1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
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_map->req() == phi_map->req(), "matching maps");
337 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
338 Node* hidden_merge_mark = root();
339 Node* region = phi_map->control();
340 MergeMemNode* phi_mem = phi_map->merged_memory();
341 MergeMemNode* ex_mem = ex_map->merged_memory();
342 if (region->in(0) != hidden_merge_mark) {
343 // The control input is not (yet) a specially-marked region in phi_map.
344 // Make it so, and build some phis.
345 region = new (C) RegionNode(2);
346 _gvn.set_type(region, Type::CONTROL);
347 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
348 region->init_req(1, phi_map->control());
349 phi_map->set_control(region);
350 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
351 record_for_igvn(io_phi);
352 _gvn.set_type(io_phi, Type::ABIO);
353 phi_map->set_i_o(io_phi);
354 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
355 Node* m = mms.memory();
356 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
357 record_for_igvn(m_phi);
358 _gvn.set_type(m_phi, Type::MEMORY);
359 mms.set_memory(m_phi);
360 }
361 }
362
363 // Either or both of phi_map and ex_map might already be converted into phis.
364 Node* ex_control = ex_map->control();
365 // if there is special marking on ex_map also, we add multiple edges from src
366 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
367 // how wide was the destination phi_map, originally?
368 uint orig_width = region->req();
369
370 if (add_multiple) {
371 add_n_reqs(region, ex_control);
372 add_n_reqs(phi_map->i_o(), ex_map->i_o());
373 } else {
374 // ex_map has no merges, so we just add single edges everywhere
375 add_one_req(region, ex_control);
376 add_one_req(phi_map->i_o(), ex_map->i_o());
377 }
378 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
379 if (mms.is_empty()) {
380 // get a copy of the base memory, and patch some inputs into it
381 const TypePtr* adr_type = mms.adr_type(C);
382 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
383 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
384 mms.set_memory(phi);
385 // Prepare to append interesting stuff onto the newly sliced phi:
386 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
387 }
388 // Append stuff from ex_map:
389 if (add_multiple) {
390 add_n_reqs(mms.memory(), mms.memory2());
391 } else {
392 add_one_req(mms.memory(), mms.memory2());
393 }
394 }
395 uint limit = ex_map->req();
396 for (uint i = TypeFunc::Parms; i < limit; i++) {
397 // Skip everything in the JVMS after tos. (The ex_oop follows.)
398 if (i == tos) i = ex_jvms->monoff();
399 Node* src = ex_map->in(i);
400 Node* dst = phi_map->in(i);
401 if (src != dst) {
402 PhiNode* phi;
403 if (dst->in(0) != region) {
404 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
405 record_for_igvn(phi);
406 _gvn.set_type(phi, phi->type());
407 phi_map->set_req(i, dst);
408 // Prepare to append interesting stuff onto the new phi:
409 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
410 } else {
411 assert(dst->is_Phi(), "nobody else uses a hidden region");
412 phi = (PhiNode*)dst;
413 }
414 if (add_multiple && src->in(0) == ex_control) {
415 // Both are phis.
416 add_n_reqs(dst, src);
417 } else {
418 while (dst->req() < region->req()) add_one_req(dst, src);
419 }
420 const Type* srctype = _gvn.type(src);
421 if (phi->type() != srctype) {
422 const Type* dsttype = phi->type()->meet(srctype);
423 if (phi->type() != dsttype) {
424 phi->set_type(dsttype);
425 _gvn.set_type(phi, dsttype);
426 }
427 }
428 }
429 }
430 }
431
432 //--------------------------use_exception_state--------------------------------
433 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
434 if (failing()) { stop(); return top(); }
435 Node* region = phi_map->control();
436 Node* hidden_merge_mark = root();
437 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
438 Node* ex_oop = clear_saved_ex_oop(phi_map);
439 if (region->in(0) == hidden_merge_mark) {
440 // Special marking for internal ex-states. Process the phis now.
441 region->set_req(0, region); // now it's an ordinary region
442 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
443 // Note: Setting the jvms also sets the bci and sp.
444 set_control(_gvn.transform(region));
445 uint tos = jvms()->stkoff() + sp();
446 for (uint i = 1; i < tos; i++) {
447 Node* x = phi_map->in(i);
448 if (x->in(0) == region) {
449 assert(x->is_Phi(), "expected a special phi");
450 phi_map->set_req(i, _gvn.transform(x));
451 }
452 }
453 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
454 Node* x = mms.memory();
455 if (x->in(0) == region) {
456 assert(x->is_Phi(), "nobody else uses a hidden region");
457 mms.set_memory(_gvn.transform(x));
458 }
459 }
460 if (ex_oop->in(0) == region) {
461 assert(ex_oop->is_Phi(), "expected a special phi");
462 ex_oop = _gvn.transform(ex_oop);
463 }
464 } else {
465 set_jvms(phi_map->jvms());
466 }
467
468 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
469 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
470 return ex_oop;
471 }
472
473 //---------------------------------java_bc-------------------------------------
474 Bytecodes::Code GraphKit::java_bc() const {
475 ciMethod* method = this->method();
476 int bci = this->bci();
477 if (method != NULL && bci != InvocationEntryBci)
478 return method->java_code_at_bci(bci);
479 else
480 return Bytecodes::_illegal;
481 }
482
483 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
484 bool must_throw) {
485 // if the exception capability is set, then we will generate code
486 // to check the JavaThread.should_post_on_exceptions flag to see
487 // if we actually need to report exception events (for this
488 // thread). If we don't need to report exception events, we will
489 // take the normal fast path provided by add_exception_events. If
490 // exception event reporting is enabled for this thread, we will
491 // take the uncommon_trap in the BuildCutout below.
492
493 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
494 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
495 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
496 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false);
497
498 // Test the should_post_on_exceptions_flag vs. 0
499 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
500 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
501
502 // Branch to slow_path if should_post_on_exceptions_flag was true
503 { BuildCutout unless(this, tst, PROB_MAX);
504 // Do not try anything fancy if we're notifying the VM on every throw.
505 // Cf. case Bytecodes::_athrow in parse2.cpp.
506 uncommon_trap(reason, Deoptimization::Action_none,
507 (ciKlass*)NULL, (char*)NULL, must_throw);
508 }
509
510 }
511
512 //------------------------------builtin_throw----------------------------------
513 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
514 bool must_throw = true;
515
516 if (env()->jvmti_can_post_on_exceptions()) {
517 // check if we must post exception events, take uncommon trap if so
518 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
519 // here if should_post_on_exceptions is false
520 // continue on with the normal codegen
521 }
522
523 // If this particular condition has not yet happened at this
524 // bytecode, then use the uncommon trap mechanism, and allow for
525 // a future recompilation if several traps occur here.
526 // If the throw is hot, try to use a more complicated inline mechanism
527 // which keeps execution inside the compiled code.
528 bool treat_throw_as_hot = false;
529 ciMethodData* md = method()->method_data();
530
531 if (ProfileTraps) {
532 if (too_many_traps(reason)) {
533 treat_throw_as_hot = true;
534 }
535 // (If there is no MDO at all, assume it is early in
536 // execution, and that any deopts are part of the
537 // startup transient, and don't need to be remembered.)
538
539 // Also, if there is a local exception handler, treat all throws
540 // as hot if there has been at least one in this method.
541 if (C->trap_count(reason) != 0
542 && method()->method_data()->trap_count(reason) != 0
543 && has_ex_handler()) {
544 treat_throw_as_hot = true;
545 }
546 }
547
548 // If this throw happens frequently, an uncommon trap might cause
549 // a performance pothole. If there is a local exception handler,
550 // and if this particular bytecode appears to be deoptimizing often,
551 // let us handle the throw inline, with a preconstructed instance.
552 // Note: If the deopt count has blown up, the uncommon trap
553 // runtime is going to flush this nmethod, not matter what.
554 if (treat_throw_as_hot
555 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
556 // If the throw is local, we use a pre-existing instance and
557 // punt on the backtrace. This would lead to a missing backtrace
558 // (a repeat of 4292742) if the backtrace object is ever asked
559 // for its backtrace.
560 // Fixing this remaining case of 4292742 requires some flavor of
561 // escape analysis. Leave that for the future.
562 ciInstance* ex_obj = NULL;
563 switch (reason) {
564 case Deoptimization::Reason_null_check:
565 ex_obj = env()->NullPointerException_instance();
566 break;
567 case Deoptimization::Reason_div0_check:
568 ex_obj = env()->ArithmeticException_instance();
569 break;
570 case Deoptimization::Reason_range_check:
571 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
572 break;
573 case Deoptimization::Reason_class_check:
574 if (java_bc() == Bytecodes::_aastore) {
575 ex_obj = env()->ArrayStoreException_instance();
576 } else {
577 ex_obj = env()->ClassCastException_instance();
578 }
579 break;
580 }
581 if (failing()) { stop(); return; } // exception allocation might fail
582 if (ex_obj != NULL) {
583 // Cheat with a preallocated exception object.
584 if (C->log() != NULL)
585 C->log()->elem("hot_throw preallocated='1' reason='%s'",
586 Deoptimization::trap_reason_name(reason));
587 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
588 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
589
590 // Clear the detail message of the preallocated exception object.
591 // Weblogic sometimes mutates the detail message of exceptions
592 // using reflection.
593 int offset = java_lang_Throwable::get_detailMessage_offset();
594 const TypePtr* adr_typ = ex_con->add_offset(offset);
595
596 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
597 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
598 // Conservatively release stores of object references.
599 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, StoreNode::release);
600
601 add_exception_state(make_exception_state(ex_node));
602 return;
603 }
604 }
605
606 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
607 // It won't be much cheaper than bailing to the interp., since we'll
608 // have to pass up all the debug-info, and the runtime will have to
609 // create the stack trace.
610
611 // Usual case: Bail to interpreter.
612 // Reserve the right to recompile if we haven't seen anything yet.
613
614 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
615 if (treat_throw_as_hot
616 && (method()->method_data()->trap_recompiled_at(bci())
617 || C->too_many_traps(reason))) {
618 // We cannot afford to take more traps here. Suffer in the interpreter.
619 if (C->log() != NULL)
620 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
621 Deoptimization::trap_reason_name(reason),
622 C->trap_count(reason));
623 action = Deoptimization::Action_none;
624 }
625
626 // "must_throw" prunes the JVM state to include only the stack, if there
627 // are no local exception handlers. This should cut down on register
628 // allocation time and code size, by drastically reducing the number
629 // of in-edges on the call to the uncommon trap.
