Divide stack into {S0,...,S1}, where S0 is set to top.
944 uint s1 = stack_slots_not_pruned;
945 stack_slots_not_pruned = 0; // for next iteration
946 if (s1 > l) s1 = l;
947 uint s0 = l - s1;
948 p += s0; // skip the tops preinstalled by add_req_batch
949 for (j = s0; j < l; j++)
950 call->set_req(p++, in_map->in(k+j));
951 } else {
952 p += l; // already set to top above by add_req_batch
953 }
954
955 // Add the Monitors
956 k = in_jvms->monoff();
957 l = in_jvms->mon_size();
958 out_jvms->set_monoff(p);
959 for (j = 0; j < l; j++)
960 call->set_req(p++, in_map->in(k+j));
961
962 // Copy any scalar object fields.
963 k = in_jvms->scloff();
964 l = in_jvms->scl_size();
965 out_jvms->set_scloff(p);
966 for (j = 0; j < l; j++)
967 call->set_req(p++, in_map->in(k+j));
968
969 // Finish the new jvms.
970 out_jvms->set_endoff(p);
971
972 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
973 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
974 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
975 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
976 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
977 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
978
979 // Update the two tail pointers in parallel.
980 out_jvms = out_jvms->caller();
981 in_jvms = in_jvms->caller();
982 }
983
984 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
985
986 // Test the correctness of JVMState::debug_xxx accessors:
987 assert(call->jvms()->debug_start() == non_debug_edges, "");
988 assert(call->jvms()->debug_end() == call->req(), "");
989 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
990 }
991
992 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
993 Bytecodes::Code code = java_bc();
994 if (code == Bytecodes::_wide) {
995 code = method()->java_code_at_bci(bci() + 1);
996 }
997
998 BasicType rtype = T_ILLEGAL;
999 int rsize = 0;
1000
1001 if (code != Bytecodes::_illegal) {
1002 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1003 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1004 if (rtype < T_CONFLICT)
1005 rsize = type2size[rtype];
1006 }
1007
1008 switch (code) {
1009 case Bytecodes::_illegal:
1010 return false;
1011
1012 case Bytecodes::_ldc:
1013 case Bytecodes::_ldc_w:
1014 case Bytecodes::_ldc2_w:
1015 inputs = 0;
1016 break;
1017
1018 case Bytecodes::_dup: inputs = 1; break;
1019 case Bytecodes::_dup_x1: inputs = 2; break;
1020 case Bytecodes::_dup_x2: inputs = 3; break;
1021 case Bytecodes::_dup2: inputs = 2; break;
1022 case Bytecodes::_dup2_x1: inputs = 3; break;
1023 case Bytecodes::_dup2_x2: inputs = 4; break;
1024 case Bytecodes::_swap: inputs = 2; break;
1025 case Bytecodes::_arraylength: inputs = 1; break;
1026
1027 case Bytecodes::_getstatic:
1028 case Bytecodes::_putstatic:
1029 case Bytecodes::_getfield:
1030 case Bytecodes::_putfield:
1031 {
1032 bool ignored_will_link;
1033 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1034 int size = field->type()->size();
1035 bool is_get = (depth >= 0), is_static = (depth & 1);
1036 inputs = (is_static ? 0 : 1);
1037 if (is_get) {
1038 depth = size - inputs;
1039 } else {
1040 inputs += size; // putxxx pops the value from the stack
1041 depth = - inputs;
1042 }
1043 }
1044 break;
1045
1046 case Bytecodes::_invokevirtual:
1047 case Bytecodes::_invokespecial:
1048 case Bytecodes::_invokestatic:
1049 case Bytecodes::_invokedynamic:
1050 case Bytecodes::_invokeinterface:
1051 {
1052 bool ignored_will_link;
1053 ciSignature* declared_signature = NULL;
1054 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1055 assert(declared_signature != NULL, "cannot be null");
1056 inputs = declared_signature->arg_size_for_bc(code);
1057 int size = declared_signature->return_type()->size();
1058 depth = size - inputs;
1059 }
1060 break;
1061
1062 case Bytecodes::_multianewarray:
1063 {
1064 ciBytecodeStream iter(method());
1065 iter.reset_to_bci(bci());
1066 iter.next();
1067 inputs = iter.get_dimensions();
1068 assert(rsize == 1, "");
1069 depth = rsize - inputs;
1070 }
1071 break;
1072
1073 case Bytecodes::_ireturn:
1074 case Bytecodes::_lreturn:
1075 case Bytecodes::_freturn:
1076 case Bytecodes::_dreturn:
1077 case Bytecodes::_areturn:
1078 assert(rsize = -depth, "");
1079 inputs = rsize;
1080 break;
1081
1082 case Bytecodes::_jsr:
1083 case Bytecodes::_jsr_w:
1084 inputs = 0;
1085 depth = 1; // S.B. depth=1, not zero
1086 break;
1087
1088 default:
1089 // bytecode produces a typed result
1090 inputs = rsize - depth;
1091 assert(inputs >= 0, "");
1092 break;
1093 }
1094
1095 #ifdef ASSERT
1096 // spot check
1097 int outputs = depth + inputs;
1098 assert(outputs >= 0, "sanity");
1099 switch (code) {
1100 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1101 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1102 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1103 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1104 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1105 }
1106 #endif //ASSERT
1107
1108 return true;
1109 }
1110
1111
1112
1113 //------------------------------basic_plus_adr---------------------------------
1114 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1115 // short-circuit a common case
1116 if (offset == intcon(0)) return ptr;
1117 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1118 }
1119
1120 Node* GraphKit::ConvI2L(Node* offset) {
1121 // short-circuit a common case
1122 jint offset_con = find_int_con(offset, Type::OffsetBot);
1123 if (offset_con != Type::OffsetBot) {
1124 return longcon((jlong) offset_con);
1125 }
1126 return _gvn.transform( new (C) ConvI2LNode(offset));
1127 }
1128
1129 Node* GraphKit::ConvI2UL(Node* offset) {
1130 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1131 if (offset_con != (juint) Type::OffsetBot) {
1132 return longcon((julong) offset_con);
1133 }
1134 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset));
1135 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) );
1136 return _gvn.transform( new (C) AndLNode(conv, mask) );
1137 }
1138
1139 Node* GraphKit::ConvL2I(Node* offset) {
1140 // short-circuit a common case
1141 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1142 if (offset_con != (jlong)Type::OffsetBot) {
1143 return intcon((int) offset_con);
1144 }
1145 return _gvn.transform( new (C) ConvL2INode(offset));
1146 }
1147
1148 //-------------------------load_object_klass-----------------------------------
1149 Node* GraphKit::load_object_klass(Node* obj) {
1150 // Special-case a fresh allocation to avoid building nodes:
1151 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1152 if (akls != NULL) return akls;
1153 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1154 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1155 }
1156
1157 //-------------------------load_array_length-----------------------------------
1158 Node* GraphKit::load_array_length(Node* array) {
1159 // Special-case a fresh allocation to avoid building nodes:
1160 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1161 Node *alen;
1162 if (alloc == NULL) {
1163 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1164 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1165 } else {
1166 alen = alloc->Ideal_length();
1167 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1168 if (ccast != alen) {
1169 alen = _gvn.transform(ccast);
1170 }
1171 }
1172 return alen;
1173 }
1174
1175 //------------------------------do_null_check----------------------------------
1176 // Helper function to do a NULL pointer check. Returned value is
1177 // the incoming address with NULL casted away. You are allowed to use the
1178 // not-null value only if you are control dependent on the test.
1179 extern int explicit_null_checks_inserted,
1180 explicit_null_checks_elided;
1181 Node* GraphKit::null_check_common(Node* value, BasicType type,
1182 // optional arguments for variations:
1183 bool assert_null,
1184 Node* *null_control) {
1185 assert(!assert_null || null_control == NULL, "not both at once");
1186 if (stopped()) return top();
1187 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1188 // For some performance testing, we may wish to suppress null checking.
1189 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1190 return value;
1191 }
1192 explicit_null_checks_inserted++;
1193
1194 // Construct NULL check
1195 Node *chk = NULL;
1196 switch(type) {
1197 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1198 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1199 case T_ARRAY : // fall through
1200 type = T_OBJECT; // simplify further tests
1201 case T_OBJECT : {
1202 const Type *t = _gvn.type( value );
1203
1204 const TypeOopPtr* tp = t->isa_oopptr();
1205 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1206 // Only for do_null_check, not any of its siblings:
1207 && !assert_null && null_control == NULL) {
1208 // Usually, any field access or invocation on an unloaded oop type
1209 // will simply fail to link, since the statically linked class is
1210 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1211 // the static class is loaded but the sharper oop type is not.
1212 // Rather than checking for this obscure case in lots of places,
1213 // we simply observe that a null check on an unloaded class
1214 // will always be followed by a nonsense operation, so we
1215 // can just issue the uncommon trap here.
1216 // Our access to the unloaded class will only be correct
1217 // after it has been loaded and initialized, which requires
1218 // a trip through the interpreter.
1219 #ifndef PRODUCT
1220 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1221 #endif
1222 uncommon_trap(Deoptimization::Reason_unloaded,
1223 Deoptimization::Action_reinterpret,
1224 tp->klass(), "!loaded");
1225 return top();
1226 }
1227
1228 if (assert_null) {
1229 // See if the type is contained in NULL_PTR.
1230 // If so, then the value is already null.
1231 if (t->higher_equal(TypePtr::NULL_PTR)) {
1232 explicit_null_checks_elided++;
1233 return value; // Elided null assert quickly!
1234 }
1235 } else {
1236 // See if mixing in the NULL pointer changes type.
1237 // If so, then the NULL pointer was not allowed in the original
1238 // type. In other words, "value" was not-null.
1239 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1240 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1241 explicit_null_checks_elided++;
1242 return value; // Elided null check quickly!
1243 }
1244 }
1245 chk = new (C) CmpPNode( value, null() );
1246 break;
1247 }
1248
1249 default:
1250 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1251 }
1252 assert(chk != NULL, "sanity check");
1253 chk = _gvn.transform(chk);
1254
1255 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1256 BoolNode *btst = new (C) BoolNode( chk, btest);
1257 Node *tst = _gvn.transform( btst );
1258
1259 //-----------
1260 // if peephole optimizations occurred, a prior test existed.
1261 // If a prior test existed, maybe it dominates as we can avoid this test.
1262 if (tst != btst && type == T_OBJECT) {
1263 // At this point we want to scan up the CFG to see if we can
1264 // find an identical test (and so avoid this test altogether).
1265 Node *cfg = control();
1266 int depth = 0;
1267 while( depth < 16 ) { // Limit search depth for speed
1268 if( cfg->Opcode() == Op_IfTrue &&
1269 cfg->in(0)->in(1) == tst ) {
1270 // Found prior test. Use "cast_not_null" to construct an identical
1271 // CastPP (and hence hash to) as already exists for the prior test.
1272 // Return that casted value.
1273 if (assert_null) {
1274 replace_in_map(value, null());
1275 return null(); // do not issue the redundant test
1276 }
1277 Node *oldcontrol = control();
1278 set_control(cfg);
1279 Node *res = cast_not_null(value);
1280 set_control(oldcontrol);
1281 explicit_null_checks_elided++;
1282 return res;
1283 }
1284 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1285 if (cfg == NULL) break; // Quit at region nodes
1286 depth++;
1287 }
1288 }
1289
1290 //-----------
1291 // Branch to failure if null
1292 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1293 Deoptimization::DeoptReason reason;
1294 if (assert_null)
1295 reason = Deoptimization::Reason_null_assert;
1296 else if (type == T_OBJECT)
1297 reason = Deoptimization::Reason_null_check;
1298 else
1299 reason = Deoptimization::Reason_div0_check;
1300
1301 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1302 // ciMethodData::has_trap_at will return a conservative -1 if any
1303 // must-be-null assertion has failed. This could cause performance
1304 // problems for a method after its first do_null_assert failure.
1305 // Consider using 'Reason_class_check' instead?
1306
1307 // To cause an implicit null check, we set the not-null probability
1308 // to the maximum (PROB_MAX). For an explicit check the probability
1309 // is set to a smaller value.
1310 if (null_control != NULL || too_many_traps(reason)) {
1311 // probability is less likely
1312 ok_prob = PROB_LIKELY_MAG(3);
1313 } else if (!assert_null &&
1314 (ImplicitNullCheckThreshold > 0) &&
1315 method() != NULL &&
1316 (method()->method_data()->trap_count(reason)
1317 >= (uint)ImplicitNullCheckThreshold)) {
1318 ok_prob = PROB_LIKELY_MAG(3);
1319 }
1320
1321 if (null_control != NULL) {
1322 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1323 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1324 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1325 if (null_true == top())
1326 explicit_null_checks_elided++;
1327 (*null_control) = null_true;
1328 } else {
1329 BuildCutout unless(this, tst, ok_prob);
1330 // Check for optimizer eliding test at parse time
1331 if (stopped()) {
1332 // Failure not possible; do not bother making uncommon trap.