630
631 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
632 }
633
634
635 //----------------------------PreserveJVMState---------------------------------
636 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
637 debug_only(kit->verify_map());
638 _kit = kit;
639 _map = kit->map(); // preserve the map
640 _sp = kit->sp();
641 kit->set_map(clone_map ? kit->clone_map() : NULL);
642 #ifdef ASSERT
643 _bci = kit->bci();
644 Parse* parser = kit->is_Parse();
645 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
646 _block = block;
647 #endif
648 }
649 PreserveJVMState::~PreserveJVMState() {
650 GraphKit* kit = _kit;
651 #ifdef ASSERT
652 assert(kit->bci() == _bci, "bci must not shift");
653 Parse* parser = kit->is_Parse();
654 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
655 assert(block == _block, "block must not shift");
656 #endif
657 kit->set_map(_map);
658 kit->set_sp(_sp);
659 }
660
661
662 //-----------------------------BuildCutout-------------------------------------
663 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
664 : PreserveJVMState(kit)
665 {
666 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
667 SafePointNode* outer_map = _map; // preserved map is caller's
668 SafePointNode* inner_map = kit->map();
669 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
670 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
671 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
672 }
673 BuildCutout::~BuildCutout() {
674 GraphKit* kit = _kit;
675 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
676 }
677
678 //---------------------------PreserveReexecuteState----------------------------
679 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
680 assert(!kit->stopped(), "must call stopped() before");
681 _kit = kit;
682 _sp = kit->sp();
683 _reexecute = kit->jvms()->_reexecute;
684 }
685 PreserveReexecuteState::~PreserveReexecuteState() {
686 if (_kit->stopped()) return;
687 _kit->jvms()->_reexecute = _reexecute;
688 _kit->set_sp(_sp);
689 }
690
691 //------------------------------clone_map--------------------------------------
692 // Implementation of PreserveJVMState
693 //
694 // Only clone_map(...) here. If this function is only used in the
695 // PreserveJVMState class we may want to get rid of this extra
696 // function eventually and do it all there.
697
698 SafePointNode* GraphKit::clone_map() {
699 if (map() == NULL) return NULL;
700
701 // Clone the memory edge first
702 Node* mem = MergeMemNode::make(C, map()->memory());
703 gvn().set_type_bottom(mem);
704
705 SafePointNode *clonemap = (SafePointNode*)map()->clone();
706 JVMState* jvms = this->jvms();
707 JVMState* clonejvms = jvms->clone_shallow(C);
708 clonemap->set_memory(mem);
709 clonemap->set_jvms(clonejvms);
710 clonejvms->set_map(clonemap);
711 record_for_igvn(clonemap);
712 gvn().set_type_bottom(clonemap);
713 return clonemap;
714 }
715
716
717 //-----------------------------set_map_clone-----------------------------------
718 void GraphKit::set_map_clone(SafePointNode* m) {
719 _map = m;
720 _map = clone_map();
721 _map->set_next_exception(NULL);
722 debug_only(verify_map());
723 }
724
725
726 //----------------------------kill_dead_locals---------------------------------
727 // Detect any locals which are known to be dead, and force them to top.
728 void GraphKit::kill_dead_locals() {
729 // Consult the liveness information for the locals. If any
730 // of them are unused, then they can be replaced by top(). This
731 // should help register allocation time and cut down on the size
732 // of the deoptimization information.
733
734 // This call is made from many of the bytecode handling
735 // subroutines called from the Big Switch in do_one_bytecode.
736 // Every bytecode which might include a slow path is responsible
737 // for killing its dead locals. The more consistent we
738 // are about killing deads, the fewer useless phis will be
739 // constructed for them at various merge points.
740
741 // bci can be -1 (InvocationEntryBci). We return the entry
742 // liveness for the method.
743
744 if (method() == NULL || method()->code_size() == 0) {
745 // We are building a graph for a call to a native method.
746 // All locals are live.
747 return;
748 }
749
750 ResourceMark rm;
751
752 // Consult the liveness information for the locals. If any
753 // of them are unused, then they can be replaced by top(). This
754 // should help register allocation time and cut down on the size
755 // of the deoptimization information.
756 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
757
758 int len = (int)live_locals.size();
759 assert(len <= jvms()->loc_size(), "too many live locals");
760 for (int local = 0; local < len; local++) {
761 if (!live_locals.at(local)) {
762 set_local(local, top());
763 }
764 }
765 }
766
767 #ifdef ASSERT
768 //-------------------------dead_locals_are_killed------------------------------
769 // Return true if all dead locals are set to top in the map.
770 // Used to assert "clean" debug info at various points.
771 bool GraphKit::dead_locals_are_killed() {
772 if (method() == NULL || method()->code_size() == 0) {
773 // No locals need to be dead, so all is as it should be.
774 return true;
775 }
776
777 // Make sure somebody called kill_dead_locals upstream.
778 ResourceMark rm;
779 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
780 if (jvms->loc_size() == 0) continue; // no locals to consult
781 SafePointNode* map = jvms->map();
782 ciMethod* method = jvms->method();
783 int bci = jvms->bci();
784 if (jvms == this->jvms()) {
785 bci = this->bci(); // it might not yet be synched
786 }
787 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
788 int len = (int)live_locals.size();
789 if (!live_locals.is_valid() || len == 0)
790 // This method is trivial, or is poisoned by a breakpoint.
791 return true;
792 assert(len == jvms->loc_size(), "live map consistent with locals map");
793 for (int local = 0; local < len; local++) {
794 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
795 if (PrintMiscellaneous && (Verbose || WizardMode)) {
796 tty->print_cr("Zombie local %d: ", local);
797 jvms->dump();
798 }
799 return false;
800 }
801 }
802 }
803 return true;
804 }
805
806 #endif //ASSERT
807
808 // Helper function for enforcing certain bytecodes to reexecute if
809 // deoptimization happens
810 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
811 ciMethod* cur_method = jvms->method();
812 int cur_bci = jvms->bci();
813 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
814 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
815 return Interpreter::bytecode_should_reexecute(code) ||
816 is_anewarray && code == Bytecodes::_multianewarray;
817 // Reexecute _multianewarray bytecode which was replaced with
818 // sequence of [a]newarray. See Parse::do_multianewarray().
819 //
820 // Note: interpreter should not have it set since this optimization
821 // is limited by dimensions and guarded by flag so in some cases
822 // multianewarray() runtime calls will be generated and
823 // the bytecode should not be reexecutes (stack will not be reset).
824 } else
825 return false;
826 }
827
828 // Helper function for adding JVMState and debug information to node
829 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
830 // Add the safepoint edges to the call (or other safepoint).
831
832 // Make sure dead locals are set to top. This
833 // should help register allocation time and cut down on the size
834 // of the deoptimization information.
835 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
836
837 // Walk the inline list to fill in the correct set of JVMState's
838 // Also fill in the associated edges for each JVMState.
839
840 // If the bytecode needs to be reexecuted we need to put
841 // the arguments back on the stack.
842 const bool should_reexecute = jvms()->should_reexecute();
843 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
844
845 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
846 // undefined if the bci is different. This is normal for Parse but it
847 // should not happen for LibraryCallKit because only one bci is processed.
848 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
849 "in LibraryCallKit the reexecute bit should not change");
850
851 // If we are guaranteed to throw, we can prune everything but the
852 // input to the current bytecode.
853 bool can_prune_locals = false;
854 uint stack_slots_not_pruned = 0;
855 int inputs = 0, depth = 0;
856 if (must_throw) {
857 assert(method() == youngest_jvms->method(), "sanity");
858 if (compute_stack_effects(inputs, depth)) {
859 can_prune_locals = true;
860 stack_slots_not_pruned = inputs;
861 }
862 }
863
864 if (env()->jvmti_can_access_local_variables()) {
865 // At any safepoint, this method can get breakpointed, which would
866 // then require an immediate deoptimization.
867 can_prune_locals = false; // do not prune locals
868 stack_slots_not_pruned = 0;
869 }
870
871 // do not scribble on the input jvms
872 JVMState* out_jvms = youngest_jvms->clone_deep(C);
873 call->set_jvms(out_jvms); // Start jvms list for call node
874
875 // For a known set of bytecodes, the interpreter should reexecute them if
876 // deoptimization happens. We set the reexecute state for them here
877 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
878 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
879 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
880 }
881
882 // Presize the call:
883 DEBUG_ONLY(uint non_debug_edges = call->req());
884 call->add_req_batch(top(), youngest_jvms->debug_depth());
885 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
886
887 // Set up edges so that the call looks like this:
888 // Call [state:] ctl io mem fptr retadr
889 // [parms:] parm0 ... parmN
890 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
891 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
892 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
893 // Note that caller debug info precedes callee debug info.
894
895 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
896 uint debug_ptr = call->req();
897
898 // Loop over the map input edges associated with jvms, add them
899 // to the call node, & reset all offsets to match call node array.
900 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
901 uint debug_end = debug_ptr;
902 uint debug_start = debug_ptr - in_jvms->debug_size();
903 debug_ptr = debug_start; // back up the ptr
904
905 uint p = debug_start; // walks forward in [debug_start, debug_end)
906 uint j, k, l;
907 SafePointNode* in_map = in_jvms->map();
908 out_jvms->set_map(call);
909
910 if (can_prune_locals) {
911 assert(in_jvms->method() == out_jvms->method(), "sanity");
912 // If the current throw can reach an exception handler in this JVMS,
913 // then we must keep everything live that can reach that handler.
914 // As a quick and dirty approximation, we look for any handlers at all.
915 if (in_jvms->method()->has_exception_handlers()) {
916 can_prune_locals = false;
917 }
918 }
919
920 // Add the Locals
921 k = in_jvms->locoff();
922 l = in_jvms->loc_size();
923 out_jvms->set_locoff(p);
924 if (!can_prune_locals) {
925 for (j = 0; j < l; j++)
926 call->set_req(p++, in_map->in(k+j));
927 } else {
928 p += l; // already set to top above by add_req_batch
929 }
930
931 // Add the Expression Stack
932 k = in_jvms->stkoff();
933 l = in_jvms->sp();
934 out_jvms->set_stkoff(p);
935 if (!can_prune_locals) {
936 for (j = 0; j < l; j++)
937 call->set_req(p++, in_map->in(k+j));
938 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
939 // Divide stack into {S0,...,S1}, where S0 is set to top.
940 uint s1 = stack_slots_not_pruned;
941 stack_slots_not_pruned = 0; // for next iteration
942 if (s1 > l) s1 = l;
943 uint s0 = l - s1;
944 p += s0; // skip the tops preinstalled by add_req_batch
945 for (j = s0; j < l; j++)
946 call->set_req(p++, in_map->in(k+j));
947 } else {
948 p += l; // already set to top above by add_req_batch
949 }
950
951 // Add the Monitors
952 k = in_jvms->monoff();
953 l = in_jvms->mon_size();
954 out_jvms->set_monoff(p);
955 for (j = 0; j < l; j++)
956 call->set_req(p++, in_map->in(k+j));
957
958 // Copy any scalar object fields.
959 k = in_jvms->scloff();
960 l = in_jvms->scl_size();
961 out_jvms->set_scloff(p);
962 for (j = 0; j < l; j++)
963 call->set_req(p++, in_map->in(k+j));
964
965 // Finish the new jvms.
966 out_jvms->set_endoff(p);
967
968 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
969 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
970 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
971 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
972 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
973 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
974
975 // Update the two tail pointers in parallel.