1333 explicit_null_checks_elided++;
1334 } else if (assert_null) {
1335 uncommon_trap(reason,
1336 Deoptimization::Action_make_not_entrant,
1337 NULL, "assert_null");
1338 } else {
1339 replace_in_map(value, zerocon(type));
1340 builtin_throw(reason);
1341 }
1342 }
1343
1344 // Must throw exception, fall-thru not possible?
1345 if (stopped()) {
1346 return top(); // No result
1347 }
1348
1349 if (assert_null) {
1350 // Cast obj to null on this path.
1351 replace_in_map(value, zerocon(type));
1352 return zerocon(type);
1353 }
1354
1355 // Cast obj to not-null on this path, if there is no null_control.
1356 // (If there is a null_control, a non-null value may come back to haunt us.)
1357 if (type == T_OBJECT) {
1358 Node* cast = cast_not_null(value, false);
1359 if (null_control == NULL || (*null_control) == top())
1360 replace_in_map(value, cast);
1361 value = cast;
1362 }
1363
1364 return value;
1365 }
1366
1367
1368 //------------------------------cast_not_null----------------------------------
1369 // Cast obj to not-null on this path
1370 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1371 const Type *t = _gvn.type(obj);
1372 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1373 // Object is already not-null?
1374 if( t == t_not_null ) return obj;
1375
1376 Node *cast = new (C) CastPPNode(obj,t_not_null);
1377 cast->init_req(0, control());
1378 cast = _gvn.transform( cast );
1379
1380 // Scan for instances of 'obj' in the current JVM mapping.
1381 // These instances are known to be not-null after the test.
1382 if (do_replace_in_map)
1383 replace_in_map(obj, cast);
1384
1385 return cast; // Return casted value
1386 }
1387
1388
1389 //--------------------------replace_in_map-------------------------------------
1390 void GraphKit::replace_in_map(Node* old, Node* neww) {
1391 if (old == neww) {
1392 return;
1393 }
1394
1395 map()->replace_edge(old, neww);
1396
1397 // Note: This operation potentially replaces any edge
1398 // on the map. This includes locals, stack, and monitors
1399 // of the current (innermost) JVM state.
1400
1401 if (!ReplaceInParentMaps) {
1402 return;
1403 }
1404
1405 // PreserveJVMState doesn't do a deep copy so we can't modify
1406 // parents
1407 if (Compile::current()->has_preserve_jvm_state()) {
1408 return;
1409 }
1410
1411 Parse* parser = is_Parse();
1412 bool progress = true;
1413 Node* ctrl = map()->in(0);
1414 // Follow the chain of parsers and see whether the update can be
1415 // done in the map of callers. We can do the replace for a caller if
1416 // the current control post dominates the control of a caller.
1417 while (parser != NULL && parser->caller() != NULL && progress) {
1418 progress = false;
1419 Node* parent_map = parser->caller()->map();
1420 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch");
1421
1422 Node* parent_ctrl = parent_map->in(0);
1423
1424 while (parent_ctrl->is_Region()) {
1425 Node* n = parent_ctrl->as_Region()->is_copy();
1426 if (n == NULL) {
1427 break;
1428 }
1429 parent_ctrl = n;
1430 }
1431
1432 for (;;) {
1433 if (ctrl == parent_ctrl) {
1434 // update the map of the exits which is the one that will be
1435 // used when compilation resume after inlining
1436 parser->exits().map()->replace_edge(old, neww);
1437 progress = true;
1438 break;
1439 }
1440 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
1441 ctrl = ctrl->in(0)->in(0);
1442 } else if (ctrl->is_Region()) {
1443 Node* n = ctrl->as_Region()->is_copy();
1444 if (n == NULL) {
1445 break;
1446 }
1447 ctrl = n;
1448 } else {
1449 break;
1450 }
1451 }
1452
1453 parser = parser->parent_parser();
1454 }
1455 }
1456
1457
1458 //=============================================================================
1459 //--------------------------------memory---------------------------------------
1460 Node* GraphKit::memory(uint alias_idx) {
1461 MergeMemNode* mem = merged_memory();
1462 Node* p = mem->memory_at(alias_idx);
1463 _gvn.set_type(p, Type::MEMORY); // must be mapped
1464 return p;
1465 }
1466
1467 //-----------------------------reset_memory------------------------------------
1468 Node* GraphKit::reset_memory() {
1469 Node* mem = map()->memory();
1470 // do not use this node for any more parsing!
1471 debug_only( map()->set_memory((Node*)NULL) );
1472 return _gvn.transform( mem );
1473 }
1474
1475 //------------------------------set_all_memory---------------------------------
1476 void GraphKit::set_all_memory(Node* newmem) {
1477 Node* mergemem = MergeMemNode::make(C, newmem);
1478 gvn().set_type_bottom(mergemem);
1479 map()->set_memory(mergemem);
1480 }
1481
1482 //------------------------------set_all_memory_call----------------------------
1483 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1484 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1485 set_all_memory(newmem);
1486 }
1487
1488 //=============================================================================
1489 //
1490 // parser factory methods for MemNodes
1491 //
1492 // These are layered on top of the factory methods in LoadNode and StoreNode,
1493 // and integrate with the parser's memory state and _gvn engine.
1494 //
1495
1496 // factory methods in "int adr_idx"
1497 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1498 int adr_idx,
1499 MemNode::MemOrd mo, bool require_atomic_access) {
1500 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1501 const TypePtr* adr_type = NULL; // debug-mode-only argument
1502 debug_only(adr_type = C->get_adr_type(adr_idx));
1503 Node* mem = memory(adr_idx);
1504 Node* ld;
1505 if (require_atomic_access && bt == T_LONG) {
1506 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo);
1507 } else {
1508 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo);
1509 }
1510 ld = _gvn.transform(ld);
1511 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1512 // Improve graph before escape analysis and boxing elimination.
1513 record_for_igvn(ld);
1514 }
1515 return ld;
1516 }
1517
1518 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1519 int adr_idx,
1520 MemNode::MemOrd mo,
1521 bool require_atomic_access) {
1522 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1523 const TypePtr* adr_type = NULL;
1524 debug_only(adr_type = C->get_adr_type(adr_idx));
1525 Node *mem = memory(adr_idx);
1526 Node* st;
1527 if (require_atomic_access && bt == T_LONG) {
1528 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1529 } else {
1530 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1531 }
1532 st = _gvn.transform(st);
1533 set_memory(st, adr_idx);
1534 // Back-to-back stores can only remove intermediate store with DU info
1535 // so push on worklist for optimizer.
1536 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1537 record_for_igvn(st);
1538
1539 return st;
1540 }
1541
1542
1543 void GraphKit::pre_barrier(bool do_load,
1544 Node* ctl,
1545 Node* obj,
1546 Node* adr,
1547 uint adr_idx,
1548 Node* val,
1549 const TypeOopPtr* val_type,
1550 Node* pre_val,
1551 BasicType bt) {
1552
1553 BarrierSet* bs = Universe::heap()->barrier_set();
1554 set_control(ctl);
1555 switch (bs->kind()) {
1556 case BarrierSet::G1SATBCT:
1557 case BarrierSet::G1SATBCTLogging:
1558 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1559 break;
1560
1561 case BarrierSet::CardTableModRef:
1562 case BarrierSet::CardTableExtension:
1563 case BarrierSet::ModRef:
1564 break;
1565
1566 case BarrierSet::Other:
1567 default :
1568 ShouldNotReachHere();
1569
1570 }
1571 }
1572
1573 bool GraphKit::can_move_pre_barrier() const {
1574 BarrierSet* bs = Universe::heap()->barrier_set();
1575 switch (bs->kind()) {
1576 case BarrierSet::G1SATBCT:
1577 case BarrierSet::G1SATBCTLogging:
1578 return true; // Can move it if no safepoint
1579
1580 case BarrierSet::CardTableModRef:
1581 case BarrierSet::CardTableExtension:
1582 case BarrierSet::ModRef:
1583 return true; // There is no pre-barrier
1584
1585 case BarrierSet::Other:
1586 default :
1587 ShouldNotReachHere();
1588 }
1589 return false;
1590 }
1591
1592 void GraphKit::post_barrier(Node* ctl,
1593 Node* store,
1594 Node* obj,
1595 Node* adr,
1596 uint adr_idx,
1597 Node* val,
1598 BasicType bt,
1599 bool use_precise) {
1600 BarrierSet* bs = Universe::heap()->barrier_set();
1601 set_control(ctl);
1602 switch (bs->kind()) {
1603 case BarrierSet::G1SATBCT:
1604 case BarrierSet::G1SATBCTLogging:
1605 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1606 break;
1607
1608 case BarrierSet::CardTableModRef:
1609 case BarrierSet::CardTableExtension:
1610 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1611 break;
1612
1613 case BarrierSet::ModRef:
1614 break;
1615
1616 case BarrierSet::Other:
1617 default :
1618 ShouldNotReachHere();
1619
1620 }
1621 }
1622
1623 Node* GraphKit::store_oop(Node* ctl,
1624 Node* obj,
1625 Node* adr,
1626 const TypePtr* adr_type,
1627 Node* val,
1628 const TypeOopPtr* val_type,
1629 BasicType bt,
1630 bool use_precise,
1631 MemNode::MemOrd mo) {
1632 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1633 // could be delayed during Parse (for example, in adjust_map_after_if()).
1634 // Execute transformation here to avoid barrier generation in such case.
1635 if (_gvn.type(val) == TypePtr::NULL_PTR)
1636 val = _gvn.makecon(TypePtr::NULL_PTR);
1637
1638 set_control(ctl);
1639 if (stopped()) return top(); // Dead path ?
1640
1641 assert(bt == T_OBJECT, "sanity");
1642 assert(val != NULL, "not dead path");
1643 uint adr_idx = C->get_alias_index(adr_type);
1644 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1645
1646 pre_barrier(true /* do_load */,
1647 control(), obj, adr, adr_idx, val, val_type,
1648 NULL /* pre_val */,
1649 bt);
1650
1651 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo);
1652 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1653 return store;
1654 }
1655
1656 // Could be an array or object we don't know at compile time (unsafe ref.)
1657 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1658 Node* obj, // containing obj
1659 Node* adr, // actual adress to store val at
1660 const TypePtr* adr_type,
1661 Node* val,
1662 BasicType bt,
1663 MemNode::MemOrd mo) {
1664 Compile::AliasType* at = C->alias_type(adr_type);
1665 const TypeOopPtr* val_type = NULL;
1666 if (adr_type->isa_instptr()) {
1667 if (at->field() != NULL) {
1668 // known field. This code is a copy of the do_put_xxx logic.
1669 ciField* field = at->field();
1670 if (!field->type()->is_loaded()) {
1671 val_type = TypeInstPtr::BOTTOM;
1672 } else {
1673 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1674 }
1675 }
1676 } else if (adr_type->isa_aryptr()) {
1677 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1678 }
1679 if (val_type == NULL) {
1680 val_type = TypeInstPtr::BOTTOM;
1681 }
1682 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo);
1683 }
1684
1685
1686 //-------------------------array_element_address-------------------------
1687 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1688 const TypeInt* sizetype) {
1689 uint shift = exact_log2(type2aelembytes(elembt));
1690 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1691
1692 // short-circuit a common case (saves lots of confusing waste motion)
1693 jint idx_con = find_int_con(idx, -1);
1694 if (idx_con >= 0) {
1695 intptr_t offset = header + ((intptr_t)idx_con << shift);
1696 return basic_plus_adr(ary, offset);
1697 }
1698
1699 // must be correct type for alignment purposes
1700 Node* base = basic_plus_adr(ary, header);
1701 #ifdef _LP64
1702 // The scaled index operand to AddP must be a clean 64-bit value.
1703 // Java allows a 32-bit int to be incremented to a negative
1704 // value, which appears in a 64-bit register as a large
1705 // positive number. Using that large positive number as an
1706 // operand in pointer arithmetic has bad consequences.
1707 // On the other hand, 32-bit overflow is rare, and the possibility
1708 // can often be excluded, if we annotate the ConvI2L node with
1709 // a type assertion that its value is known to be a small positive
1710 // number. (The prior range check has ensured this.)
1711 // This assertion is used by ConvI2LNode::Ideal.
1712 int index_max = max_jint - 1; // array size is max_jint, index is one less
1713 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1714 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1715 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1716 #endif
1717 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1718 return basic_plus_adr(ary, base, scale);
1719 }
1720
1721 //-------------------------load_array_element-------------------------
1722 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1723 const Type* elemtype = arytype->elem();
1724 BasicType elembt = elemtype->array_element_basic_type();
1725 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1726 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1727 return ld;
1728 }
1729
1730 //-------------------------set_arguments_for_java_call-------------------------
1731 // Arguments (pre-popped from the stack) are taken from the JVMS.
1732 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1733 // Add the call arguments:
1734 uint nargs = call->method()->arg_size();
1735 for (uint i = 0; i < nargs; i++) {
1736 Node* arg = argument(i);
1737 call->init_req(i + TypeFunc::Parms, arg);
1738 }
1739 }
1740
1741 //---------------------------set_edges_for_java_call---------------------------
1742 // Connect a newly created call into the current JVMS.
1743 // A return value node (if any) is returned from set_edges_for_java_call.