976 out_jvms = out_jvms->caller();
977 in_jvms = in_jvms->caller();
978 }
979
980 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
981
982 // Test the correctness of JVMState::debug_xxx accessors:
983 assert(call->jvms()->debug_start() == non_debug_edges, "");
984 assert(call->jvms()->debug_end() == call->req(), "");
985 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
986 }
987
988 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
989 Bytecodes::Code code = java_bc();
990 if (code == Bytecodes::_wide) {
991 code = method()->java_code_at_bci(bci() + 1);
992 }
993
994 BasicType rtype = T_ILLEGAL;
995 int rsize = 0;
996
997 if (code != Bytecodes::_illegal) {
998 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
999 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1000 if (rtype < T_CONFLICT)
1001 rsize = type2size[rtype];
1002 }
1003
1004 switch (code) {
1005 case Bytecodes::_illegal:
1006 return false;
1007
1008 case Bytecodes::_ldc:
1009 case Bytecodes::_ldc_w:
1010 case Bytecodes::_ldc2_w:
1011 inputs = 0;
1012 break;
1013
1014 case Bytecodes::_dup: inputs = 1; break;
1015 case Bytecodes::_dup_x1: inputs = 2; break;
1016 case Bytecodes::_dup_x2: inputs = 3; break;
1017 case Bytecodes::_dup2: inputs = 2; break;
1018 case Bytecodes::_dup2_x1: inputs = 3; break;
1019 case Bytecodes::_dup2_x2: inputs = 4; break;
1020 case Bytecodes::_swap: inputs = 2; break;
1021 case Bytecodes::_arraylength: inputs = 1; break;
1022
1023 case Bytecodes::_getstatic:
1024 case Bytecodes::_putstatic:
1025 case Bytecodes::_getfield:
1026 case Bytecodes::_putfield:
1027 {
1028 bool ignored_will_link;
1029 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1030 int size = field->type()->size();
1031 bool is_get = (depth >= 0), is_static = (depth & 1);
1032 inputs = (is_static ? 0 : 1);
1033 if (is_get) {
1034 depth = size - inputs;
1035 } else {
1036 inputs += size; // putxxx pops the value from the stack
1037 depth = - inputs;
1038 }
1039 }
1040 break;
1041
1042 case Bytecodes::_invokevirtual:
1043 case Bytecodes::_invokespecial:
1044 case Bytecodes::_invokestatic:
1045 case Bytecodes::_invokedynamic:
1046 case Bytecodes::_invokeinterface:
1047 {
1048 bool ignored_will_link;
1049 ciSignature* declared_signature = NULL;
1050 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1051 assert(declared_signature != NULL, "cannot be null");
1052 inputs = declared_signature->arg_size_for_bc(code);
1053 int size = declared_signature->return_type()->size();
1054 depth = size - inputs;
1055 }
1056 break;
1057
1058 case Bytecodes::_multianewarray:
1059 {
1060 ciBytecodeStream iter(method());
1061 iter.reset_to_bci(bci());
1062 iter.next();
1063 inputs = iter.get_dimensions();
1064 assert(rsize == 1, "");
1065 depth = rsize - inputs;
1066 }
1067 break;
1068
1069 case Bytecodes::_ireturn:
1070 case Bytecodes::_lreturn:
1071 case Bytecodes::_freturn:
1072 case Bytecodes::_dreturn:
1073 case Bytecodes::_areturn:
1074 assert(rsize = -depth, "");
1075 inputs = rsize;
1076 break;
1077
1078 case Bytecodes::_jsr:
1079 case Bytecodes::_jsr_w:
1080 inputs = 0;
1081 depth = 1; // S.B. depth=1, not zero
1082 break;
1083
1084 default:
1085 // bytecode produces a typed result
1086 inputs = rsize - depth;
1087 assert(inputs >= 0, "");
1088 break;
1089 }
1090
1091 #ifdef ASSERT
1092 // spot check
1093 int outputs = depth + inputs;
1094 assert(outputs >= 0, "sanity");
1095 switch (code) {
1096 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1097 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1098 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1099 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1100 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1101 }
1102 #endif //ASSERT
1103
1104 return true;
1105 }
1106
1107
1108
1109 //------------------------------basic_plus_adr---------------------------------
1110 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1111 // short-circuit a common case
1112 if (offset == intcon(0)) return ptr;
1113 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1114 }
1115
1116 Node* GraphKit::ConvI2L(Node* offset) {
1117 // short-circuit a common case
1118 jint offset_con = find_int_con(offset, Type::OffsetBot);
1119 if (offset_con != Type::OffsetBot) {
1120 return longcon((jlong) offset_con);
1121 }
1122 return _gvn.transform( new (C) ConvI2LNode(offset));
1123 }
1124
1125 Node* GraphKit::ConvI2UL(Node* offset) {
1126 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1127 if (offset_con != (juint) Type::OffsetBot) {
1128 return longcon((julong) offset_con);
1129 }
1130 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset));
1131 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) );
1132 return _gvn.transform( new (C) AndLNode(conv, mask) );
1133 }
1134
1135 Node* GraphKit::ConvL2I(Node* offset) {
1136 // short-circuit a common case
1137 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1138 if (offset_con != (jlong)Type::OffsetBot) {
1139 return intcon((int) offset_con);
1140 }
1141 return _gvn.transform( new (C) ConvL2INode(offset));
1142 }
1143
1144 //-------------------------load_object_klass-----------------------------------
1145 Node* GraphKit::load_object_klass(Node* obj) {
1146 // Special-case a fresh allocation to avoid building nodes:
1147 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1148 if (akls != NULL) return akls;
1149 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1150 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1151 }
1152
1153 //-------------------------load_array_length-----------------------------------
1154 Node* GraphKit::load_array_length(Node* array) {
1155 // Special-case a fresh allocation to avoid building nodes:
1156 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1157 Node *alen;
1158 if (alloc == NULL) {
1159 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1160 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1161 } else {
1162 alen = alloc->Ideal_length();
1163 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1164 if (ccast != alen) {
1165 alen = _gvn.transform(ccast);
1166 }
1167 }
1168 return alen;
1169 }
1170
1171 //------------------------------do_null_check----------------------------------
1172 // Helper function to do a NULL pointer check. Returned value is
1173 // the incoming address with NULL casted away. You are allowed to use the
1174 // not-null value only if you are control dependent on the test.
1175 extern int explicit_null_checks_inserted,
1176 explicit_null_checks_elided;
1177 Node* GraphKit::null_check_common(Node* value, BasicType type,
1178 // optional arguments for variations:
1179 bool assert_null,
1180 Node* *null_control) {
1181 assert(!assert_null || null_control == NULL, "not both at once");
1182 if (stopped()) return top();
1183 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1184 // For some performance testing, we may wish to suppress null checking.
1185 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1186 return value;
1187 }
1188 explicit_null_checks_inserted++;
1189
1190 // Construct NULL check
1191 Node *chk = NULL;
1192 switch(type) {
1193 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1194 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1195 case T_ARRAY : // fall through
1196 type = T_OBJECT; // simplify further tests
1197 case T_OBJECT : {
1198 const Type *t = _gvn.type( value );
1199
1200 const TypeOopPtr* tp = t->isa_oopptr();
1201 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1202 // Only for do_null_check, not any of its siblings:
1203 && !assert_null && null_control == NULL) {
1204 // Usually, any field access or invocation on an unloaded oop type
1205 // will simply fail to link, since the statically linked class is
1206 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1207 // the static class is loaded but the sharper oop type is not.
1208 // Rather than checking for this obscure case in lots of places,
1209 // we simply observe that a null check on an unloaded class
1210 // will always be followed by a nonsense operation, so we
1211 // can just issue the uncommon trap here.
1212 // Our access to the unloaded class will only be correct
1213 // after it has been loaded and initialized, which requires
1214 // a trip through the interpreter.
1215 #ifndef PRODUCT
1216 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1217 #endif
1218 uncommon_trap(Deoptimization::Reason_unloaded,
1219 Deoptimization::Action_reinterpret,
1220 tp->klass(), "!loaded");
1221 return top();
1222 }
1223
1224 if (assert_null) {
1225 // See if the type is contained in NULL_PTR.
1226 // If so, then the value is already null.
1227 if (t->higher_equal(TypePtr::NULL_PTR)) {
1228 explicit_null_checks_elided++;
1229 return value; // Elided null assert quickly!
1230 }
1231 } else {
1232 // See if mixing in the NULL pointer changes type.
1233 // If so, then the NULL pointer was not allowed in the original
1234 // type. In other words, "value" was not-null.
1235 if (t->meet(TypePtr::NULL_PTR) != t) {
1236 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1237 explicit_null_checks_elided++;
1238 return value; // Elided null check quickly!
1239 }
1240 }
1241 chk = new (C) CmpPNode( value, null() );
1242 break;
1243 }
1244
1245 default:
1246 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1247 }
1248 assert(chk != NULL, "sanity check");
1249 chk = _gvn.transform(chk);
1250
1251 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1252 BoolNode *btst = new (C) BoolNode( chk, btest);
1253 Node *tst = _gvn.transform( btst );
1254
1255 //-----------
1256 // if peephole optimizations occurred, a prior test existed.
1257 // If a prior test existed, maybe it dominates as we can avoid this test.
1258 if (tst != btst && type == T_OBJECT) {
1259 // At this point we want to scan up the CFG to see if we can
1260 // find an identical test (and so avoid this test altogether).
1261 Node *cfg = control();
1262 int depth = 0;
1263 while( depth < 16 ) { // Limit search depth for speed
1264 if( cfg->Opcode() == Op_IfTrue &&
1265 cfg->in(0)->in(1) == tst ) {
1266 // Found prior test. Use "cast_not_null" to construct an identical
1267 // CastPP (and hence hash to) as already exists for the prior test.
1268 // Return that casted value.
1269 if (assert_null) {
1270 replace_in_map(value, null());
1271 return null(); // do not issue the redundant test
1272 }
1273 Node *oldcontrol = control();
1274 set_control(cfg);
1275 Node *res = cast_not_null(value);
1276 set_control(oldcontrol);
1277 explicit_null_checks_elided++;
1278 return res;
1279 }
1280 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1281 if (cfg == NULL) break; // Quit at region nodes
1282 depth++;
1283 }
1284 }
1285
1286 //-----------
1287 // Branch to failure if null
1288 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1289 Deoptimization::DeoptReason reason;
1290 if (assert_null)
1291 reason = Deoptimization::Reason_null_assert;
1292 else if (type == T_OBJECT)
1293 reason = Deoptimization::Reason_null_check;
1294 else
1295 reason = Deoptimization::Reason_div0_check;
1296
1297 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1298 // ciMethodData::has_trap_at will return a conservative -1 if any
1299 // must-be-null assertion has failed. This could cause performance
1300 // problems for a method after its first do_null_assert failure.
1301 // Consider using 'Reason_class_check' instead?
1302
1303 // To cause an implicit null check, we set the not-null probability
1304 // to the maximum (PROB_MAX). For an explicit check the probability
1305 // is set to a smaller value.
1306 if (null_control != NULL || too_many_traps(reason)) {
1307 // probability is less likely
1308 ok_prob = PROB_LIKELY_MAG(3);
1309 } else if (!assert_null &&
1310 (ImplicitNullCheckThreshold > 0) &&
1311 method() != NULL &&
1312 (method()->method_data()->trap_count(reason)
1313 >= (uint)ImplicitNullCheckThreshold)) {
1314 ok_prob = PROB_LIKELY_MAG(3);
1315 }
1316
1317 if (null_control != NULL) {
1318 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1319 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1320 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1321 if (null_true == top())
1322 explicit_null_checks_elided++;
1323 (*null_control) = null_true;
1324 } else {
1325 BuildCutout unless(this, tst, ok_prob);
1326 // Check for optimizer eliding test at parse time
1327 if (stopped()) {
1328 // Failure not possible; do not bother making uncommon trap.
1329 explicit_null_checks_elided++;
1330 } else if (assert_null) {
1331 uncommon_trap(reason,
1332 Deoptimization::Action_make_not_entrant,
1333 NULL, "assert_null");
1334 } else {
1335 replace_in_map(value, zerocon(type));
1336 builtin_throw(reason);
1337 }
1338 }
1339
1340 // Must throw exception, fall-thru not possible?
1341 if (stopped()) {
1342 return top(); // No result
1343 }
1344
1345 if (assert_null) {
1346 // Cast obj to null on this path.
1347 replace_in_map(value, zerocon(type));
1348 return zerocon(type);
1349 }
1350
1351 // Cast obj to not-null on this path, if there is no null_control.
1352 // (If there is a null_control, a non-null value may come back to haunt us.)
1353 if (type == T_OBJECT) {
1354 Node* cast = cast_not_null(value, false);
1355 if (null_control == NULL || (*null_control) == top())
1356 replace_in_map(value, cast);
1357 value = cast;
1358 }
1359
1360 return value;
1361 }
1362
1363
1364 //------------------------------cast_not_null----------------------------------
1365 // Cast obj to not-null on this path
1366 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1367 const Type *t = _gvn.type(obj);
1368 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1369 // Object is already not-null?
1370 if( t == t_not_null ) return obj;
1371
1372 Node *cast = new (C) CastPPNode(obj,t_not_null);
1373 cast->init_req(0, control());
1374 cast = _gvn.transform( cast );
1375
1376 // Scan for instances of 'obj' in the current JVM mapping.
1377 // These instances are known to be not-null after the test.
1378 if (do_replace_in_map)
1379 replace_in_map(obj, cast);
1380
1381 return cast; // Return casted value
1382 }
1383
1384
1385 //--------------------------replace_in_map-------------------------------------
1386 void GraphKit::replace_in_map(Node* old, Node* neww) {
1387 this->map()->replace_edge(old, neww);
1388
1389 // Note: This operation potentially replaces any edge
1390 // on the map. This includes locals, stack, and monitors
1391 // of the current (innermost) JVM state.
1392
1393 // We can consider replacing in caller maps.
1394 // The idea would be that an inlined function's null checks
1395 // can be shared with the entire inlining tree.
1396 // The expense of doing this is that the PreserveJVMState class
1397 // would have to preserve caller states too, with a deep copy.
1398 }
1399
1400
1401 //=============================================================================
1402 //--------------------------------memory---------------------------------------
1403 Node* GraphKit::memory(uint alias_idx) {
1404 MergeMemNode* mem = merged_memory();
1405 Node* p = mem->memory_at(alias_idx);
1406 _gvn.set_type(p, Type::MEMORY); // must be mapped
1407 return p;
1408 }
1409
1410 //-----------------------------reset_memory------------------------------------
1411 Node* GraphKit::reset_memory() {
1412 Node* mem = map()->memory();
1413 // do not use this node for any more parsing!
1414 debug_only( map()->set_memory((Node*)NULL) );
1415 return _gvn.transform( mem );
1416 }
1417
1418 //------------------------------set_all_memory---------------------------------
1419 void GraphKit::set_all_memory(Node* newmem) {
1420 Node* mergemem = MergeMemNode::make(C, newmem);
1421 gvn().set_type_bottom(mergemem);
1422 map()->set_memory(mergemem);
1423 }
1424
1425 //------------------------------set_all_memory_call----------------------------
1426 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1427 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1428 set_all_memory(newmem);
1429 }
1430
1431 //=============================================================================
1432 //
1433 // parser factory methods for MemNodes
1434 //
1435 // These are layered on top of the factory methods in LoadNode and StoreNode,
1436 // and integrate with the parser's memory state and _gvn engine.
1437 //
1438
1439 // factory methods in "int adr_idx"
1440 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1441 int adr_idx,
1442 bool require_atomic_access, LoadNode::Sem sem) {
1443 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1444 const TypePtr* adr_type = NULL; // debug-mode-only argument
1445 debug_only(adr_type = C->get_adr_type(adr_idx));
1446 Node* mem = memory(adr_idx);
1447 Node* ld;
1448 if (require_atomic_access && bt == T_LONG) {
1449 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, sem);
1450 } else {
1451 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, sem);
1452 }
1453 return _gvn.transform(ld);
1454 }
1455
1456 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1457 int adr_idx,
1458 bool require_atomic_access,
1459 StoreNode::Sem sem) {
1460 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1461 const TypePtr* adr_type = NULL;
1462 debug_only(adr_type = C->get_adr_type(adr_idx));
1463 Node *mem = memory(adr_idx);
1464 Node* st;
1465 if (require_atomic_access && bt == T_LONG) {
1466 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, sem);
1467 } else {
1468 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, sem);
1469 }
1470 st = _gvn.transform(st);
1471 set_memory(st, adr_idx);
1472 // Back-to-back stores can only remove intermediate store with DU info
1473 // so push on worklist for optimizer.
1474 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1475 record_for_igvn(st);
1476
1477 return st;
1478 }
1479
1480
1481 void GraphKit::pre_barrier(bool do_load,
1482 Node* ctl,
1483 Node* obj,
1484 Node* adr,
1485 uint adr_idx,
1486 Node* val,
1487 const TypeOopPtr* val_type,
1488 Node* pre_val,
1489 BasicType bt) {
1490
1491 BarrierSet* bs = Universe::heap()->barrier_set();
1492 set_control(ctl);
1493 switch (bs->kind()) {
1494 case BarrierSet::G1SATBCT:
1495 case BarrierSet::G1SATBCTLogging:
1496 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1497 break;
1498
1499 case BarrierSet::CardTableModRef:
1500 case BarrierSet::CardTableExtension:
1501 case BarrierSet::ModRef:
1502 break;
1503
1504 case BarrierSet::Other:
1505 default :
1506 ShouldNotReachHere();
1507
1508 }
1509 }
1510
1511 void GraphKit::post_barrier(Node* ctl,
1512 Node* store,
1513 Node* obj,
1514 Node* adr,
1515 uint adr_idx,
1516 Node* val,
1517 BasicType bt,
1518 bool use_precise) {
1519 BarrierSet* bs = Universe::heap()->barrier_set();
1520 set_control(ctl);
1521 switch (bs->kind()) {
1522 case BarrierSet::G1SATBCT:
1523 case BarrierSet::G1SATBCTLogging:
1524 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1525 break;
1526
1527 case BarrierSet::CardTableModRef:
1528 case BarrierSet::CardTableExtension:
1529 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1530 break;
1531
1532 case BarrierSet::ModRef:
1533 break;
1534
1535 case BarrierSet::Other:
1536 default :
1537 ShouldNotReachHere();
1538
1539 }
1540 }
1541
1542 Node* GraphKit::store_oop(Node* ctl,
1543 Node* obj,
1544 Node* adr,
1545 const TypePtr* adr_type,
1546 Node* val,
1547 const TypeOopPtr* val_type,
1548 BasicType bt,
1549 bool use_precise,
1550 StoreNode::Sem sem) {
1551 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1552 // could be delayed during Parse (for example, in adjust_map_after_if()).
1553 // Execute transformation here to avoid barrier generation in such case.
1554 if (_gvn.type(val) == TypePtr::NULL_PTR)
1555 val = _gvn.makecon(TypePtr::NULL_PTR);
1556
1557 set_control(ctl);
1558 if (stopped()) return top(); // Dead path ?
1559
1560 assert(bt == T_OBJECT, "sanity");
1561 assert(val != NULL, "not dead path");
1562 uint adr_idx = C->get_alias_index(adr_type);
1563 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1564
1565 pre_barrier(true /* do_load */,
1566 control(), obj, adr, adr_idx, val, val_type,
1567 NULL /* pre_val */,
1568 bt);
1569
1570 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, false, sem);
1571 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1572 return store;
1573 }
1574
1575 // Could be an array or object we don't know at compile time (unsafe ref.)
1576 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1577 Node* obj, // containing obj
1578 Node* adr, // actual adress to store val at
1579 const TypePtr* adr_type,
1580 Node* val,
1581 BasicType bt,
1582 StoreNode::Sem sem) {
1583 Compile::AliasType* at = C->alias_type(adr_type);
1584 const TypeOopPtr* val_type = NULL;
1585 if (adr_type->isa_instptr()) {
1586 if (at->field() != NULL) {
1587 // known field. This code is a copy of the do_put_xxx logic.
1588 ciField* field = at->field();
1589 if (!field->type()->is_loaded()) {
1590 val_type = TypeInstPtr::BOTTOM;
1591 } else {
1592 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1593 }
1594 }
1595 } else if (adr_type->isa_aryptr()) {
1596 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1597 }
1598 if (val_type == NULL) {
1599 val_type = TypeInstPtr::BOTTOM;
1600 }
1601 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, sem);
1602 }
1603
1604
1605 //-------------------------array_element_address-------------------------
1606 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1607 const TypeInt* sizetype) {
1608 uint shift = exact_log2(type2aelembytes(elembt));
1609 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1610
1611 // short-circuit a common case (saves lots of confusing waste motion)
1612 jint idx_con = find_int_con(idx, -1);
1613 if (idx_con >= 0) {
1614 intptr_t offset = header + ((intptr_t)idx_con << shift);
1615 return basic_plus_adr(ary, offset);
1616 }
1617
1618 // must be correct type for alignment purposes
1619 Node* base = basic_plus_adr(ary, header);
1620 #ifdef _LP64
1621 // The scaled index operand to AddP must be a clean 64-bit value.
1622 // Java allows a 32-bit int to be incremented to a negative
1623 // value, which appears in a 64-bit register as a large
1624 // positive number. Using that large positive number as an
1625 // operand in pointer arithmetic has bad consequences.
1626 // On the other hand, 32-bit overflow is rare, and the possibility
1627 // can often be excluded, if we annotate the ConvI2L node with
1628 // a type assertion that its value is known to be a small positive
1629 // number. (The prior range check has ensured this.)
1630 // This assertion is used by ConvI2LNode::Ideal.
1631 int index_max = max_jint - 1; // array size is max_jint, index is one less
1632 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1633 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1634 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1635 #endif
1636 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1637 return basic_plus_adr(ary, base, scale);
1638 }
1639
1640 //-------------------------load_array_element-------------------------
1641 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1642 const Type* elemtype = arytype->elem();
1643 BasicType elembt = elemtype->array_element_basic_type();
1644 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1645 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1646 return ld;
1647 }
1648
1649 //-------------------------set_arguments_for_java_call-------------------------
1650 // Arguments (pre-popped from the stack) are taken from the JVMS.
1651 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1652 // Add the call arguments:
1653 uint nargs = call->method()->arg_size();
1654 for (uint i = 0; i < nargs; i++) {
1655 Node* arg = argument(i);
1656 call->init_req(i + TypeFunc::Parms, arg);
1657 }
1658 }
1659
1660 //---------------------------set_edges_for_java_call---------------------------
1661 // Connect a newly created call into the current JVMS.
1662 // A return value node (if any) is returned from set_edges_for_java_call.
1663 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1664
1665 // Add the predefined inputs:
1666 call->init_req( TypeFunc::Control, control() );
1667 call->init_req( TypeFunc::I_O , i_o() );
1668 call->init_req( TypeFunc::Memory , reset_memory() );
1669 call->init_req( TypeFunc::FramePtr, frameptr() );
1670 call->init_req( TypeFunc::ReturnAdr, top() );
1671
1672 add_safepoint_edges(call, must_throw);
1673
1674 Node* xcall = _gvn.transform(call);
1675
1676 if (xcall == top()) {
1677 set_control(top());
1678 return;
1679 }
1680 assert(xcall == call, "call identity is stable");
1681
1682 // Re-use the current map to produce the result.