1744 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1745
1746 // Add the predefined inputs:
1747 call->init_req( TypeFunc::Control, control() );
1748 call->init_req( TypeFunc::I_O , i_o() );
1749 call->init_req( TypeFunc::Memory , reset_memory() );
1750 call->init_req( TypeFunc::FramePtr, frameptr() );
1751 call->init_req( TypeFunc::ReturnAdr, top() );
1752
1753 add_safepoint_edges(call, must_throw);
1754
1755 Node* xcall = _gvn.transform(call);
1756
1757 if (xcall == top()) {
1758 set_control(top());
1759 return;
1760 }
1761 assert(xcall == call, "call identity is stable");
1762
1763 // Re-use the current map to produce the result.
1764
1765 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1766 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1767 set_all_memory_call(xcall, separate_io_proj);
1768
1769 //return xcall; // no need, caller already has it
1770 }
1771
1772 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1773 if (stopped()) return top(); // maybe the call folded up?
1774
1775 // Capture the return value, if any.
1776 Node* ret;
1777 if (call->method() == NULL ||
1778 call->method()->return_type()->basic_type() == T_VOID)
1779 ret = top();
1780 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1781
1782 // Note: Since any out-of-line call can produce an exception,
1783 // we always insert an I_O projection from the call into the result.
1784
1785 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1786
1787 if (separate_io_proj) {
1788 // The caller requested separate projections be used by the fall
1789 // through and exceptional paths, so replace the projections for
1790 // the fall through path.
1791 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1792 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1793 }
1794 return ret;
1795 }
1796
1797 //--------------------set_predefined_input_for_runtime_call--------------------
1798 // Reading and setting the memory state is way conservative here.
1799 // The real problem is that I am not doing real Type analysis on memory,
1800 // so I cannot distinguish card mark stores from other stores. Across a GC
1801 // point the Store Barrier and the card mark memory has to agree. I cannot
1802 // have a card mark store and its barrier split across the GC point from
1803 // either above or below. Here I get that to happen by reading ALL of memory.
1804 // A better answer would be to separate out card marks from other memory.
1805 // For now, return the input memory state, so that it can be reused
1806 // after the call, if this call has restricted memory effects.
1807 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1808 // Set fixed predefined input arguments
1809 Node* memory = reset_memory();
1810 call->init_req( TypeFunc::Control, control() );
1811 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1812 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1813 call->init_req( TypeFunc::FramePtr, frameptr() );
1814 call->init_req( TypeFunc::ReturnAdr, top() );
1815 return memory;
1816 }
1817
1818 //-------------------set_predefined_output_for_runtime_call--------------------
1819 // Set control and memory (not i_o) from the call.
1820 // If keep_mem is not NULL, use it for the output state,
1821 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1822 // If hook_mem is NULL, this call produces no memory effects at all.
1823 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1824 // then only that memory slice is taken from the call.
1825 // In the last case, we must put an appropriate memory barrier before
1826 // the call, so as to create the correct anti-dependencies on loads
1827 // preceding the call.
1828 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1829 Node* keep_mem,
1830 const TypePtr* hook_mem) {
1831 // no i/o
1832 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1833 if (keep_mem) {
1834 // First clone the existing memory state
1835 set_all_memory(keep_mem);
1836 if (hook_mem != NULL) {
1837 // Make memory for the call
1838 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1839 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1840 // We also use hook_mem to extract specific effects from arraycopy stubs.
1841 set_memory(mem, hook_mem);
1842 }
1843 // ...else the call has NO memory effects.
1844
1845 // Make sure the call advertises its memory effects precisely.
1846 // This lets us build accurate anti-dependences in gcm.cpp.
1847 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1848 "call node must be constructed correctly");
1849 } else {
1850 assert(hook_mem == NULL, "");
1851 // This is not a "slow path" call; all memory comes from the call.
1852 set_all_memory_call(call);
1853 }
1854 }
1855
1856
1857 // Replace the call with the current state of the kit.
1858 void GraphKit::replace_call(CallNode* call, Node* result) {
1859 JVMState* ejvms = NULL;
1860 if (has_exceptions()) {
1861 ejvms = transfer_exceptions_into_jvms();
1862 }
1863
1864 SafePointNode* final_state = stop();
1865
1866 // Find all the needed outputs of this call
1867 CallProjections callprojs;
1868 call->extract_projections(&callprojs, true);
1869
1870 Node* init_mem = call->in(TypeFunc::Memory);
1871 Node* final_mem = final_state->in(TypeFunc::Memory);
1872 Node* final_ctl = final_state->in(TypeFunc::Control);
1873 Node* final_io = final_state->in(TypeFunc::I_O);
1874
1875 // Replace all the old call edges with the edges from the inlining result
1876 if (callprojs.fallthrough_catchproj != NULL) {
1877 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1878 }
1879 if (callprojs.fallthrough_memproj != NULL) {
1880 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1881 }
1882 if (callprojs.fallthrough_ioproj != NULL) {
1883 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1884 }
1885
1886 // Replace the result with the new result if it exists and is used
1887 if (callprojs.resproj != NULL && result != NULL) {
1888 C->gvn_replace_by(callprojs.resproj, result);
1889 }
1890
1891 if (ejvms == NULL) {
1892 // No exception edges to simply kill off those paths
1893 if (callprojs.catchall_catchproj != NULL) {
1894 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1895 }
1896 if (callprojs.catchall_memproj != NULL) {
1897 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1898 }
1899 if (callprojs.catchall_ioproj != NULL) {
1900 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1901 }
1902 // Replace the old exception object with top
1903 if (callprojs.exobj != NULL) {
1904 C->gvn_replace_by(callprojs.exobj, C->top());
1905 }
1906 } else {
1907 GraphKit ekit(ejvms);
1908
1909 // Load my combined exception state into the kit, with all phis transformed:
1910 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1911
1912 Node* ex_oop = ekit.use_exception_state(ex_map);
1913 if (callprojs.catchall_catchproj != NULL) {
1914 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1915 }
1916 if (callprojs.catchall_memproj != NULL) {
1917 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1918 }
1919 if (callprojs.catchall_ioproj != NULL) {
1920 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1921 }
1922
1923 // Replace the old exception object with the newly created one
1924 if (callprojs.exobj != NULL) {
1925 C->gvn_replace_by(callprojs.exobj, ex_oop);
1926 }
1927 }
1928
1929 // Disconnect the call from the graph
1930 call->disconnect_inputs(NULL, C);
1931 C->gvn_replace_by(call, C->top());
1932
1933 // Clean up any MergeMems that feed other MergeMems since the
1934 // optimizer doesn't like that.
1935 if (final_mem->is_MergeMem()) {
1936 Node_List wl;
1937 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1938 Node* m = i.get();
1939 if (m->is_MergeMem() && !wl.contains(m)) {
1940 wl.push(m);
1941 }
1942 }
1943 while (wl.size() > 0) {
1944 _gvn.transform(wl.pop());
1945 }
1946 }
1947 }
1948
1949
1950 //------------------------------increment_counter------------------------------
1951 // for statistics: increment a VM counter by 1
1952
1953 void GraphKit::increment_counter(address counter_addr) {
1954 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1955 increment_counter(adr1);
1956 }
1957
1958 void GraphKit::increment_counter(Node* counter_addr) {
1959 int adr_type = Compile::AliasIdxRaw;
1960 Node* ctrl = control();
1961 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1962 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1963 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1964 }
1965
1966
1967 //------------------------------uncommon_trap----------------------------------
1968 // Bail out to the interpreter in mid-method. Implemented by calling the
1969 // uncommon_trap blob. This helper function inserts a runtime call with the
1970 // right debug info.
1971 void GraphKit::uncommon_trap(int trap_request,
1972 ciKlass* klass, const char* comment,
1973 bool must_throw,
1974 bool keep_exact_action) {
1975 if (failing()) stop();
1976 if (stopped()) return; // trap reachable?
1977
1978 // Note: If ProfileTraps is true, and if a deopt. actually
1979 // occurs here, the runtime will make sure an MDO exists. There is
1980 // no need to call method()->ensure_method_data() at this point.
1981
1982 // Set the stack pointer to the right value for reexecution:
1983 set_sp(reexecute_sp());
1984
1985 #ifdef ASSERT
1986 if (!must_throw) {
1987 // Make sure the stack has at least enough depth to execute
1988 // the current bytecode.
1989 int inputs, ignored_depth;
1990 if (compute_stack_effects(inputs, ignored_depth)) {
1991 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1992 Bytecodes::name(java_bc()), sp(), inputs));
1993 }
1994 }
1995 #endif
1996
1997 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1998 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1999
2000 switch (action) {
2001 case Deoptimization::Action_maybe_recompile:
2002 case Deoptimization::Action_reinterpret:
2003 // Temporary fix for 6529811 to allow virtual calls to be sure they
2004 // get the chance to go from mono->bi->mega
2005 if (!keep_exact_action &&
2006 Deoptimization::trap_request_index(trap_request) < 0 &&
2007 too_many_recompiles(reason)) {
2008 // This BCI is causing too many recompilations.
2009 action = Deoptimization::Action_none;
2010 trap_request = Deoptimization::make_trap_request(reason, action);
2011 } else {
2012 C->set_trap_can_recompile(true);
2013 }
2014 break;
2015 case Deoptimization::Action_make_not_entrant:
2016 C->set_trap_can_recompile(true);
2017 break;
2018 #ifdef ASSERT
2019 case Deoptimization::Action_none:
2020 case Deoptimization::Action_make_not_compilable:
2021 break;
2022 default:
2023 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
2024 break;
2025 #endif
2026 }
2027
2028 if (TraceOptoParse) {
2029 char buf[100];
2030 tty->print_cr("Uncommon trap %s at bci:%d",
2031 Deoptimization::format_trap_request(buf, sizeof(buf),
2032 trap_request), bci());
2033 }
2034
2035 CompileLog* log = C->log();
2036 if (log != NULL) {
2037 int kid = (klass == NULL)? -1: log->identify(klass);
2038 log->begin_elem("uncommon_trap bci='%d'", bci());
2039 char buf[100];
2040 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2041 trap_request));
2042 if (kid >= 0) log->print(" klass='%d'", kid);
2043 if (comment != NULL) log->print(" comment='%s'", comment);
2044 log->end_elem();
2045 }
2046
2047 // Make sure any guarding test views this path as very unlikely
2048 Node *i0 = control()->in(0);
2049 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2050 IfNode *iff = i0->as_If();
2051 float f = iff->_prob; // Get prob
2052 if (control()->Opcode() == Op_IfTrue) {
2053 if (f > PROB_UNLIKELY_MAG(4))
2054 iff->_prob = PROB_MIN;
2055 } else {
2056 if (f < PROB_LIKELY_MAG(4))
2057 iff->_prob = PROB_MAX;
2058 }
2059 }
2060
2061 // Clear out dead values from the debug info.
2062 kill_dead_locals();
2063
2064 // Now insert the uncommon trap subroutine call
2065 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2066 const TypePtr* no_memory_effects = NULL;
2067 // Pass the index of the class to be loaded
2068 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2069 (must_throw ? RC_MUST_THROW : 0),
2070 OptoRuntime::uncommon_trap_Type(),
2071 call_addr, "uncommon_trap", no_memory_effects,
2072 intcon(trap_request));
2073 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2074 "must extract request correctly from the graph");
2075 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2076
2077 call->set_req(TypeFunc::ReturnAdr, returnadr());
2078 // The debug info is the only real input to this call.
2079
2080 // Halt-and-catch fire here. The above call should never return!
2081 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2082 _gvn.set_type_bottom(halt);
2083 root()->add_req(halt);
2084
2085 stop_and_kill_map();
2086 }
2087
2088
2089 //--------------------------just_allocated_object------------------------------
2090 // Report the object that was just allocated.
2091 // It must be the case that there are no intervening safepoints.
2092 // We use this to determine if an object is so "fresh" that
2093 // it does not require card marks.
2094 Node* GraphKit::just_allocated_object(Node* current_control) {
2095 if (C->recent_alloc_ctl() == current_control)
2096 return C->recent_alloc_obj();
2097 return NULL;
2098 }
2099
2100
2101 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2102 // (Note: TypeFunc::make has a cache that makes this fast.)
2103 const TypeFunc* tf = TypeFunc::make(dest_method);
2104 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2105 for (int j = 0; j < nargs; j++) {
2106 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2107 if( targ->basic_type() == T_DOUBLE ) {
2108 // If any parameters are doubles, they must be rounded before
2109 // the call, dstore_rounding does gvn.transform
2110 Node *arg = argument(j);
2111 arg = dstore_rounding(arg);
2112 set_argument(j, arg);
2113 }
2114 }
2115 }
2116
2117 /**
2118 * Record profiling data exact_kls for Node n with the type system so
2119 * that it can propagate it (speculation)
2120 *
2121 * @param n node that the type applies to
2122 * @param exact_kls type from profiling
2123 *
2124 * @return node with improved type
2125 */
2126 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2127 const Type* current_type = _gvn.type(n);
2128 assert(UseTypeSpeculation, "type speculation must be on");
2129
2130 const TypeOopPtr* speculative = current_type->speculative();
2131
2132 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2133 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2134 const TypeOopPtr* xtype = tklass->as_instance_type();
2135 assert(xtype->klass_is_exact(), "Should be exact");
2136 // record the new speculative type's depth
2137 speculative = xtype->with_inline_depth(jvms()->depth());
2138 }
2139
2140 if (speculative != current_type->speculative()) {
2141 // Build a type with a speculative type (what we think we know
2142 // about the type but will need a guard when we use it)
2143 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2144 // We're changing the type, we need a new CheckCast node to carry
2145 // the new type. The new type depends on the control: what
2146 // profiling tells us is only valid from here as far as we can
2147 // tell.