1683
1684 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1685 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1686 set_all_memory_call(xcall, separate_io_proj);
1687
1688 //return xcall; // no need, caller already has it
1689 }
1690
1691 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1692 if (stopped()) return top(); // maybe the call folded up?
1693
1694 // Capture the return value, if any.
1695 Node* ret;
1696 if (call->method() == NULL ||
1697 call->method()->return_type()->basic_type() == T_VOID)
1698 ret = top();
1699 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1700
1701 // Note: Since any out-of-line call can produce an exception,
1702 // we always insert an I_O projection from the call into the result.
1703
1704 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1705
1706 if (separate_io_proj) {
1707 // The caller requested separate projections be used by the fall
1708 // through and exceptional paths, so replace the projections for
1709 // the fall through path.
1710 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1711 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1712 }
1713 return ret;
1714 }
1715
1716 //--------------------set_predefined_input_for_runtime_call--------------------
1717 // Reading and setting the memory state is way conservative here.
1718 // The real problem is that I am not doing real Type analysis on memory,
1719 // so I cannot distinguish card mark stores from other stores. Across a GC
1720 // point the Store Barrier and the card mark memory has to agree. I cannot
1721 // have a card mark store and its barrier split across the GC point from
1722 // either above or below. Here I get that to happen by reading ALL of memory.
1723 // A better answer would be to separate out card marks from other memory.
1724 // For now, return the input memory state, so that it can be reused
1725 // after the call, if this call has restricted memory effects.
1726 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1727 // Set fixed predefined input arguments
1728 Node* memory = reset_memory();
1729 call->init_req( TypeFunc::Control, control() );
1730 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1731 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1732 call->init_req( TypeFunc::FramePtr, frameptr() );
1733 call->init_req( TypeFunc::ReturnAdr, top() );
1734 return memory;
1735 }
1736
1737 //-------------------set_predefined_output_for_runtime_call--------------------
1738 // Set control and memory (not i_o) from the call.
1739 // If keep_mem is not NULL, use it for the output state,
1740 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1741 // If hook_mem is NULL, this call produces no memory effects at all.
1742 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1743 // then only that memory slice is taken from the call.
1744 // In the last case, we must put an appropriate memory barrier before
1745 // the call, so as to create the correct anti-dependencies on loads
1746 // preceding the call.
1747 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1748 Node* keep_mem,
1749 const TypePtr* hook_mem) {
1750 // no i/o
1751 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1752 if (keep_mem) {
1753 // First clone the existing memory state
1754 set_all_memory(keep_mem);
1755 if (hook_mem != NULL) {
1756 // Make memory for the call
1757 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1758 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1759 // We also use hook_mem to extract specific effects from arraycopy stubs.
1760 set_memory(mem, hook_mem);
1761 }
1762 // ...else the call has NO memory effects.
1763
1764 // Make sure the call advertises its memory effects precisely.
1765 // This lets us build accurate anti-dependences in gcm.cpp.
1766 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1767 "call node must be constructed correctly");
1768 } else {
1769 assert(hook_mem == NULL, "");
1770 // This is not a "slow path" call; all memory comes from the call.
1771 set_all_memory_call(call);
1772 }
1773 }
1774
1775
1776 // Replace the call with the current state of the kit.
1777 void GraphKit::replace_call(CallNode* call, Node* result) {
1778 JVMState* ejvms = NULL;
1779 if (has_exceptions()) {
1780 ejvms = transfer_exceptions_into_jvms();
1781 }
1782
1783 SafePointNode* final_state = stop();
1784
1785 // Find all the needed outputs of this call
1786 CallProjections callprojs;
1787 call->extract_projections(&callprojs, true);
1788
1789 Node* init_mem = call->in(TypeFunc::Memory);
1790 Node* final_mem = final_state->in(TypeFunc::Memory);
1791 Node* final_ctl = final_state->in(TypeFunc::Control);
1792 Node* final_io = final_state->in(TypeFunc::I_O);
1793
1794 // Replace all the old call edges with the edges from the inlining result
1795 if (callprojs.fallthrough_catchproj != NULL) {
1796 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1797 }
1798 if (callprojs.fallthrough_memproj != NULL) {
1799 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1800 }
1801 if (callprojs.fallthrough_ioproj != NULL) {
1802 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1803 }
1804
1805 // Replace the result with the new result if it exists and is used
1806 if (callprojs.resproj != NULL && result != NULL) {
1807 C->gvn_replace_by(callprojs.resproj, result);
1808 }
1809
1810 if (ejvms == NULL) {
1811 // No exception edges to simply kill off those paths
1812 if (callprojs.catchall_catchproj != NULL) {
1813 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1814 }
1815 if (callprojs.catchall_memproj != NULL) {
1816 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1817 }
1818 if (callprojs.catchall_ioproj != NULL) {
1819 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1820 }
1821 // Replace the old exception object with top
1822 if (callprojs.exobj != NULL) {
1823 C->gvn_replace_by(callprojs.exobj, C->top());
1824 }
1825 } else {
1826 GraphKit ekit(ejvms);
1827
1828 // Load my combined exception state into the kit, with all phis transformed:
1829 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1830
1831 Node* ex_oop = ekit.use_exception_state(ex_map);
1832 if (callprojs.catchall_catchproj != NULL) {
1833 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1834 }
1835 if (callprojs.catchall_memproj != NULL) {
1836 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1837 }
1838 if (callprojs.catchall_ioproj != NULL) {
1839 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1840 }
1841
1842 // Replace the old exception object with the newly created one
1843 if (callprojs.exobj != NULL) {
1844 C->gvn_replace_by(callprojs.exobj, ex_oop);
1845 }
1846 }
1847
1848 // Disconnect the call from the graph
1849 call->disconnect_inputs(NULL, C);
1850 C->gvn_replace_by(call, C->top());
1851
1852 // Clean up any MergeMems that feed other MergeMems since the
1853 // optimizer doesn't like that.
1854 if (final_mem->is_MergeMem()) {
1855 Node_List wl;
1856 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1857 Node* m = i.get();
1858 if (m->is_MergeMem() && !wl.contains(m)) {
1859 wl.push(m);
1860 }
1861 }
1862 while (wl.size() > 0) {
1863 _gvn.transform(wl.pop());
1864 }
1865 }
1866 }
1867
1868
1869 //------------------------------increment_counter------------------------------
1870 // for statistics: increment a VM counter by 1
1871
1872 void GraphKit::increment_counter(address counter_addr) {
1873 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1874 increment_counter(adr1);
1875 }
1876
1877 void GraphKit::increment_counter(Node* counter_addr) {
1878 int adr_type = Compile::AliasIdxRaw;
1879 Node* ctrl = control();
1880 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type);
1881 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1882 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type );
1883 }
1884
1885
1886 //------------------------------uncommon_trap----------------------------------
1887 // Bail out to the interpreter in mid-method. Implemented by calling the
1888 // uncommon_trap blob. This helper function inserts a runtime call with the
1889 // right debug info.
1890 void GraphKit::uncommon_trap(int trap_request,
1891 ciKlass* klass, const char* comment,
1892 bool must_throw,
1893 bool keep_exact_action) {
1894 if (failing()) stop();
1895 if (stopped()) return; // trap reachable?
1896
1897 // Note: If ProfileTraps is true, and if a deopt. actually
1898 // occurs here, the runtime will make sure an MDO exists. There is
1899 // no need to call method()->ensure_method_data() at this point.
1900
1901 // Set the stack pointer to the right value for reexecution:
1902 set_sp(reexecute_sp());
1903
1904 #ifdef ASSERT
1905 if (!must_throw) {
1906 // Make sure the stack has at least enough depth to execute
1907 // the current bytecode.
1908 int inputs, ignored_depth;
1909 if (compute_stack_effects(inputs, ignored_depth)) {
1910 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1911 Bytecodes::name(java_bc()), sp(), inputs));
1912 }
1913 }
1914 #endif
1915
1916 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1917 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1918
1919 switch (action) {
1920 case Deoptimization::Action_maybe_recompile:
1921 case Deoptimization::Action_reinterpret:
1922 // Temporary fix for 6529811 to allow virtual calls to be sure they
1923 // get the chance to go from mono->bi->mega
1924 if (!keep_exact_action &&
1925 Deoptimization::trap_request_index(trap_request) < 0 &&
1926 too_many_recompiles(reason)) {
1927 // This BCI is causing too many recompilations.
1928 action = Deoptimization::Action_none;
1929 trap_request = Deoptimization::make_trap_request(reason, action);
1930 } else {
1931 C->set_trap_can_recompile(true);
1932 }
1933 break;
1934 case Deoptimization::Action_make_not_entrant:
1935 C->set_trap_can_recompile(true);
1936 break;
1937 #ifdef ASSERT
1938 case Deoptimization::Action_none:
1939 case Deoptimization::Action_make_not_compilable:
1940 break;
1941 default:
1942 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
1943 break;
1944 #endif
1945 }
1946
1947 if (TraceOptoParse) {
1948 char buf[100];
1949 tty->print_cr("Uncommon trap %s at bci:%d",
1950 Deoptimization::format_trap_request(buf, sizeof(buf),
1951 trap_request), bci());
1952 }
1953
1954 CompileLog* log = C->log();
1955 if (log != NULL) {
1956 int kid = (klass == NULL)? -1: log->identify(klass);
1957 log->begin_elem("uncommon_trap bci='%d'", bci());
1958 char buf[100];
1959 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1960 trap_request));
1961 if (kid >= 0) log->print(" klass='%d'", kid);
1962 if (comment != NULL) log->print(" comment='%s'", comment);
1963 log->end_elem();
1964 }
1965
1966 // Make sure any guarding test views this path as very unlikely
1967 Node *i0 = control()->in(0);
1968 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1969 IfNode *iff = i0->as_If();
1970 float f = iff->_prob; // Get prob
1971 if (control()->Opcode() == Op_IfTrue) {
1972 if (f > PROB_UNLIKELY_MAG(4))
1973 iff->_prob = PROB_MIN;
1974 } else {
1975 if (f < PROB_LIKELY_MAG(4))
1976 iff->_prob = PROB_MAX;
1977 }
1978 }
1979
1980 // Clear out dead values from the debug info.
1981 kill_dead_locals();
1982
1983 // Now insert the uncommon trap subroutine call
1984 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
1985 const TypePtr* no_memory_effects = NULL;
1986 // Pass the index of the class to be loaded
1987 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1988 (must_throw ? RC_MUST_THROW : 0),
1989 OptoRuntime::uncommon_trap_Type(),
1990 call_addr, "uncommon_trap", no_memory_effects,
1991 intcon(trap_request));
1992 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1993 "must extract request correctly from the graph");
1994 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1995
1996 call->set_req(TypeFunc::ReturnAdr, returnadr());
1997 // The debug info is the only real input to this call.
1998
1999 // Halt-and-catch fire here. The above call should never return!
2000 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2001 _gvn.set_type_bottom(halt);
2002 root()->add_req(halt);
2003
2004 stop_and_kill_map();
2005 }
2006
2007
2008 //--------------------------just_allocated_object------------------------------
2009 // Report the object that was just allocated.
2010 // It must be the case that there are no intervening safepoints.
2011 // We use this to determine if an object is so "fresh" that
2012 // it does not require card marks.