2148 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2149 cast = _gvn.transform(cast);
2150 replace_in_map(n, cast);
2151 n = cast;
2152 }
2153
2154 return n;
2155 }
2156
2157 /**
2158 * Record profiling data from receiver profiling at an invoke with the
2159 * type system so that it can propagate it (speculation)
2160 *
2161 * @param n receiver node
2162 *
2163 * @return node with improved type
2164 */
2165 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2166 if (!UseTypeSpeculation) {
2167 return n;
2168 }
2169 ciKlass* exact_kls = profile_has_unique_klass();
2170 return record_profile_for_speculation(n, exact_kls);
2171 }
2172
2173 /**
2174 * Record profiling data from argument profiling at an invoke with the
2175 * type system so that it can propagate it (speculation)
2176 *
2177 * @param dest_method target method for the call
2178 * @param bc what invoke bytecode is this?
2179 */
2180 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2181 if (!UseTypeSpeculation) {
2182 return;
2183 }
2184 const TypeFunc* tf = TypeFunc::make(dest_method);
2185 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2186 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2187 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2188 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2189 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2190 ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2191 if (better_type != NULL) {
2192 record_profile_for_speculation(argument(j), better_type);
2193 }
2194 i++;
2195 }
2196 }
2197 }
2198
2199 /**
2200 * Record profiling data from parameter profiling at an invoke with
2201 * the type system so that it can propagate it (speculation)
2202 */
2203 void GraphKit::record_profiled_parameters_for_speculation() {
2204 if (!UseTypeSpeculation) {
2205 return;
2206 }
2207 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2208 if (_gvn.type(local(i))->isa_oopptr()) {
2209 ciKlass* better_type = method()->parameter_profiled_type(j);
2210 if (better_type != NULL) {
2211 record_profile_for_speculation(local(i), better_type);
2212 }
2213 j++;
2214 }
2215 }
2216 }
2217
2218 void GraphKit::round_double_result(ciMethod* dest_method) {
2219 // A non-strict method may return a double value which has an extended
2220 // exponent, but this must not be visible in a caller which is 'strict'
2221 // If a strict caller invokes a non-strict callee, round a double result
2222
2223 BasicType result_type = dest_method->return_type()->basic_type();
2224 assert( method() != NULL, "must have caller context");
2225 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2226 // Destination method's return value is on top of stack
2227 // dstore_rounding() does gvn.transform
2228 Node *result = pop_pair();
2229 result = dstore_rounding(result);
2230 push_pair(result);
2231 }
2232 }
2233
2234 // rounding for strict float precision conformance
2235 Node* GraphKit::precision_rounding(Node* n) {
2236 return UseStrictFP && _method->flags().is_strict()
2237 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2238 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2239 : n;
2240 }
2241
2242 // rounding for strict double precision conformance
2243 Node* GraphKit::dprecision_rounding(Node *n) {
2244 return UseStrictFP && _method->flags().is_strict()
2245 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2246 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2247 : n;
2248 }
2249
2250 // rounding for non-strict double stores
2251 Node* GraphKit::dstore_rounding(Node* n) {
2252 return Matcher::strict_fp_requires_explicit_rounding
2253 && UseSSE <= 1
2254 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2255 : n;
2256 }
2257
2258 //=============================================================================
2259 // Generate a fast path/slow path idiom. Graph looks like:
2260 // [foo] indicates that 'foo' is a parameter
2261 //
2262 // [in] NULL
2263 // \ /
2264 // CmpP
2265 // Bool ne
2266 // If
2267 // / \
2268 // True False-<2>
2269 // / |
2270 // / cast_not_null
2271 // Load | | ^
2272 // [fast_test] | |
2273 // gvn to opt_test | |
2274 // / \ | <1>
2275 // True False |
2276 // | \\ |
2277 // [slow_call] \[fast_result]
2278 // Ctl Val \ \
2279 // | \ \
2280 // Catch <1> \ \
2281 // / \ ^ \ \
2282 // Ex No_Ex | \ \
2283 // | \ \ | \ <2> \
2284 // ... \ [slow_res] | | \ [null_result]
2285 // \ \--+--+--- | |
2286 // \ | / \ | /
2287 // --------Region Phi
2288 //
2289 //=============================================================================
2290 // Code is structured as a series of driver functions all called 'do_XXX' that
2291 // call a set of helper functions. Helper functions first, then drivers.
2292
2293 //------------------------------null_check_oop---------------------------------
2294 // Null check oop. Set null-path control into Region in slot 3.
2295 // Make a cast-not-nullness use the other not-null control. Return cast.
2296 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2297 bool never_see_null, bool safe_for_replace) {
2298 // Initial NULL check taken path
2299 (*null_control) = top();
2300 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2301
2302 // Generate uncommon_trap:
2303 if (never_see_null && (*null_control) != top()) {
2304 // If we see an unexpected null at a check-cast we record it and force a
2305 // recompile; the offending check-cast will be compiled to handle NULLs.
2306 // If we see more than one offending BCI, then all checkcasts in the
2307 // method will be compiled to handle NULLs.
2308 PreserveJVMState pjvms(this);
2309 set_control(*null_control);
2310 replace_in_map(value, null());
2311 uncommon_trap(Deoptimization::Reason_null_check,
2312 Deoptimization::Action_make_not_entrant);
2313 (*null_control) = top(); // NULL path is dead
2314 }
2315 if ((*null_control) == top() && safe_for_replace) {
2316 replace_in_map(value, cast);
2317 }
2318
2319 // Cast away null-ness on the result
2320 return cast;
2321 }
2322
2323 //------------------------------opt_iff----------------------------------------
2324 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2325 // Return slow-path control.
2326 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2327 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2328
2329 // Fast path taken; set region slot 2
2330 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2331 region->init_req(2,fast_taken); // Capture fast-control
2332
2333 // Fast path not-taken, i.e. slow path
2334 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2335 return slow_taken;
2336 }
2337
2338 //-----------------------------make_runtime_call-------------------------------
2339 Node* GraphKit::make_runtime_call(int flags,
2340 const TypeFunc* call_type, address call_addr,
2341 const char* call_name,
2342 const TypePtr* adr_type,
2343 // The following parms are all optional.
2344 // The first NULL ends the list.
2345 Node* parm0, Node* parm1,
2346 Node* parm2, Node* parm3,
2347 Node* parm4, Node* parm5,
2348 Node* parm6, Node* parm7) {
2349 // Slow-path call
2350 bool is_leaf = !(flags & RC_NO_LEAF);
2351 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2352 if (call_name == NULL) {
2353 assert(!is_leaf, "must supply name for leaf");
2354 call_name = OptoRuntime::stub_name(call_addr);
2355 }
2356 CallNode* call;
2357 if (!is_leaf) {
2358 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2359 bci(), adr_type);
2360 } else if (flags & RC_NO_FP) {
2361 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2362 } else {
2363 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2364 }
2365
2366 // The following is similar to set_edges_for_java_call,
2367 // except that the memory effects of the call are restricted to AliasIdxRaw.
2368
2369 // Slow path call has no side-effects, uses few values
2370 bool wide_in = !(flags & RC_NARROW_MEM);
2371 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2372
2373 Node* prev_mem = NULL;
2374 if (wide_in) {
2375 prev_mem = set_predefined_input_for_runtime_call(call);
2376 } else {
2377 assert(!wide_out, "narrow in => narrow out");
2378 Node* narrow_mem = memory(adr_type);
2379 prev_mem = reset_memory();
2380 map()->set_memory(narrow_mem);
2381 set_predefined_input_for_runtime_call(call);
2382 }
2383
2384 // Hook each parm in order. Stop looking at the first NULL.
2385 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2386 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2387 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2388 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2389 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2390 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2391 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2392 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2393 /* close each nested if ===> */ } } } } } } } }
2394 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2395
2396 if (!is_leaf) {
2397 // Non-leaves can block and take safepoints:
2398 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2399 }
2400 // Non-leaves can throw exceptions:
2401 if (has_io) {
2402 call->set_req(TypeFunc::I_O, i_o());
2403 }
2404
2405 if (flags & RC_UNCOMMON) {
2406 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2407 // (An "if" probability corresponds roughly to an unconditional count.
2408 // Sort of.)
2409 call->set_cnt(PROB_UNLIKELY_MAG(4));
2410 }
2411
2412 Node* c = _gvn.transform(call);
2413 assert(c == call, "cannot disappear");
2414
2415 if (wide_out) {
2416 // Slow path call has full side-effects.
2417 set_predefined_output_for_runtime_call(call);
2418 } else {
2419 // Slow path call has few side-effects, and/or sets few values.
2420 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2421 }
2422
2423 if (has_io) {
2424 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2425 }
2426 return call;
2427
2428 }
2429
2430 //------------------------------merge_memory-----------------------------------
2431 // Merge memory from one path into the current memory state.
2432 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2433 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2434 Node* old_slice = mms.force_memory();
2435 Node* new_slice = mms.memory2();
2436 if (old_slice != new_slice) {
2437 PhiNode* phi;
2438 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2439 phi = new_slice->as_Phi();
2440 #ifdef ASSERT
2441 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2442 old_slice = old_slice->in(new_path);
2443 // Caller is responsible for ensuring that any pre-existing
2444 // phis are already aware of old memory.
2445 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2446 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2447 #endif
2448 mms.set_memory(phi);
2449 } else {
2450 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2451 _gvn.set_type(phi, Type::MEMORY);
2452 phi->set_req(new_path, new_slice);
2453 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2454 }
2455 }
2456 }
2457 }
2458
2459 //------------------------------make_slow_call_ex------------------------------
2460 // Make the exception handler hookups for the slow call
2461 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2462 if (stopped()) return;
2463
2464 // Make a catch node with just two handlers: fall-through and catch-all
2465 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2466 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2467 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2468 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2469
2470 { PreserveJVMState pjvms(this);
2471 set_control(excp);
2472 set_i_o(i_o);
2473
2474 if (excp != top()) {
2475 // Create an exception state also.
2476 // Use an exact type if the caller has specified a specific exception.
2477 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2478 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2479 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2480 }
2481 }
2482
2483 // Get the no-exception control from the CatchNode.
2484 set_control(norm);
2485 }
2486
2487
2488 //-------------------------------gen_subtype_check-----------------------------
2489 // Generate a subtyping check. Takes as input the subtype and supertype.
2490 // Returns 2 values: sets the default control() to the true path and returns
2491 // the false path. Only reads invariant memory; sets no (visible) memory.
2492 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2493 // but that's not exposed to the optimizer. This call also doesn't take in an
2494 // Object; if you wish to check an Object you need to load the Object's class
2495 // prior to coming here.
2496 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2497 // Fast check for identical types, perhaps identical constants.
2498 // The types can even be identical non-constants, in cases
2499 // involving Array.newInstance, Object.clone, etc.
2500 if (subklass == superklass)
2501 return top(); // false path is dead; no test needed.
2502
2503 if (_gvn.type(superklass)->singleton()) {
2504 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2505 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2506
2507 // In the common case of an exact superklass, try to fold up the
2508 // test before generating code. You may ask, why not just generate
2509 // the code and then let it fold up? The answer is that the generated
2510 // code will necessarily include null checks, which do not always
2511 // completely fold away. If they are also needless, then they turn
2512 // into a performance loss. Example:
2513 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2514 // Here, the type of 'fa' is often exact, so the store check
2515 // of fa[1]=x will fold up, without testing the nullness of x.
2516 switch (static_subtype_check(superk, subk)) {
2517 case SSC_always_false:
2518 {
2519 Node* always_fail = control();
2520 set_control(top());
2521 return always_fail;
2522 }
2523 case SSC_always_true:
2524 return top();
2525 case SSC_easy_test:
2526 {
2527 // Just do a direct pointer compare and be done.
2528 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2529 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2530 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2531 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2532 return _gvn.transform( new(C) IfFalseNode(iff) );
2533 }
2534 case SSC_full_test:
2535 break;
2536 default:
2537 ShouldNotReachHere();
2538 }
2539 }
2540
2541 // %%% Possible further optimization: Even if the superklass is not exact,
2542 // if the subklass is the unique subtype of the superklass, the check
2543 // will always succeed. We could leave a dependency behind to ensure this.
2544
2545 // First load the super-klass's check-offset
2546 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2547 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
2548 TypeInt::INT, MemNode::unordered));
2549 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2550 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2551
2552 // Load from the sub-klass's super-class display list, or a 1-word cache of
2553 // the secondary superclass list, or a failing value with a sentinel offset
2554 // if the super-klass is an interface or exceptionally deep in the Java
2555 // hierarchy and we have to scan the secondary superclass list the hard way.
2556 // Worst-case type is a little odd: NULL is allowed as a result (usually
2557 // klass loads can never produce a NULL).
2558 Node *chk_off_X = ConvI2X(chk_off);
2559 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2560 // For some types like interfaces the following loadKlass is from a 1-word
2561 // cache which is mutable so can't use immutable memory. Other
2562 // types load from the super-class display table which is immutable.
2563 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2564 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2565
2566 // Compile speed common case: ARE a subtype and we canNOT fail
2567 if( superklass == nkls )
2568 return top(); // false path is dead; no test needed.
2569
2570 // See if we get an immediate positive hit. Happens roughly 83% of the
2571 // time. Test to see if the value loaded just previously from the subklass
2572 // is exactly the superklass.
2573 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2574 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2575 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2576 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2577 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2578
2579 // Compile speed common case: Check for being deterministic right now. If
2580 // chk_off is a constant and not equal to cacheoff then we are NOT a
2581 // subklass. In this case we need exactly the 1 test above and we can
2582 // return those results immediately.
2583 if (!might_be_cache) {
2584 Node* not_subtype_ctrl = control();
2585 set_control(iftrue1); // We need exactly the 1 test above
2586 return not_subtype_ctrl;
2587 }
2588
2589 // Gather the various success & failures here
2590 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2591 record_for_igvn(r_ok_subtype);
2592 RegionNode *r_not_subtype = new (C) RegionNode(3);
2593 record_for_igvn(r_not_subtype);
2594
2595 r_ok_subtype->init_req(1, iftrue1);
2596
2597 // Check for immediate negative hit. Happens roughly 11% of the time (which
2598 // is roughly 63% of the remaining cases). Test to see if the loaded
2599 // check-offset points into the subklass display list or the 1-element
2600 // cache. If it points to the display (and NOT the cache) and the display
2601 // missed then it's not a subtype.
2602 Node *cacheoff = _gvn.intcon(cacheoff_con);
2603 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2604 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2605 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2606 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2607 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2608
2609 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2610 // No performance impact (too rare) but allows sharing of secondary arrays
2611 // which has some footprint reduction.
2612 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2613 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2614 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2615 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2616 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2617
2618 // -- Roads not taken here: --
2619 // We could also have chosen to perform the self-check at the beginning
2620 // of this code sequence, as the assembler does. This would not pay off
2621 // the same way, since the optimizer, unlike the assembler, can perform
2622 // static type analysis to fold away many successful self-checks.
2623 // Non-foldable self checks work better here in second position, because
2624 // the initial primary superclass check subsumes a self-check for most
2625 // types. An exception would be a secondary type like array-of-interface,
2626 // which does not appear in its own primary supertype display.
2627 // Finally, we could have chosen to move the self-check into the
2628 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2629 // dependent manner. But it is worthwhile to have the check here,
2630 // where it can be perhaps be optimized. The cost in code space is
2631 // small (register compare, branch).
2632
2633 // Now do a linear scan of the secondary super-klass array. Again, no real
2634 // performance impact (too rare) but it's gotta be done.
2635 // Since the code is rarely used, there is no penalty for moving it
2636 // out of line, and it can only improve I-cache density.
2637 // The decision to inline or out-of-line this final check is platform
2638 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2639 Node* psc = _gvn.transform(
2640 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2641
2642 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2643 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2644 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2645 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2646 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2647
2648 // Return false path; set default control to true path.
2649 set_control( _gvn.transform(r_ok_subtype) );
2650 return _gvn.transform(r_not_subtype);
2651 }
2652
2653 //----------------------------static_subtype_check-----------------------------
2654 // Shortcut important common cases when superklass is exact:
2655 // (0) superklass is java.lang.Object (can occur in reflective code)
2656 // (1) subklass is already limited to a subtype of superklass => always ok
2657 // (2) subklass does not overlap with superklass => always fail
2658 // (3) superklass has NO subtypes and we can check with a simple compare.
2659 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2660 if (StressReflectiveCode) {
2661 return SSC_full_test; // Let caller generate the general case.
2662 }
2663
2664 if (superk == env()->Object_klass()) {
2665 return SSC_always_true; // (0) this test cannot fail
2666 }
2667
2668 ciType* superelem = superk;
2669 if (superelem->is_array_klass())
2670 superelem = superelem->as_array_klass()->base_element_type();
2671
2672 if (!subk->is_interface()) { // cannot trust static interface types yet
2673 if (subk->is_subtype_of(superk)) {
2674 return SSC_always_true; // (1) false path dead; no dynamic test needed
2675 }
2676 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2677 !superk->is_subtype_of(subk)) {
2678 return SSC_always_false;
2679 }
2680 }
2681
2682 // If casting to an instance klass, it must have no subtypes
2683 if (superk->is_interface()) {
2684 // Cannot trust interfaces yet.
2685 // %%% S.B. superk->nof_implementors() == 1
2686 } else if (superelem->is_instance_klass()) {
2687 ciInstanceKlass* ik = superelem->as_instance_klass();
2688 if (!ik->has_subklass() && !ik->is_interface()) {
2689 if (!ik->is_final()) {
2690 // Add a dependency if there is a chance of a later subclass.
2691 C->dependencies()->assert_leaf_type(ik);
2692 }
2693 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2694 }
2695 } else {
2696 // A primitive array type has no subtypes.
2697 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2698 }
2699
2700 return SSC_full_test;
2701 }
2702
2703 // Profile-driven exact type check:
2704 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2705 float prob,
2706 Node* *casted_receiver) {
2707 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2708 Node* recv_klass = load_object_klass(receiver);
2709 Node* want_klass = makecon(tklass);
2710 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2711 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2712 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2713 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2714 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2715
2716 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2717 assert(recv_xtype->klass_is_exact(), "");
2718
2719 // Subsume downstream occurrences of receiver with a cast to
2720 // recv_xtype, since now we know what the type will be.
2721 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2722 (*casted_receiver) = _gvn.transform(cast);
2723 // (User must make the replace_in_map call.)
2724
2725 return fail;
2726 }
2727
2728
2729 //------------------------------seems_never_null-------------------------------
2730 // Use null_seen information if it is available from the profile.
2731 // If we see an unexpected null at a type check we record it and force a
2732 // recompile; the offending check will be recompiled to handle NULLs.
2733 // If we see several offending BCIs, then all checks in the
2734 // method will be recompiled.
2735 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2736 if (UncommonNullCast // Cutout for this technique
2737 && obj != null() // And not the -Xcomp stupid case?
2738 && !too_many_traps(Deoptimization::Reason_null_check)
2739 ) {
2740 if (data == NULL)
2741 // Edge case: no mature data. Be optimistic here.
2742 return true;
2743 // If the profile has not seen a null, assume it won't happen.
2744 assert(java_bc() == Bytecodes::_checkcast ||
2745 java_bc() == Bytecodes::_instanceof ||
2746 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2747 return !data->as_BitData()->null_seen();
2748 }
2749 return false;
2750 }
2751
2752 //------------------------maybe_cast_profiled_receiver-------------------------
2753 // If the profile has seen exactly one type, narrow to exactly that type.
2754 // Subsequent type checks will always fold up.
2755 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2756 ciKlass* require_klass,
2757 ciKlass* spec_klass,
2758 bool safe_for_replace) {
2759 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2760
2761 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2762
2763 // Make sure we haven't already deoptimized from this tactic.
2764 if (too_many_traps(reason))
2765 return NULL;
2766
2767 // (No, this isn't a call, but it's enough like a virtual call
2768 // to use the same ciMethod accessor to get the profile info...)
2769 // If we have a speculative type use it instead of profiling (which
2770 // may not help us)
2771 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2772 if (exact_kls != NULL) {// no cast failures here
2773 if (require_klass == NULL ||
2774 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2775 // If we narrow the type to match what the type profile sees or
2776 // the speculative type, we can then remove the rest of the
2777 // cast.
2778 // This is a win, even if the exact_kls is very specific,
2779 // because downstream operations, such as method calls,
2780 // will often benefit from the sharper type.
2781 Node* exact_obj = not_null_obj; // will get updated in place...
2782 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2783 &exact_obj);
2784 { PreserveJVMState pjvms(this);
2785 set_control(slow_ctl);
2786 uncommon_trap(reason,
2787 Deoptimization::Action_maybe_recompile);
2788 }
2789 if (safe_for_replace) {
2790 replace_in_map(not_null_obj, exact_obj);
2791 }
2792 return exact_obj;
2793 }
2794 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2795 }
2796
2797 return NULL;
2798 }
2799
2800 /**
2801 * Cast obj to type and emit guard unless we had too many traps here
2802 * already
2803 *
2804 * @param obj node being casted
2805 * @param type type to cast the node to
2806 * @param not_null true if we know node cannot be null
2807 */
2808 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2809 ciKlass* type,
2810 bool not_null) {
2811 // type == NULL if profiling tells us this object is always null
2812 if (type != NULL) {
2813 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2814 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2815 if (!too_many_traps(null_reason) &&
2816 !too_many_traps(class_reason)) {
2817 Node* not_null_obj = NULL;
2818 // not_null is true if we know the object is not null and
2819 // there's no need for a null check
2820 if (!not_null) {
2821 Node* null_ctl = top();
2822 not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2823 assert(null_ctl->is_top(), "no null control here");
2824 } else {
2825 not_null_obj = obj;
2826 }
2827
2828 Node* exact_obj = not_null_obj;
2829 ciKlass* exact_kls = type;
2830 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2831 &exact_obj);
2832 {
2833 PreserveJVMState pjvms(this);
2834 set_control(slow_ctl);
2835 uncommon_trap(class_reason,
2836 Deoptimization::Action_maybe_recompile);
2837 }
2838 replace_in_map(not_null_obj, exact_obj);
2839 obj = exact_obj;
2840 }
2841 } else {
2842 if (!too_many_traps(Deoptimization::Reason_null_assert)) {
2843 Node* exact_obj = null_assert(obj);
2844 replace_in_map(obj, exact_obj);
2845 obj = exact_obj;
2846 }
2847 }
2848 return obj;
2849 }
2850
2851 //-------------------------------gen_instanceof--------------------------------
2852 // Generate an instance-of idiom. Used by both the instance-of bytecode
2853 // and the reflective instance-of call.
2854 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2855 kill_dead_locals(); // Benefit all the uncommon traps
2856 assert( !stopped(), "dead parse path should be checked in callers" );
2857 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2858 "must check for not-null not-dead klass in callers");
2859
2860 // Make the merge point
2861 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2862 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2863 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2864 C->set_has_split_ifs(true); // Has chance for split-if optimization
2865
2866 ciProfileData* data = NULL;
2867 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2868 data = method()->method_data()->bci_to_data(bci());
2869 }
2870 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2871 && seems_never_null(obj, data));
2872
2873 // Null check; get casted pointer; set region slot 3
2874 Node* null_ctl = top();
2875 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2876
2877 // If not_null_obj is dead, only null-path is taken
2878 if (stopped()) { // Doing instance-of on a NULL?
2879 set_control(null_ctl);
2880 return intcon(0);
2881 }
2882 region->init_req(_null_path, null_ctl);
2883 phi ->init_req(_null_path, intcon(0)); // Set null path value
2884 if (null_ctl == top()) {
2885 // Do this eagerly, so that pattern matches like is_diamond_phi
2886 // will work even during parsing.
2887 assert(_null_path == PATH_LIMIT-1, "delete last");
2888 region->del_req(_null_path);
2889 phi ->del_req(_null_path);
2890 }
2891
2892 // Do we know the type check always succeed?
2893 bool known_statically = false;
2894 if (_gvn.type(superklass)->singleton()) {
2895 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2896 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2897 if (subk != NULL && subk->is_loaded()) {
2898 int static_res = static_subtype_check(superk, subk);
2899 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2900 }
2901 }
2902
2903 if (known_statically && UseTypeSpeculation) {
2904 // If we know the type check always succeeds then we don't use the
2905 // profiling data at this bytecode. Don't lose it, feed it to the
2906 // type system as a speculative type.
2907 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2908 } else {
2909 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2910 // We may not have profiling here or it may not help us. If we
2911 // have a speculative type use it to perform an exact cast.
2912 ciKlass* spec_obj_type = obj_type->speculative_type();
2913 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2914 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2915 if (stopped()) { // Profile disagrees with this path.
2916 set_control(null_ctl); // Null is the only remaining possibility.
2917 return intcon(0);
2918 }
2919 if (cast_obj != NULL) {
2920 not_null_obj = cast_obj;
2921 }
2922 }
2923 }
2924
2925 // Load the object's klass
2926 Node* obj_klass = load_object_klass(not_null_obj);
2927
2928 // Generate the subtype check
2929 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2930
2931 // Plug in the success path to the general merge in slot 1.
2932 region->init_req(_obj_path, control());
2933 phi ->init_req(_obj_path, intcon(1));
2934
2935 // Plug in the failing path to the general merge in slot 2.
2936 region->init_req(_fail_path, not_subtype_ctrl);
2937 phi ->init_req(_fail_path, intcon(0));
2938
2939 // Return final merged results
2940 set_control( _gvn.transform(region) );
2941 record_for_igvn(region);
2942 return _gvn.transform(phi);
2943 }
2944
2945 //-------------------------------gen_checkcast---------------------------------
2946 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2947 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2948 // uncommon-trap paths work. Adjust stack after this call.