2013 Node* GraphKit::just_allocated_object(Node* current_control) {
2014 if (C->recent_alloc_ctl() == current_control)
2015 return C->recent_alloc_obj();
2016 return NULL;
2017 }
2018
2019
2020 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2021 // (Note: TypeFunc::make has a cache that makes this fast.)
2022 const TypeFunc* tf = TypeFunc::make(dest_method);
2023 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2024 for (int j = 0; j < nargs; j++) {
2025 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2026 if( targ->basic_type() == T_DOUBLE ) {
2027 // If any parameters are doubles, they must be rounded before
2028 // the call, dstore_rounding does gvn.transform
2029 Node *arg = argument(j);
2030 arg = dstore_rounding(arg);
2031 set_argument(j, arg);
2032 }
2033 }
2034 }
2035
2036 void GraphKit::round_double_result(ciMethod* dest_method) {
2037 // A non-strict method may return a double value which has an extended
2038 // exponent, but this must not be visible in a caller which is 'strict'
2039 // If a strict caller invokes a non-strict callee, round a double result
2040
2041 BasicType result_type = dest_method->return_type()->basic_type();
2042 assert( method() != NULL, "must have caller context");
2043 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2044 // Destination method's return value is on top of stack
2045 // dstore_rounding() does gvn.transform
2046 Node *result = pop_pair();
2047 result = dstore_rounding(result);
2048 push_pair(result);
2049 }
2050 }
2051
2052 // rounding for strict float precision conformance
2053 Node* GraphKit::precision_rounding(Node* n) {
2054 return UseStrictFP && _method->flags().is_strict()
2055 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2056 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2057 : n;
2058 }
2059
2060 // rounding for strict double precision conformance
2061 Node* GraphKit::dprecision_rounding(Node *n) {
2062 return UseStrictFP && _method->flags().is_strict()
2063 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2064 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2065 : n;
2066 }
2067
2068 // rounding for non-strict double stores
2069 Node* GraphKit::dstore_rounding(Node* n) {
2070 return Matcher::strict_fp_requires_explicit_rounding
2071 && UseSSE <= 1
2072 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2073 : n;
2074 }
2075
2076 //=============================================================================
2077 // Generate a fast path/slow path idiom. Graph looks like:
2078 // [foo] indicates that 'foo' is a parameter
2079 //
2080 // [in] NULL
2081 // \ /
2082 // CmpP
2083 // Bool ne
2084 // If
2085 // / \
2086 // True False-<2>
2087 // / |
2088 // / cast_not_null
2089 // Load | | ^
2090 // [fast_test] | |
2091 // gvn to opt_test | |
2092 // / \ | <1>
2093 // True False |
2094 // | \\ |
2095 // [slow_call] \[fast_result]
2096 // Ctl Val \ \
2097 // | \ \
2098 // Catch <1> \ \
2099 // / \ ^ \ \
2100 // Ex No_Ex | \ \
2101 // | \ \ | \ <2> \
2102 // ... \ [slow_res] | | \ [null_result]
2103 // \ \--+--+--- | |
2104 // \ | / \ | /
2105 // --------Region Phi
2106 //
2107 //=============================================================================
2108 // Code is structured as a series of driver functions all called 'do_XXX' that
2109 // call a set of helper functions. Helper functions first, then drivers.
2110
2111 //------------------------------null_check_oop---------------------------------
2112 // Null check oop. Set null-path control into Region in slot 3.
2113 // Make a cast-not-nullness use the other not-null control. Return cast.
2114 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2115 bool never_see_null) {
2116 // Initial NULL check taken path
2117 (*null_control) = top();
2118 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2119
2120 // Generate uncommon_trap:
2121 if (never_see_null && (*null_control) != top()) {
2122 // If we see an unexpected null at a check-cast we record it and force a
2123 // recompile; the offending check-cast will be compiled to handle NULLs.
2124 // If we see more than one offending BCI, then all checkcasts in the
2125 // method will be compiled to handle NULLs.
2126 PreserveJVMState pjvms(this);
2127 set_control(*null_control);
2128 replace_in_map(value, null());
2129 uncommon_trap(Deoptimization::Reason_null_check,
2130 Deoptimization::Action_make_not_entrant);
2131 (*null_control) = top(); // NULL path is dead
2132 }
2133
2134 // Cast away null-ness on the result
2135 return cast;
2136 }
2137
2138 //------------------------------opt_iff----------------------------------------
2139 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2140 // Return slow-path control.
2141 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2142 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2143
2144 // Fast path taken; set region slot 2
2145 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2146 region->init_req(2,fast_taken); // Capture fast-control
2147
2148 // Fast path not-taken, i.e. slow path
2149 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2150 return slow_taken;
2151 }
2152
2153 //-----------------------------make_runtime_call-------------------------------
2154 Node* GraphKit::make_runtime_call(int flags,
2155 const TypeFunc* call_type, address call_addr,
2156 const char* call_name,
2157 const TypePtr* adr_type,
2158 // The following parms are all optional.
2159 // The first NULL ends the list.
2160 Node* parm0, Node* parm1,
2161 Node* parm2, Node* parm3,
2162 Node* parm4, Node* parm5,
2163 Node* parm6, Node* parm7) {
2164 // Slow-path call
2165 bool is_leaf = !(flags & RC_NO_LEAF);
2166 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2167 if (call_name == NULL) {
2168 assert(!is_leaf, "must supply name for leaf");
2169 call_name = OptoRuntime::stub_name(call_addr);
2170 }
2171 CallNode* call;
2172 if (!is_leaf) {
2173 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2174 bci(), adr_type);
2175 } else if (flags & RC_NO_FP) {
2176 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2177 } else {
2178 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2179 }
2180
2181 // The following is similar to set_edges_for_java_call,
2182 // except that the memory effects of the call are restricted to AliasIdxRaw.
2183
2184 // Slow path call has no side-effects, uses few values
2185 bool wide_in = !(flags & RC_NARROW_MEM);
2186 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2187
2188 Node* prev_mem = NULL;
2189 if (wide_in) {
2190 prev_mem = set_predefined_input_for_runtime_call(call);
2191 } else {
2192 assert(!wide_out, "narrow in => narrow out");
2193 Node* narrow_mem = memory(adr_type);
2194 prev_mem = reset_memory();
2195 map()->set_memory(narrow_mem);
2196 set_predefined_input_for_runtime_call(call);
2197 }
2198
2199 // Hook each parm in order. Stop looking at the first NULL.
2200 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2201 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2202 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2203 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2204 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2205 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2206 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2207 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2208 /* close each nested if ===> */ } } } } } } } }
2209 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2210
2211 if (!is_leaf) {
2212 // Non-leaves can block and take safepoints:
2213 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2214 }
2215 // Non-leaves can throw exceptions:
2216 if (has_io) {
2217 call->set_req(TypeFunc::I_O, i_o());
2218 }
2219
2220 if (flags & RC_UNCOMMON) {
2221 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2222 // (An "if" probability corresponds roughly to an unconditional count.
2223 // Sort of.)
2224 call->set_cnt(PROB_UNLIKELY_MAG(4));
2225 }
2226
2227 Node* c = _gvn.transform(call);
2228 assert(c == call, "cannot disappear");
2229
2230 if (wide_out) {
2231 // Slow path call has full side-effects.
2232 set_predefined_output_for_runtime_call(call);
2233 } else {
2234 // Slow path call has few side-effects, and/or sets few values.
2235 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2236 }
2237
2238 if (has_io) {
2239 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2240 }
2241 return call;
2242
2243 }
2244
2245 //------------------------------merge_memory-----------------------------------
2246 // Merge memory from one path into the current memory state.
2247 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2248 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2249 Node* old_slice = mms.force_memory();
2250 Node* new_slice = mms.memory2();
2251 if (old_slice != new_slice) {
2252 PhiNode* phi;
2253 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2254 phi = new_slice->as_Phi();
2255 #ifdef ASSERT
2256 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2257 old_slice = old_slice->in(new_path);
2258 // Caller is responsible for ensuring that any pre-existing
2259 // phis are already aware of old memory.
2260 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2261 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2262 #endif
2263 mms.set_memory(phi);
2264 } else {
2265 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2266 _gvn.set_type(phi, Type::MEMORY);
2267 phi->set_req(new_path, new_slice);
2268 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2269 }
2270 }
2271 }
2272 }
2273
2274 //------------------------------make_slow_call_ex------------------------------
2275 // Make the exception handler hookups for the slow call
2276 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2277 if (stopped()) return;
2278
2279 // Make a catch node with just two handlers: fall-through and catch-all
2280 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2281 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2282 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2283 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2284
2285 { PreserveJVMState pjvms(this);
2286 set_control(excp);
2287 set_i_o(i_o);
2288
2289 if (excp != top()) {
2290 // Create an exception state also.
2291 // Use an exact type if the caller has specified a specific exception.
2292 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2293 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2294 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2295 }
2296 }
2297
2298 // Get the no-exception control from the CatchNode.
2299 set_control(norm);
2300 }
2301
2302
2303 //-------------------------------gen_subtype_check-----------------------------
2304 // Generate a subtyping check. Takes as input the subtype and supertype.
2305 // Returns 2 values: sets the default control() to the true path and returns
2306 // the false path. Only reads invariant memory; sets no (visible) memory.
2307 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2308 // but that's not exposed to the optimizer. This call also doesn't take in an
2309 // Object; if you wish to check an Object you need to load the Object's class
2310 // prior to coming here.
2311 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2312 // Fast check for identical types, perhaps identical constants.
2313 // The types can even be identical non-constants, in cases
2314 // involving Array.newInstance, Object.clone, etc.
2315 if (subklass == superklass)
2316 return top(); // false path is dead; no test needed.
2317
2318 if (_gvn.type(superklass)->singleton()) {
2319 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2320 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2321
2322 // In the common case of an exact superklass, try to fold up the
2323 // test before generating code. You may ask, why not just generate
2324 // the code and then let it fold up? The answer is that the generated
2325 // code will necessarily include null checks, which do not always
2326 // completely fold away. If they are also needless, then they turn
2327 // into a performance loss. Example:
2328 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2329 // Here, the type of 'fa' is often exact, so the store check
2330 // of fa[1]=x will fold up, without testing the nullness of x.
2331 switch (static_subtype_check(superk, subk)) {
2332 case SSC_always_false:
2333 {
2334 Node* always_fail = control();
2335 set_control(top());
2336 return always_fail;
2337 }
2338 case SSC_always_true:
2339 return top();
2340 case SSC_easy_test:
2341 {
2342 // Just do a direct pointer compare and be done.
2343 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2344 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2345 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2346 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2347 return _gvn.transform( new(C) IfFalseNode(iff) );
2348 }
2349 case SSC_full_test:
2350 break;
2351 default:
2352 ShouldNotReachHere();
2353 }
2354 }
2355
2356 // %%% Possible further optimization: Even if the superklass is not exact,
2357 // if the subklass is the unique subtype of the superklass, the check
2358 // will always succeed. We could leave a dependency behind to ensure this.
2359
2360 // First load the super-klass's check-offset
2361 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2362 Node *chk_off = _gvn.transform( new (C) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2363 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2364 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2365
2366 // Load from the sub-klass's super-class display list, or a 1-word cache of
2367 // the secondary superclass list, or a failing value with a sentinel offset
2368 // if the super-klass is an interface or exceptionally deep in the Java
2369 // hierarchy and we have to scan the secondary superclass list the hard way.
2370 // Worst-case type is a little odd: NULL is allowed as a result (usually
2371 // klass loads can never produce a NULL).
2372 Node *chk_off_X = ConvI2X(chk_off);
2373 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2374 // For some types like interfaces the following loadKlass is from a 1-word
2375 // cache which is mutable so can't use immutable memory. Other
2376 // types load from the super-class display table which is immutable.
2377 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2378 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2379
2380 // Compile speed common case: ARE a subtype and we canNOT fail
2381 if( superklass == nkls )
2382 return top(); // false path is dead; no test needed.
2383
2384 // See if we get an immediate positive hit. Happens roughly 83% of the
2385 // time. Test to see if the value loaded just previously from the subklass
2386 // is exactly the superklass.
2387 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2388 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2389 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2390 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2391 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2392
2393 // Compile speed common case: Check for being deterministic right now. If
2394 // chk_off is a constant and not equal to cacheoff then we are NOT a
2395 // subklass. In this case we need exactly the 1 test above and we can
2396 // return those results immediately.
2397 if (!might_be_cache) {
2398 Node* not_subtype_ctrl = control();
2399 set_control(iftrue1); // We need exactly the 1 test above
2400 return not_subtype_ctrl;
2401 }
2402
2403 // Gather the various success & failures here
2404 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2405 record_for_igvn(r_ok_subtype);
2406 RegionNode *r_not_subtype = new (C) RegionNode(3);
2407 record_for_igvn(r_not_subtype);
2408
2409 r_ok_subtype->init_req(1, iftrue1);
2410
2411 // Check for immediate negative hit. Happens roughly 11% of the time (which
2412 // is roughly 63% of the remaining cases). Test to see if the loaded
2413 // check-offset points into the subklass display list or the 1-element
2414 // cache. If it points to the display (and NOT the cache) and the display
2415 // missed then it's not a subtype.
2416 Node *cacheoff = _gvn.intcon(cacheoff_con);
2417 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2418 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2419 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2420 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2421 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2422
2423 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2424 // No performance impact (too rare) but allows sharing of secondary arrays
2425 // which has some footprint reduction.
2426 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2427 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2428 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2429 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2430 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2431
2432 // -- Roads not taken here: --
2433 // We could also have chosen to perform the self-check at the beginning
2434 // of this code sequence, as the assembler does. This would not pay off
2435 // the same way, since the optimizer, unlike the assembler, can perform
2436 // static type analysis to fold away many successful self-checks.
2437 // Non-foldable self checks work better here in second position, because
2438 // the initial primary superclass check subsumes a self-check for most
2439 // types. An exception would be a secondary type like array-of-interface,
2440 // which does not appear in its own primary supertype display.
2441 // Finally, we could have chosen to move the self-check into the
2442 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2443 // dependent manner. But it is worthwhile to have the check here,
2444 // where it can be perhaps be optimized. The cost in code space is
2445 // small (register compare, branch).
2446
2447 // Now do a linear scan of the secondary super-klass array. Again, no real
2448 // performance impact (too rare) but it's gotta be done.
2449 // Since the code is rarely used, there is no penalty for moving it
2450 // out of line, and it can only improve I-cache density.
2451 // The decision to inline or out-of-line this final check is platform
2452 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2453 Node* psc = _gvn.transform(
2454 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2455
2456 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2457 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2458 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2459 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2460 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2461
2462 // Return false path; set default control to true path.
2463 set_control( _gvn.transform(r_ok_subtype) );
2464 return _gvn.transform(r_not_subtype);
2465 }
2466
2467 //----------------------------static_subtype_check-----------------------------
2468 // Shortcut important common cases when superklass is exact:
2469 // (0) superklass is java.lang.Object (can occur in reflective code)
2470 // (1) subklass is already limited to a subtype of superklass => always ok
2471 // (2) subklass does not overlap with superklass => always fail
2472 // (3) superklass has NO subtypes and we can check with a simple compare.
2473 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2474 if (StressReflectiveCode) {
2475 return SSC_full_test; // Let caller generate the general case.
2476 }
2477
2478 if (superk == env()->Object_klass()) {
2479 return SSC_always_true; // (0) this test cannot fail
2480 }
2481
2482 ciType* superelem = superk;
2483 if (superelem->is_array_klass())
2484 superelem = superelem->as_array_klass()->base_element_type();
2485
2486 if (!subk->is_interface()) { // cannot trust static interface types yet
2487 if (subk->is_subtype_of(superk)) {
2488 return SSC_always_true; // (1) false path dead; no dynamic test needed
2489 }
2490 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2491 !superk->is_subtype_of(subk)) {
2492 return SSC_always_false;
2493 }
2494 }
2495
2496 // If casting to an instance klass, it must have no subtypes
2497 if (superk->is_interface()) {
2498 // Cannot trust interfaces yet.
2499 // %%% S.B. superk->nof_implementors() == 1
2500 } else if (superelem->is_instance_klass()) {
2501 ciInstanceKlass* ik = superelem->as_instance_klass();
2502 if (!ik->has_subklass() && !ik->is_interface()) {
2503 if (!ik->is_final()) {
2504 // Add a dependency if there is a chance of a later subclass.
2505 C->dependencies()->assert_leaf_type(ik);
2506 }
2507 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2508 }
2509 } else {
2510 // A primitive array type has no subtypes.
2511 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2512 }
2513
2514 return SSC_full_test;
2515 }
2516
2517 // Profile-driven exact type check:
2518 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2519 float prob,
2520 Node* *casted_receiver) {
2521 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2522 Node* recv_klass = load_object_klass(receiver);
2523 Node* want_klass = makecon(tklass);
2524 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2525 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2526 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2527 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2528 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2529
2530 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2531 assert(recv_xtype->klass_is_exact(), "");
2532
2533 // Subsume downstream occurrences of receiver with a cast to
2534 // recv_xtype, since now we know what the type will be.
2535 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2536 (*casted_receiver) = _gvn.transform(cast);
2537 // (User must make the replace_in_map call.)
2538
2539 return fail;
2540 }
2541
2542
2543 //------------------------------seems_never_null-------------------------------
2544 // Use null_seen information if it is available from the profile.
2545 // If we see an unexpected null at a type check we record it and force a
2546 // recompile; the offending check will be recompiled to handle NULLs.
2547 // If we see several offending BCIs, then all checks in the
2548 // method will be recompiled.
2549 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2550 if (UncommonNullCast // Cutout for this technique
2551 && obj != null() // And not the -Xcomp stupid case?
2552 && !too_many_traps(Deoptimization::Reason_null_check)
2553 ) {
2554 if (data == NULL)
2555 // Edge case: no mature data. Be optimistic here.
2556 return true;
2557 // If the profile has not seen a null, assume it won't happen.
2558 assert(java_bc() == Bytecodes::_checkcast ||
2559 java_bc() == Bytecodes::_instanceof ||
2560 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2561 return !data->as_BitData()->null_seen();
2562 }
2563 return false;
2564 }
2565
2566 //------------------------maybe_cast_profiled_receiver-------------------------
2567 // If the profile has seen exactly one type, narrow to exactly that type.
2568 // Subsequent type checks will always fold up.
2569 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2570 ciProfileData* data,
2571 ciKlass* require_klass) {
2572 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2573 if (data == NULL) return NULL;
2574
2575 // Make sure we haven't already deoptimized from this tactic.
2576 if (too_many_traps(Deoptimization::Reason_class_check))
2577 return NULL;
2578
2579 // (No, this isn't a call, but it's enough like a virtual call
2580 // to use the same ciMethod accessor to get the profile info...)
2581 ciCallProfile profile = method()->call_profile_at_bci(bci());
2582 if (profile.count() >= 0 && // no cast failures here
2583 profile.has_receiver(0) &&
2584 profile.morphism() == 1) {
2585 ciKlass* exact_kls = profile.receiver(0);
2586 if (require_klass == NULL ||
2587 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2588 // If we narrow the type to match what the type profile sees,
2589 // we can then remove the rest of the cast.
2590 // This is a win, even if the exact_kls is very specific,
2591 // because downstream operations, such as method calls,
2592 // will often benefit from the sharper type.
2593 Node* exact_obj = not_null_obj; // will get updated in place...
2594 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2595 &exact_obj);
2596 { PreserveJVMState pjvms(this);
2597 set_control(slow_ctl);
2598 uncommon_trap(Deoptimization::Reason_class_check,
2599 Deoptimization::Action_maybe_recompile);
2600 }
2601 replace_in_map(not_null_obj, exact_obj);
2602 return exact_obj;
2603 }
2604 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2605 }
2606
2607 return NULL;
2608 }
2609
2610
2611 //-------------------------------gen_instanceof--------------------------------
2612 // Generate an instance-of idiom. Used by both the instance-of bytecode
2613 // and the reflective instance-of call.
2614 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) {
2615 kill_dead_locals(); // Benefit all the uncommon traps
2616 assert( !stopped(), "dead parse path should be checked in callers" );
2617 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2618 "must check for not-null not-dead klass in callers");
2619
2620 // Make the merge point
2621 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2622 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2623 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2624 C->set_has_split_ifs(true); // Has chance for split-if optimization
2625
2626 ciProfileData* data = NULL;
2627 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2628 data = method()->method_data()->bci_to_data(bci());
2629 }
2630 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2631 && seems_never_null(obj, data));
2632
2633 // Null check; get casted pointer; set region slot 3
2634 Node* null_ctl = top();
2635 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2636
2637 // If not_null_obj is dead, only null-path is taken
2638 if (stopped()) { // Doing instance-of on a NULL?
2639 set_control(null_ctl);
2640 return intcon(0);
2641 }
2642 region->init_req(_null_path, null_ctl);
2643 phi ->init_req(_null_path, intcon(0)); // Set null path value
2644 if (null_ctl == top()) {
2645 // Do this eagerly, so that pattern matches like is_diamond_phi
2646 // will work even during parsing.
2647 assert(_null_path == PATH_LIMIT-1, "delete last");
2648 region->del_req(_null_path);
2649 phi ->del_req(_null_path);
2650 }
2651
2652 if (ProfileDynamicTypes && data != NULL) {
2653 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL);
2654 if (stopped()) { // Profile disagrees with this path.
2655 set_control(null_ctl); // Null is the only remaining possibility.
2656 return intcon(0);
2657 }
2658 if (cast_obj != NULL)
2659 not_null_obj = cast_obj;
2660 }
2661
2662 // Load the object's klass
2663 Node* obj_klass = load_object_klass(not_null_obj);
2664
2665 // Generate the subtype check
2666 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2667
2668 // Plug in the success path to the general merge in slot 1.
2669 region->init_req(_obj_path, control());
2670 phi ->init_req(_obj_path, intcon(1));
2671
2672 // Plug in the failing path to the general merge in slot 2.
2673 region->init_req(_fail_path, not_subtype_ctrl);
2674 phi ->init_req(_fail_path, intcon(0));
2675
2676 // Return final merged results
2677 set_control( _gvn.transform(region) );
2678 record_for_igvn(region);
2679 return _gvn.transform(phi);
2680 }
2681
2682 //-------------------------------gen_checkcast---------------------------------
2683 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2684 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2685 // uncommon-trap paths work. Adjust stack after this call.
2686 // If failure_control is supplied and not null, it is filled in with
2687 // the control edge for the cast failure. Otherwise, an appropriate
2688 // uncommon trap or exception is thrown.