2949 // If failure_control is supplied and not null, it is filled in with
2950 // the control edge for the cast failure. Otherwise, an appropriate
2951 // uncommon trap or exception is thrown.
2952 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2953 Node* *failure_control) {
2954 kill_dead_locals(); // Benefit all the uncommon traps
2955 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2956 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2957
2958 // Fast cutout: Check the case that the cast is vacuously true.
2959 // This detects the common cases where the test will short-circuit
2960 // away completely. We do this before we perform the null check,
2961 // because if the test is going to turn into zero code, we don't
2962 // want a residual null check left around. (Causes a slowdown,
2963 // for example, in some objArray manipulations, such as a[i]=a[j].)
2964 if (tk->singleton()) {
2965 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2966 if (objtp != NULL && objtp->klass() != NULL) {
2967 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2968 case SSC_always_true:
2969 // If we know the type check always succeed then we don't use
2970 // the profiling data at this bytecode. Don't lose it, feed it
2971 // to the type system as a speculative type.
2972 return record_profiled_receiver_for_speculation(obj);
2973 case SSC_always_false:
2974 // It needs a null check because a null will *pass* the cast check.
2975 // A non-null value will always produce an exception.
2976 return null_assert(obj);
2977 }
2978 }
2979 }
2980
2981 ciProfileData* data = NULL;
2982 bool safe_for_replace = false;
2983 if (failure_control == NULL) { // use MDO in regular case only
2984 assert(java_bc() == Bytecodes::_aastore ||
2985 java_bc() == Bytecodes::_checkcast,
2986 "interpreter profiles type checks only for these BCs");
2987 data = method()->method_data()->bci_to_data(bci());
2988 safe_for_replace = true;
2989 }
2990
2991 // Make the merge point
2992 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2993 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2994 Node* phi = new (C) PhiNode(region, toop);
2995 C->set_has_split_ifs(true); // Has chance for split-if optimization
2996
2997 // Use null-cast information if it is available
2998 bool never_see_null = ((failure_control == NULL) // regular case only
2999 && seems_never_null(obj, data));
3000
3001 // Null check; get casted pointer; set region slot 3
3002 Node* null_ctl = top();
3003 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
3004
3005 // If not_null_obj is dead, only null-path is taken
3006 if (stopped()) { // Doing instance-of on a NULL?
3007 set_control(null_ctl);
3008 return null();
3009 }
3010 region->init_req(_null_path, null_ctl);
3011 phi ->init_req(_null_path, null()); // Set null path value
3012 if (null_ctl == top()) {
3013 // Do this eagerly, so that pattern matches like is_diamond_phi
3014 // will work even during parsing.
3015 assert(_null_path == PATH_LIMIT-1, "delete last");
3016 region->del_req(_null_path);
3017 phi ->del_req(_null_path);
3018 }
3019
3020 Node* cast_obj = NULL;
3021 if (tk->klass_is_exact()) {
3022 // The following optimization tries to statically cast the speculative type of the object
3023 // (for example obtained during profiling) to the type of the superklass and then do a
3024 // dynamic check that the type of the object is what we expect. To work correctly
3025 // for checkcast and aastore the type of superklass should be exact.
3026 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3027 // We may not have profiling here or it may not help us. If we have
3028 // a speculative type use it to perform an exact cast.
3029 ciKlass* spec_obj_type = obj_type->speculative_type();
3030 if (spec_obj_type != NULL ||
3031 (data != NULL &&
3032 // Counter has never been decremented (due to cast failure).
3033 // ...This is a reasonable thing to expect. It is true of
3034 // all casts inserted by javac to implement generic types.
3035 data->as_CounterData()->count() >= 0)) {
3036 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3037 if (cast_obj != NULL) {
3038 if (failure_control != NULL) // failure is now impossible
3039 (*failure_control) = top();
3040 // adjust the type of the phi to the exact klass:
3041 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3042 }
3043 }
3044 }
3045
3046 if (cast_obj == NULL) {
3047 // Load the object's klass
3048 Node* obj_klass = load_object_klass(not_null_obj);
3049
3050 // Generate the subtype check
3051 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3052
3053 // Plug in success path into the merge
3054 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3055 not_null_obj, toop));
3056 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3057 if (failure_control == NULL) {
3058 if (not_subtype_ctrl != top()) { // If failure is possible
3059 PreserveJVMState pjvms(this);
3060 set_control(not_subtype_ctrl);
3061 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3062 }
3063 } else {
3064 (*failure_control) = not_subtype_ctrl;
3065 }
3066 }
3067
3068 region->init_req(_obj_path, control());
3069 phi ->init_req(_obj_path, cast_obj);
3070
3071 // A merge of NULL or Casted-NotNull obj
3072 Node* res = _gvn.transform(phi);
3073
3074 // Note I do NOT always 'replace_in_map(obj,result)' here.
3075 // if( tk->klass()->can_be_primary_super() )
3076 // This means that if I successfully store an Object into an array-of-String
3077 // I 'forget' that the Object is really now known to be a String. I have to
3078 // do this because we don't have true union types for interfaces - if I store
3079 // a Baz into an array-of-Interface and then tell the optimizer it's an
3080 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3081 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3082 // replace_in_map( obj, res );
3083
3084 // Return final merged results
3085 set_control( _gvn.transform(region) );
3086 record_for_igvn(region);
3087 return res;
3088 }
3089
3090 //------------------------------next_monitor-----------------------------------
3091 // What number should be given to the next monitor?
3092 int GraphKit::next_monitor() {
3093 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3094 int next = current + C->sync_stack_slots();
3095 // Keep the toplevel high water mark current:
3096 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3097 return current;
3098 }
3099
3100 //------------------------------insert_mem_bar---------------------------------
3101 // Memory barrier to avoid floating things around
3102 // The membar serves as a pinch point between both control and all memory slices.
3103 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3104 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3105 mb->init_req(TypeFunc::Control, control());
3106 mb->init_req(TypeFunc::Memory, reset_memory());
3107 Node* membar = _gvn.transform(mb);
3108 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3109 set_all_memory_call(membar);
3110 return membar;
3111 }
3112
3113 //-------------------------insert_mem_bar_volatile----------------------------
3114 // Memory barrier to avoid floating things around
3115 // The membar serves as a pinch point between both control and memory(alias_idx).
3116 // If you want to make a pinch point on all memory slices, do not use this
3117 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3118 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3119 // When Parse::do_put_xxx updates a volatile field, it appends a series
3120 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3121 // The first membar is on the same memory slice as the field store opcode.
3122 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3123 // All the other membars (for other volatile slices, including AliasIdxBot,
3124 // which stands for all unknown volatile slices) are control-dependent
3125 // on the first membar. This prevents later volatile loads or stores
3126 // from sliding up past the just-emitted store.
3127
3128 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3129 mb->set_req(TypeFunc::Control,control());
3130 if (alias_idx == Compile::AliasIdxBot) {
3131 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3132 } else {
3133 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3134 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3135 }
3136 Node* membar = _gvn.transform(mb);
3137 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3138 if (alias_idx == Compile::AliasIdxBot) {
3139 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3140 } else {
3141 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3142 }
3143 return membar;
3144 }
3145
3146 //------------------------------shared_lock------------------------------------
3147 // Emit locking code.
3148 FastLockNode* GraphKit::shared_lock(Node* obj) {
3149 // bci is either a monitorenter bc or InvocationEntryBci
3150 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3151 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3152
3153 if( !GenerateSynchronizationCode )
3154 return NULL; // Not locking things?
3155 if (stopped()) // Dead monitor?
3156 return NULL;
3157
3158 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3159
3160 // Box the stack location
3161 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3162 Node* mem = reset_memory();
3163
3164 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3165 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3166 // Create the counters for this fast lock.
3167 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3168 }
3169
3170 // Create the rtm counters for this fast lock if needed.
3171 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3172
3173 // Add monitor to debug info for the slow path. If we block inside the
3174 // slow path and de-opt, we need the monitor hanging around
3175 map()->push_monitor( flock );
3176
3177 const TypeFunc *tf = LockNode::lock_type();
3178 LockNode *lock = new (C) LockNode(C, tf);
3179
3180 lock->init_req( TypeFunc::Control, control() );
3181 lock->init_req( TypeFunc::Memory , mem );
3182 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3183 lock->init_req( TypeFunc::FramePtr, frameptr() );
3184 lock->init_req( TypeFunc::ReturnAdr, top() );
3185
3186 lock->init_req(TypeFunc::Parms + 0, obj);
3187 lock->init_req(TypeFunc::Parms + 1, box);
3188 lock->init_req(TypeFunc::Parms + 2, flock);
3189 add_safepoint_edges(lock);
3190
3191 lock = _gvn.transform( lock )->as_Lock();
3192
3193 // lock has no side-effects, sets few values
3194 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3195
3196 insert_mem_bar(Op_MemBarAcquireLock);
3197
3198 // Add this to the worklist so that the lock can be eliminated
3199 record_for_igvn(lock);
3200
3201 #ifndef PRODUCT
3202 if (PrintLockStatistics) {
3203 // Update the counter for this lock. Don't bother using an atomic
3204 // operation since we don't require absolute accuracy.
3205 lock->create_lock_counter(map()->jvms());
3206 increment_counter(lock->counter()->addr());
3207 }
3208 #endif
3209
3210 return flock;
3211 }
3212
3213
3214 //------------------------------shared_unlock----------------------------------
3215 // Emit unlocking code.
3216 void GraphKit::shared_unlock(Node* box, Node* obj) {
3217 // bci is either a monitorenter bc or InvocationEntryBci
3218 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3219 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3220
3221 if( !GenerateSynchronizationCode )
3222 return;
3223 if (stopped()) { // Dead monitor?
3224 map()->pop_monitor(); // Kill monitor from debug info
3225 return;
3226 }
3227
3228 // Memory barrier to avoid floating things down past the locked region
3229 insert_mem_bar(Op_MemBarReleaseLock);
3230
3231 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3232 UnlockNode *unlock = new (C) UnlockNode(C, tf);
3233 uint raw_idx = Compile::AliasIdxRaw;
3234 unlock->init_req( TypeFunc::Control, control() );
3235 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3236 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3237 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3238 unlock->init_req( TypeFunc::ReturnAdr, top() );
3239
3240 unlock->init_req(TypeFunc::Parms + 0, obj);
3241 unlock->init_req(TypeFunc::Parms + 1, box);
3242 unlock = _gvn.transform(unlock)->as_Unlock();
3243
3244 Node* mem = reset_memory();
3245
3246 // unlock has no side-effects, sets few values
3247 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3248
3249 // Kill monitor from debug info
3250 map()->pop_monitor( );
3251 }
3252
3253 //-------------------------------get_layout_helper-----------------------------
3254 // If the given klass is a constant or known to be an array,
3255 // fetch the constant layout helper value into constant_value
3256 // and return (Node*)NULL. Otherwise, load the non-constant
3257 // layout helper value, and return the node which represents it.
3258 // This two-faced routine is useful because allocation sites
3259 // almost always feature constant types.
3260 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3261 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3262 if (!StressReflectiveCode && inst_klass != NULL) {
3263 ciKlass* klass = inst_klass->klass();
3264 bool xklass = inst_klass->klass_is_exact();
3265 if (xklass || klass->is_array_klass()) {
3266 jint lhelper = klass->layout_helper();
3267 if (lhelper != Klass::_lh_neutral_value) {
3268 constant_value = lhelper;
3269 return (Node*) NULL;
3270 }
3271 }
3272 }
3273 constant_value = Klass::_lh_neutral_value; // put in a known value
3274 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3275 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3276 }
3277
3278 // We just put in an allocate/initialize with a big raw-memory effect.
3279 // Hook selected additional alias categories on the initialization.
3280 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3281 MergeMemNode* init_in_merge,
3282 Node* init_out_raw) {
3283 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3284 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3285
3286 Node* prevmem = kit.memory(alias_idx);
3287 init_in_merge->set_memory_at(alias_idx, prevmem);
3288 kit.set_memory(init_out_raw, alias_idx);
3289 }
3290
3291 //---------------------------set_output_for_allocation-------------------------
3292 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3293 const TypeOopPtr* oop_type) {
3294 int rawidx = Compile::AliasIdxRaw;
3295 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3296 add_safepoint_edges(alloc);
3297 Node* allocx = _gvn.transform(alloc);
3298 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3299 // create memory projection for i_o
3300 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3301 make_slow_call_ex(allocx, env()->Throwable_klass(), true);
3302
3303 // create a memory projection as for the normal control path
3304 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3305 set_memory(malloc, rawidx);
3306
3307 // a normal slow-call doesn't change i_o, but an allocation does
3308 // we create a separate i_o projection for the normal control path
3309 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3310 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3311
3312 // put in an initialization barrier
3313 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3314 rawoop)->as_Initialize();
3315 assert(alloc->initialization() == init, "2-way macro link must work");
3316 assert(init ->allocation() == alloc, "2-way macro link must work");
3317 {
3318 // Extract memory strands which may participate in the new object's
3319 // initialization, and source them from the new InitializeNode.
3320 // This will allow us to observe initializations when they occur,
3321 // and link them properly (as a group) to the InitializeNode.
3322 assert(init->in(InitializeNode::Memory) == malloc, "");
3323 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3324 init->set_req(InitializeNode::Memory, minit_in);
3325 record_for_igvn(minit_in); // fold it up later, if possible
3326 Node* minit_out = memory(rawidx);
3327 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3328 if (oop_type->isa_aryptr()) {
3329 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3330 int elemidx = C->get_alias_index(telemref);
3331 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3332 } else if (oop_type->isa_instptr()) {
3333 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3334 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3335 ciField* field = ik->nonstatic_field_at(i);
3336 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3337 continue; // do not bother to track really large numbers of fields
3338 // Find (or create) the alias category for this field:
3339 int fieldidx = C->alias_type(field)->index();
3340 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3341 }
3342 }
3343 }
3344
3345 // Cast raw oop to the real thing...
3346 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3347 javaoop = _gvn.transform(javaoop);
3348 C->set_recent_alloc(control(), javaoop);
3349 assert(just_allocated_object(control()) == javaoop, "just allocated");
3350
3351 #ifdef ASSERT
3352 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3353 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3354 "Ideal_allocation works");
3355 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3356 "Ideal_allocation works");
3357 if (alloc->is_AllocateArray()) {
3358 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3359 "Ideal_allocation works");
3360 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3361 "Ideal_allocation works");
3362 } else {
3363 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3364 }
3365 }
3366 #endif //ASSERT
3367
3368 return javaoop;
3369 }
3370
3371 //---------------------------new_instance--------------------------------------
3372 // This routine takes a klass_node which may be constant (for a static type)
3373 // or may be non-constant (for reflective code). It will work equally well
3374 // for either, and the graph will fold nicely if the optimizer later reduces
3375 // the type to a constant.
3376 // The optional arguments are for specialized use by intrinsics:
3377 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3378 // - If 'return_size_val', report the the total object size to the caller.
3379 Node* GraphKit::new_instance(Node* klass_node,
3380 Node* extra_slow_test,
3381 Node* *return_size_val) {
3382 // Compute size in doublewords
3383 // The size is always an integral number of doublewords, represented
3384 // as a positive bytewise size stored in the klass's layout_helper.
3385 // The layout_helper also encodes (in a low bit) the need for a slow path.
3386 jint layout_con = Klass::_lh_neutral_value;
3387 Node* layout_val = get_layout_helper(klass_node, layout_con);
3388 int layout_is_con = (layout_val == NULL);
3389
3390 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3391 // Generate the initial go-slow test. It's either ALWAYS (return a
3392 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3393 // case) a computed value derived from the layout_helper.
3394 Node* initial_slow_test = NULL;
3395 if (layout_is_con) {
3396 assert(!StressReflectiveCode, "stress mode does not use these paths");
3397 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3398 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3399
3400 } else { // reflective case
3401 // This reflective path is used by Unsafe.allocateInstance.
3402 // (It may be stress-tested by specifying StressReflectiveCode.)
3403 // Basically, we want to get into the VM is there's an illegal argument.
3404 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3405 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3406 if (extra_slow_test != intcon(0)) {
3407 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3408 }
3409 // (Macro-expander will further convert this to a Bool, if necessary.)
3410 }
3411
3412 // Find the size in bytes. This is easy; it's the layout_helper.
3413 // The size value must be valid even if the slow path is taken.
3414 Node* size = NULL;
3415 if (layout_is_con) {
3416 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3417 } else { // reflective case
3418 // This reflective path is used by clone and Unsafe.allocateInstance.
3419 size = ConvI2X(layout_val);
3420
3421 // Clear the low bits to extract layout_helper_size_in_bytes:
3422 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3423 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3424 size = _gvn.transform( new (C) AndXNode(size, mask) );
3425 }
3426 if (return_size_val != NULL) {
3427 (*return_size_val) = size;
3428 }
3429
3430 // This is a precise notnull oop of the klass.
3431 // (Actually, it need not be precise if this is a reflective allocation.)
3432 // It's what we cast the result to.
3433 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3434 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3435 const TypeOopPtr* oop_type = tklass->as_instance_type();
3436
3437 // Now generate allocation code
3438
3439 // The entire memory state is needed for slow path of the allocation
3440 // since GC and deoptimization can happened.
3441 Node *mem = reset_memory();
3442 set_all_memory(mem); // Create new memory state
3443
3444 AllocateNode* alloc
3445 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3446 control(), mem, i_o(),
3447 size, klass_node,
3448 initial_slow_test);
3449
3450 return set_output_for_allocation(alloc, oop_type);
3451 }
3452
3453 //-------------------------------new_array-------------------------------------
3454 // helper for both newarray and anewarray
3455 // The 'length' parameter is (obviously) the length of the array.
3456 // See comments on new_instance for the meaning of the other arguments.
3457 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3458 Node* length, // number of array elements
3459 int nargs, // number of arguments to push back for uncommon trap
3460 Node* *return_size_val) {
3461 jint layout_con = Klass::_lh_neutral_value;
3462 Node* layout_val = get_layout_helper(klass_node, layout_con);
3463 int layout_is_con = (layout_val == NULL);
3464
3465 if (!layout_is_con && !StressReflectiveCode &&
3466 !too_many_traps(Deoptimization::Reason_class_check)) {
3467 // This is a reflective array creation site.
3468 // Optimistically assume that it is a subtype of Object[],
3469 // so that we can fold up all the address arithmetic.
3470 layout_con = Klass::array_layout_helper(T_OBJECT);
3471 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3472 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3473 { BuildCutout unless(this, bol_lh, PROB_MAX);
3474 inc_sp(nargs);
3475 uncommon_trap(Deoptimization::Reason_class_check,
3476 Deoptimization::Action_maybe_recompile);
3477 }
3478 layout_val = NULL;
3479 layout_is_con = true;
3480 }
3481
3482 // Generate the initial go-slow test. Make sure we do not overflow
3483 // if length is huge (near 2Gig) or negative! We do not need
3484 // exact double-words here, just a close approximation of needed
3485 // double-words. We can't add any offset or rounding bits, lest we
3486 // take a size -1 of bytes and make it positive. Use an unsigned
3487 // compare, so negative sizes look hugely positive.
3488 int fast_size_limit = FastAllocateSizeLimit;
3489 if (layout_is_con) {
3490 assert(!StressReflectiveCode, "stress mode does not use these paths");
3491 // Increase the size limit if we have exact knowledge of array type.
3492 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3493 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3494 }
3495
3496 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3497 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3498 if (initial_slow_test->is_Bool()) {
3499 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3500 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3501 }
3502
3503 // --- Size Computation ---
3504 // array_size = round_to_heap(array_header + (length << elem_shift));
3505 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3506 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3507 // The rounding mask is strength-reduced, if possible.
3508 int round_mask = MinObjAlignmentInBytes - 1;
3509 Node* header_size = NULL;
3510 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3511 // (T_BYTE has the weakest alignment and size restrictions...)
3512 if (layout_is_con) {
3513 int hsize = Klass::layout_helper_header_size(layout_con);
3514 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3515 BasicType etype = Klass::layout_helper_element_type(layout_con);
3516 if ((round_mask & ~right_n_bits(eshift)) == 0)
3517 round_mask = 0; // strength-reduce it if it goes away completely
3518 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3519 assert(header_size_min <= hsize, "generic minimum is smallest");
3520 header_size_min = hsize;
3521 header_size = intcon(hsize + round_mask);
3522 } else {
3523 Node* hss = intcon(Klass::_lh_header_size_shift);
3524 Node* hsm = intcon(Klass::_lh_header_size_mask);
3525 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3526 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3527 Node* mask = intcon(round_mask);
3528 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3529 }
3530
3531 Node* elem_shift = NULL;
3532 if (layout_is_con) {
3533 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3534 if (eshift != 0)
3535 elem_shift = intcon(eshift);
3536 } else {
3537 // There is no need to mask or shift this value.
3538 // The semantics of LShiftINode include an implicit mask to 0x1F.
3539 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3540 elem_shift = layout_val;
3541 }
3542
3543 // Transition to native address size for all offset calculations:
3544 Node* lengthx = ConvI2X(length);
3545 Node* headerx = ConvI2X(header_size);
3546 #ifdef _LP64
3547 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3548 if (tllen != NULL && tllen->_lo < 0) {
3549 // Add a manual constraint to a positive range. Cf. array_element_address.
3550 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3551 if (size_max > tllen->_hi) size_max = tllen->_hi;
3552 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3553 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3554 }
3555 }
3556 #endif
3557
3558 // Combine header size (plus rounding) and body size. Then round down.
3559 // This computation cannot overflow, because it is used only in two
3560 // places, one where the length is sharply limited, and the other
3561 // after a successful allocation.
3562 Node* abody = lengthx;
3563 if (elem_shift != NULL)
3564 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3565 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3566 if (round_mask != 0) {
3567 Node* mask = MakeConX(~round_mask);
3568 size = _gvn.transform( new(C) AndXNode(size, mask) );
3569 }
3570 // else if round_mask == 0, the size computation is self-rounding
3571
3572 if (return_size_val != NULL) {
3573 // This is the size
3574 (*return_size_val) = size;
3575 }
3576
3577 // Now generate allocation code
3578
3579 // The entire memory state is needed for slow path of the allocation
3580 // since GC and deoptimization can happened.
3581 Node *mem = reset_memory();
3582 set_all_memory(mem); // Create new memory state
3583
3584 // Create the AllocateArrayNode and its result projections
3585 AllocateArrayNode* alloc
3586 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3587 control(), mem, i_o(),
3588 size, klass_node,
3589 initial_slow_test,
3590 length);
3591
3592 // Cast to correct type. Note that the klass_node may be constant or not,
3593 // and in the latter case the actual array type will be inexact also.
3594 // (This happens via a non-constant argument to inline_native_newArray.)
3595 // In any case, the value of klass_node provides the desired array type.
3596 const TypeInt* length_type = _gvn.find_int_type(length);
3597 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3598 if (ary_type->isa_aryptr() && length_type != NULL) {
3599 // Try to get a better type than POS for the size
3600 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3601 }
3602
3603 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3604
3605 // Cast length on remaining path to be as narrow as possible
3606 if (map()->find_edge(length) >= 0) {
3607 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3608 if (ccast != length) {
3609 _gvn.set_type_bottom(ccast);
3610 record_for_igvn(ccast);
3611 replace_in_map(length, ccast);
3612 }
3613 }
3614
3615 return javaoop;
3616 }
3617
3618 // The following "Ideal_foo" functions are placed here because they recognize
3619 // the graph shapes created by the functions immediately above.
3620
3621 //---------------------------Ideal_allocation----------------------------------
3622 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3623 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3624 if (ptr == NULL) { // reduce dumb test in callers
3625 return NULL;
3626 }
3627 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3628 ptr = ptr->in(1);
3629 if (ptr == NULL) return NULL;
3630 }
3631 // Return NULL for allocations with several casts:
3632 // j.l.reflect.Array.newInstance(jobject, jint)
3633 // Object.clone()
3634 // to keep more precise type from last cast.
3635 if (ptr->is_Proj()) {
3636 Node* allo = ptr->in(0);
3637 if (allo != NULL && allo->is_Allocate()) {
3638 return allo->as_Allocate();
3639 }
3640 }
3641 // Report failure to match.
3642 return NULL;
3643 }
3644
3645 // Fancy version which also strips off an offset (and reports it to caller).
3646 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3647 intptr_t& offset) {
3648 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3649 if (base == NULL) return NULL;
3650 return Ideal_allocation(base, phase);
3651 }
3652
3653 // Trace Initialize <- Proj[Parm] <- Allocate
3654 AllocateNode* InitializeNode::allocation() {
3655 Node* rawoop = in(InitializeNode::RawAddress);
3656 if (rawoop->is_Proj()) {
3657 Node* alloc = rawoop->in(0);
3658 if (alloc->is_Allocate()) {
3659 return alloc->as_Allocate();
3660 }
3661 }
3662 return NULL;
3663 }
3664
3665 // Trace Allocate -> Proj[Parm] -> Initialize
3666 InitializeNode* AllocateNode::initialization() {
3667 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3668 if (rawoop == NULL) return NULL;
3669 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3670 Node* init = rawoop->fast_out(i);
3671 if (init->is_Initialize()) {
3672 assert(init->as_Initialize()->allocation() == this, "2-way link");
3673 return init->as_Initialize();
3674 }
3675 }
3676 return NULL;
3677 }
3678
3679 //----------------------------- loop predicates ---------------------------
3680
3681 //------------------------------add_predicate_impl----------------------------
3682 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3683 // Too many traps seen?
3684 if (too_many_traps(reason)) {
3685 #ifdef ASSERT
3686 if (TraceLoopPredicate) {
3687 int tc = C->trap_count(reason);
3688 tty->print("too many traps=%s tcount=%d in ",
3689 Deoptimization::trap_reason_name(reason), tc);
3690 method()->print(); // which method has too many predicate traps
3691 tty->cr();
3692 }
3693 #endif
3694 // We cannot afford to take more traps here,
3695 // do not generate predicate.
3696 return;
3697 }
3698
3699 Node *cont = _gvn.intcon(1);
3700 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3701 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3702 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3703 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3704 C->add_predicate_opaq(opq);
3705 {
3706 PreserveJVMState pjvms(this);
3707 set_control(iffalse);
3708 inc_sp(nargs);
3709 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3710 }
3711 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3712 set_control(iftrue);
3713 }
3714
3715 //------------------------------add_predicate---------------------------------
3716 void GraphKit::add_predicate(int nargs) {
3717 if (UseLoopPredicate) {
3718 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3719 }
3720 // loop's limit check predicate should be near the loop.
3721 if (LoopLimitCheck) {
3722 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3723 }
3724 }
3725
3726 //----------------------------- store barriers ----------------------------
3727 #define __ ideal.
3728
3729 void GraphKit::sync_kit(IdealKit& ideal) {
3730 set_all_memory(__ merged_memory());
3731 set_i_o(__ i_o());
3732 set_control(__ ctrl());
3733 }
3734
3735 void GraphKit::final_sync(IdealKit& ideal) {
3736 // Final sync IdealKit and graphKit.
3737 sync_kit(ideal);
3738 }
3739
3740 // vanilla/CMS post barrier
3741 // Insert a write-barrier store. This is to let generational GC work; we have
3742 // to flag all oop-stores before the next GC point.
3743 void GraphKit::write_barrier_post(Node* oop_store,
3744 Node* obj,
3745 Node* adr,
3746 uint adr_idx,
3747 Node* val,
3748 bool use_precise) {
3749 // No store check needed if we're storing a NULL or an old object
3750 // (latter case is probably a string constant). The concurrent
3751 // mark sweep garbage collector, however, needs to have all nonNull
3752 // oop updates flagged via card-marks.
3753 if (val != NULL && val->is_Con()) {
3754 // must be either an oop or NULL
3755 const Type* t = val->bottom_type();
3756 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3757 // stores of null never (?) need barriers
3758 return;
3759 }
3760
3761 if (use_ReduceInitialCardMarks()
3762 && obj == just_allocated_object(control())) {
3763 // We can skip marks on a freshly-allocated object in Eden.
3764 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3765 // That routine informs GC to take appropriate compensating steps,
3766 // upon a slow-path allocation, so as to make this card-mark
3767 // elision safe.
3768 return;
3769 }
3770
3771 if (!use_precise) {
3772 // All card marks for a (non-array) instance are in one place:
3773 adr = obj;
3774 }
3775 // (Else it's an array (or unknown), and we want more precise card marks.)
3776 assert(adr != NULL, "");
3777
3778 IdealKit ideal(this, true);
3779
3780 // Convert the pointer to an int prior to doing math on it
3781 Node* cast = __ CastPX(__ ctrl(), adr);
3782
3783 // Divide by card size
3784 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3785 "Only one we handle so far.");
3786 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3787
3788 // Combine card table base and card offset
3789 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3790
3791 // Get the alias_index for raw card-mark memory
3792 int adr_type = Compile::AliasIdxRaw;
3793 Node* zero = __ ConI(0); // Dirty card value
3794 BasicType bt = T_BYTE;
3795
3796 if (UseCondCardMark) {
3797 // The classic GC reference write barrier is typically implemented
3798 // as a store into the global card mark table. Unfortunately
3799 // unconditional stores can result in false sharing and excessive
3800 // coherence traffic as well as false transactional aborts.
3801 // UseCondCardMark enables MP "polite" conditional card mark
3802 // stores. In theory we could relax the load from ctrl() to
3803 // no_ctrl, but that doesn't buy much latitude.
3804 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3805 __ if_then(card_val, BoolTest::ne, zero);
3806 }
3807
3808 // Smash zero into card
3809 if( !UseConcMarkSweepGC ) {
3810 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release);
3811 } else {
3812 // Specialized path for CM store barrier
3813 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3814 }
3815
3816 if (UseCondCardMark) {
3817 __ end_if();
3818 }
3819
3820 // Final sync IdealKit and GraphKit.
3821 final_sync(ideal);
3822 }
3823
3824 // G1 pre/post barriers
3825 void GraphKit::g1_write_barrier_pre(bool do_load,
3826 Node* obj,
3827 Node* adr,
3828 uint alias_idx,
3829 Node* val,
3830 const TypeOopPtr* val_type,
3831 Node* pre_val,
3832 BasicType bt) {
3833
3834 // Some sanity checks
3835 // Note: val is unused in this routine.
3836
3837 if (do_load) {
3838 // We need to generate the load of the previous value
3839 assert(obj != NULL, "must have a base");
3840 assert(adr != NULL, "where are loading from?");
3841 assert(pre_val == NULL, "loaded already?");
3842 assert(val_type != NULL, "need a type");
3843 } else {
3844 // In this case both val_type and alias_idx are unused.
3845 assert(pre_val != NULL, "must be loaded already");
3846 // Nothing to be done if pre_val is null.
3847 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3848 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3849 }
3850 assert(bt == T_OBJECT, "or we shouldn't be here");
3851
3852 IdealKit ideal(this, true);
3853
3854 Node* tls = __ thread(); // ThreadLocalStorage
3855
3856 Node* no_ctrl = NULL;
3857 Node* no_base = __ top();
3858 Node* zero = __ ConI(0);
3859 Node* zeroX = __ ConX(0);
3860
3861 float likely = PROB_LIKELY(0.999);
3862 float unlikely = PROB_UNLIKELY(0.999);
3863
3864 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3865 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3866
3867 // Offsets into the thread
3868 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3869 PtrQueue::byte_offset_of_active());
3870 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3871 PtrQueue::byte_offset_of_index());
3872 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3873 PtrQueue::byte_offset_of_buf());
3874
3875 // Now the actual pointers into the thread
3876 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3877 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3878 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3879
3880 // Now some of the values
3881 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3882
3883 // if (!marking)
3884 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3885 BasicType index_bt = TypeX_X->basic_type();
3886 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3887 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3888
3889 if (do_load) {
3890 // load original value
3891 // alias_idx correct??
3892 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3893 }
3894
3895 // if (pre_val != NULL)
3896 __ if_then(pre_val, BoolTest::ne, null()); {
3897 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3898
3899 // is the queue for this thread full?
3900 __ if_then(index, BoolTest::ne, zeroX, likely); {
3901
3902 // decrement the index
3903 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3904
3905 // Now get the buffer location we will log the previous value into and store it
3906 Node *log_addr = __ AddP(no_base, buffer, next_index);
3907 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
3908 // update the index
3909 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
3910
3911 } __ else_(); {
3912
3913 // logging buffer is full, call the runtime
3914 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3915 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3916 } __ end_if(); // (!index)
3917 } __ end_if(); // (pre_val != NULL)
3918 } __ end_if(); // (!marking)
3919
3920 // Final sync IdealKit and GraphKit.
3921 final_sync(ideal);
3922 }
3923
3924 //
3925 // Update the card table and add card address to the queue
3926 //
3927 void GraphKit::g1_mark_card(IdealKit& ideal,
3928 Node* card_adr,
3929 Node* oop_store,
3930 uint oop_alias_idx,
3931 Node* index,
3932 Node* index_adr,
3933 Node* buffer,
3934 const TypeFunc* tf) {
3935
3936 Node* zero = __ ConI(0);
3937 Node* zeroX = __ ConX(0);
3938 Node* no_base = __ top();
3939 BasicType card_bt = T_BYTE;
3940 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3941 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3942
3943 // Now do the queue work
3944 __ if_then(index, BoolTest::ne, zeroX); {
3945
3946 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3947 Node* log_addr = __ AddP(no_base, buffer, next_index);
3948
3949 // Order, see storeCM.
3950 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
3951 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
3952
3953 } __ else_(); {
3954 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3955 } __ end_if();
3956
3957 }
3958
3959 void GraphKit::g1_write_barrier_post(Node* oop_store,
3960 Node* obj,
3961 Node* adr,
3962 uint alias_idx,
3963 Node* val,
3964 BasicType bt,
3965 bool use_precise) {
3966 // If we are writing a NULL then we need no post barrier
3967
3968 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3969 // Must be NULL
3970 const Type* t = val->bottom_type();
3971 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3972 // No post barrier if writing NULLx
3973 return;
3974 }
3975
3976 if (!use_precise) {
3977 // All card marks for a (non-array) instance are in one place:
3978 adr = obj;
3979 }
3980 // (Else it's an array (or unknown), and we want more precise card marks.)
3981 assert(adr != NULL, "");
3982
3983 IdealKit ideal(this, true);
3984
3985 Node* tls = __ thread(); // ThreadLocalStorage
3986
3987 Node* no_base = __ top();
3988 float likely = PROB_LIKELY(0.999);
3989 float unlikely = PROB_UNLIKELY(0.999);
3990 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
3991 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
3992 Node* zeroX = __ ConX(0);
3993
3994 // Get the alias_index for raw card-mark memory
3995 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3996
3997 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3998
3999 // Offsets into the thread
4000 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4001 PtrQueue::byte_offset_of_index());
4002 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4003 PtrQueue::byte_offset_of_buf());
4004
4005 // Pointers into the thread
4006
4007 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4008 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4009
4010 // Now some values
4011 // Use ctrl to avoid hoisting these values past a safepoint, which could
4012 // potentially reset these fields in the JavaThread.
4013 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4014 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4015
4016 // Convert the store obj pointer to an int prior to doing math on it
4017 // Must use ctrl to prevent "integerized oop" existing across safepoint
4018 Node* cast = __ CastPX(__ ctrl(), adr);
4019
4020 // Divide pointer by card size
4021 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4022
4023 // Combine card table base and card offset
4024 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4025
4026 // If we know the value being stored does it cross regions?
4027
4028 if (val != NULL) {
4029 // Does the store cause us to cross regions?
4030
4031 // Should be able to do an unsigned compare of region_size instead of
4032 // and extra shift. Do we have an unsigned compare??
4033 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4034 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4035
4036 // if (xor_res == 0) same region so skip
4037 __ if_then(xor_res, BoolTest::ne, zeroX); {
4038
4039 // No barrier if we are storing a NULL
4040 __ if_then(val, BoolTest::ne, null(), unlikely); {
4041
4042 // Ok must mark the card if not already dirty
4043
4044 // load the original value of the card
4045 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4046
4047 __ if_then(card_val, BoolTest::ne, young_card); {
4048 sync_kit(ideal);
4049 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4050 insert_mem_bar(Op_MemBarVolatile, oop_store);
4051 __ sync_kit(this);
4052
4053 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4054 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4055 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4056 } __ end_if();
4057 } __ end_if();
4058 } __ end_if();
4059 } __ end_if();
4060 } else {
4061 // Object.clone() instrinsic uses this path.
4062 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4063 }
4064
4065 // Final sync IdealKit and GraphKit.
4066 final_sync(ideal);
4067 }
4068 #undef __
4069
4070
4071
4072 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4073 if (java_lang_String::has_offset_field()) {
4074 int offset_offset = java_lang_String::offset_offset_in_bytes();
4075 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4076 false, NULL, 0);
4077 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4078 int offset_field_idx = C->get_alias_index(offset_field_type);
4079 return make_load(ctrl,
4080 basic_plus_adr(str, str, offset_offset),
4081 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
4082 } else {
4083 return intcon(0);
4084 }
4085 }
4086
4087 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4088 if (java_lang_String::has_count_field()) {
4089 int count_offset = java_lang_String::count_offset_in_bytes();
4090 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4091 false, NULL, 0);
4092 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4093 int count_field_idx = C->get_alias_index(count_field_type);
4094 return make_load(ctrl,
4095 basic_plus_adr(str, str, count_offset),
4096 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
4097 } else {
4098 return load_array_length(load_String_value(ctrl, str));
4099 }
4100 }
4101
4102 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4103 int value_offset = java_lang_String::value_offset_in_bytes();
4104 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4105 false, NULL, 0);
4106 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4107 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4108 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4109 ciTypeArrayKlass::make(T_CHAR), true, 0);
4110 int value_field_idx = C->get_alias_index(value_field_type);
4111 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4112 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4113 // String.value field is known to be @Stable.
4114 if (UseImplicitStableValues) {
4115 load = cast_array_to_stable(load, value_type);
4116 }
4117 return load;
4118 }
4119
4120 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4121 int offset_offset = java_lang_String::offset_offset_in_bytes();
4122 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4123 false, NULL, 0);
4124 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4125 int offset_field_idx = C->get_alias_index(offset_field_type);
4126 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4127 value, T_INT, offset_field_idx, MemNode::unordered);
4128 }
4129
4130 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4131 int value_offset = java_lang_String::value_offset_in_bytes();
4132 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4133 false, NULL, 0);
4134 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4135
4136 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4137 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
4138 }
4139
4140 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4141 int count_offset = java_lang_String::count_offset_in_bytes();
4142 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4143 false, NULL, 0);
4144 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4145 int count_field_idx = C->get_alias_index(count_field_type);
4146 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4147 value, T_INT, count_field_idx, MemNode::unordered);
4148 }
4149
4150 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4151 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4152 // assumption of CCP analysis.
4153 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4154 }