2689 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2690 Node* *failure_control) {
2691 kill_dead_locals(); // Benefit all the uncommon traps
2692 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2693 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2694
2695 // Fast cutout: Check the case that the cast is vacuously true.
2696 // This detects the common cases where the test will short-circuit
2697 // away completely. We do this before we perform the null check,
2698 // because if the test is going to turn into zero code, we don't
2699 // want a residual null check left around. (Causes a slowdown,
2700 // for example, in some objArray manipulations, such as a[i]=a[j].)
2701 if (tk->singleton()) {
2702 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2703 if (objtp != NULL && objtp->klass() != NULL) {
2704 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2705 case SSC_always_true:
2706 return obj;
2707 case SSC_always_false:
2708 // It needs a null check because a null will *pass* the cast check.
2709 // A non-null value will always produce an exception.
2710 return null_assert(obj);
2711 }
2712 }
2713 }
2714
2715 ciProfileData* data = NULL;
2716 if (failure_control == NULL) { // use MDO in regular case only
2717 assert(java_bc() == Bytecodes::_aastore ||
2718 java_bc() == Bytecodes::_checkcast,
2719 "interpreter profiles type checks only for these BCs");
2720 data = method()->method_data()->bci_to_data(bci());
2721 }
2722
2723 // Make the merge point
2724 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2725 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2726 Node* phi = new (C) PhiNode(region, toop);
2727 C->set_has_split_ifs(true); // Has chance for split-if optimization
2728
2729 // Use null-cast information if it is available
2730 bool never_see_null = ((failure_control == NULL) // regular case only
2731 && seems_never_null(obj, data));
2732
2733 // Null check; get casted pointer; set region slot 3
2734 Node* null_ctl = top();
2735 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2736
2737 // If not_null_obj is dead, only null-path is taken
2738 if (stopped()) { // Doing instance-of on a NULL?
2739 set_control(null_ctl);
2740 return null();
2741 }
2742 region->init_req(_null_path, null_ctl);
2743 phi ->init_req(_null_path, null()); // Set null path value
2744 if (null_ctl == top()) {
2745 // Do this eagerly, so that pattern matches like is_diamond_phi
2746 // will work even during parsing.
2747 assert(_null_path == PATH_LIMIT-1, "delete last");
2748 region->del_req(_null_path);
2749 phi ->del_req(_null_path);
2750 }
2751
2752 Node* cast_obj = NULL;
2753 if (tk->klass_is_exact()) {
2754 // The following optimization tries to statically cast the speculative type of the object
2755 // (for example obtained during profiling) to the type of the superklass and then do a
2756 // dynamic check that the type of the object is what we expect. To work correctly
2757 // for checkcast and aastore the type of superklass should be exact.
2758 if (data != NULL &&
2759 // Counter has never been decremented (due to cast failure).
2760 // ...This is a reasonable thing to expect. It is true of
2761 // all casts inserted by javac to implement generic types.
2762 data->as_CounterData()->count() >= 0) {
2763 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass());
2764 if (cast_obj != NULL) {
2765 if (failure_control != NULL) // failure is now impossible
2766 (*failure_control) = top();
2767 // adjust the type of the phi to the exact klass:
2768 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR));
2769 }
2770 }
2771 }
2772
2773 if (cast_obj == NULL) {
2774 // Load the object's klass
2775 Node* obj_klass = load_object_klass(not_null_obj);
2776
2777 // Generate the subtype check
2778 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2779
2780 // Plug in success path into the merge
2781 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
2782 not_null_obj, toop));
2783 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2784 if (failure_control == NULL) {
2785 if (not_subtype_ctrl != top()) { // If failure is possible
2786 PreserveJVMState pjvms(this);
2787 set_control(not_subtype_ctrl);
2788 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2789 }
2790 } else {
2791 (*failure_control) = not_subtype_ctrl;
2792 }
2793 }
2794
2795 region->init_req(_obj_path, control());
2796 phi ->init_req(_obj_path, cast_obj);
2797
2798 // A merge of NULL or Casted-NotNull obj
2799 Node* res = _gvn.transform(phi);
2800
2801 // Note I do NOT always 'replace_in_map(obj,result)' here.
2802 // if( tk->klass()->can_be_primary_super() )
2803 // This means that if I successfully store an Object into an array-of-String
2804 // I 'forget' that the Object is really now known to be a String. I have to
2805 // do this because we don't have true union types for interfaces - if I store
2806 // a Baz into an array-of-Interface and then tell the optimizer it's an
2807 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2808 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2809 // replace_in_map( obj, res );
2810
2811 // Return final merged results
2812 set_control( _gvn.transform(region) );
2813 record_for_igvn(region);
2814 return res;
2815 }
2816
2817 //------------------------------next_monitor-----------------------------------
2818 // What number should be given to the next monitor?
2819 int GraphKit::next_monitor() {
2820 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2821 int next = current + C->sync_stack_slots();
2822 // Keep the toplevel high water mark current:
2823 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2824 return current;
2825 }
2826
2827 //------------------------------insert_mem_bar---------------------------------
2828 // Memory barrier to avoid floating things around
2829 // The membar serves as a pinch point between both control and all memory slices.
2830 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2831 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2832 mb->init_req(TypeFunc::Control, control());
2833 mb->init_req(TypeFunc::Memory, reset_memory());
2834 Node* membar = _gvn.transform(mb);
2835 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
2836 set_all_memory_call(membar);
2837 return membar;
2838 }
2839
2840 //-------------------------insert_mem_bar_volatile----------------------------
2841 // Memory barrier to avoid floating things around
2842 // The membar serves as a pinch point between both control and memory(alias_idx).
2843 // If you want to make a pinch point on all memory slices, do not use this
2844 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2845 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2846 // When Parse::do_put_xxx updates a volatile field, it appends a series
2847 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2848 // The first membar is on the same memory slice as the field store opcode.
2849 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2850 // All the other membars (for other volatile slices, including AliasIdxBot,
2851 // which stands for all unknown volatile slices) are control-dependent
2852 // on the first membar. This prevents later volatile loads or stores
2853 // from sliding up past the just-emitted store.
2854
2855 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2856 mb->set_req(TypeFunc::Control,control());
2857 if (alias_idx == Compile::AliasIdxBot) {
2858 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2859 } else {
2860 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2861 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2862 }
2863 Node* membar = _gvn.transform(mb);
2864 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
2865 if (alias_idx == Compile::AliasIdxBot) {
2866 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
2867 } else {
2868 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2869 }
2870 return membar;
2871 }
2872
2873 //------------------------------shared_lock------------------------------------
2874 // Emit locking code.
2875 FastLockNode* GraphKit::shared_lock(Node* obj) {
2876 // bci is either a monitorenter bc or InvocationEntryBci
2877 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2878 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2879
2880 if( !GenerateSynchronizationCode )
2881 return NULL; // Not locking things?
2882 if (stopped()) // Dead monitor?
2883 return NULL;
2884
2885 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2886
2887 // Box the stack location
2888 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
2889 Node* mem = reset_memory();
2890
2891 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
2892 if (PrintPreciseBiasedLockingStatistics) {
2893 // Create the counters for this fast lock.
2894 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2895 }
2896 // Add monitor to debug info for the slow path. If we block inside the
2897 // slow path and de-opt, we need the monitor hanging around
2898 map()->push_monitor( flock );
2899
2900 const TypeFunc *tf = LockNode::lock_type();
2901 LockNode *lock = new (C) LockNode(C, tf);
2902
2903 lock->init_req( TypeFunc::Control, control() );
2904 lock->init_req( TypeFunc::Memory , mem );
2905 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2906 lock->init_req( TypeFunc::FramePtr, frameptr() );
2907 lock->init_req( TypeFunc::ReturnAdr, top() );
2908
2909 lock->init_req(TypeFunc::Parms + 0, obj);
2910 lock->init_req(TypeFunc::Parms + 1, box);
2911 lock->init_req(TypeFunc::Parms + 2, flock);
2912 add_safepoint_edges(lock);
2913
2914 lock = _gvn.transform( lock )->as_Lock();
2915
2916 // lock has no side-effects, sets few values
2917 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2918
2919 insert_mem_bar(Op_MemBarAcquireLock);
2920
2921 // Add this to the worklist so that the lock can be eliminated
2922 record_for_igvn(lock);
2923
2924 #ifndef PRODUCT
2925 if (PrintLockStatistics) {
2926 // Update the counter for this lock. Don't bother using an atomic
2927 // operation since we don't require absolute accuracy.
2928 lock->create_lock_counter(map()->jvms());
2929 increment_counter(lock->counter()->addr());
2930 }
2931 #endif
2932
2933 return flock;
2934 }
2935
2936
2937 //------------------------------shared_unlock----------------------------------
2938 // Emit unlocking code.
2939 void GraphKit::shared_unlock(Node* box, Node* obj) {
2940 // bci is either a monitorenter bc or InvocationEntryBci
2941 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2942 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2943
2944 if( !GenerateSynchronizationCode )
2945 return;
2946 if (stopped()) { // Dead monitor?
2947 map()->pop_monitor(); // Kill monitor from debug info
2948 return;
2949 }
2950
2951 // Memory barrier to avoid floating things down past the locked region
2952 insert_mem_bar(Op_MemBarReleaseLock);
2953
2954 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2955 UnlockNode *unlock = new (C) UnlockNode(C, tf);
2956 uint raw_idx = Compile::AliasIdxRaw;
2957 unlock->init_req( TypeFunc::Control, control() );
2958 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2959 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2960 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2961 unlock->init_req( TypeFunc::ReturnAdr, top() );
2962
2963 unlock->init_req(TypeFunc::Parms + 0, obj);
2964 unlock->init_req(TypeFunc::Parms + 1, box);
2965 unlock = _gvn.transform(unlock)->as_Unlock();
2966
2967 Node* mem = reset_memory();
2968
2969 // unlock has no side-effects, sets few values
2970 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2971
2972 // Kill monitor from debug info
2973 map()->pop_monitor( );
2974 }
2975
2976 //-------------------------------get_layout_helper-----------------------------
2977 // If the given klass is a constant or known to be an array,
2978 // fetch the constant layout helper value into constant_value
2979 // and return (Node*)NULL. Otherwise, load the non-constant
2980 // layout helper value, and return the node which represents it.
2981 // This two-faced routine is useful because allocation sites
2982 // almost always feature constant types.
2983 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2984 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2985 if (!StressReflectiveCode && inst_klass != NULL) {
2986 ciKlass* klass = inst_klass->klass();
2987 bool xklass = inst_klass->klass_is_exact();
2988 if (xklass || klass->is_array_klass()) {
2989 jint lhelper = klass->layout_helper();
2990 if (lhelper != Klass::_lh_neutral_value) {
2991 constant_value = lhelper;
2992 return (Node*) NULL;
2993 }
2994 }
2995 }
2996 constant_value = Klass::_lh_neutral_value; // put in a known value
2997 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
2998 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2999 }
3000
3001 // We just put in an allocate/initialize with a big raw-memory effect.
3002 // Hook selected additional alias categories on the initialization.
3003 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3004 MergeMemNode* init_in_merge,
3005 Node* init_out_raw) {
3006 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3007 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3008
3009 Node* prevmem = kit.memory(alias_idx);
3010 init_in_merge->set_memory_at(alias_idx, prevmem);
3011 kit.set_memory(init_out_raw, alias_idx);
3012 }
3013
3014 //---------------------------set_output_for_allocation-------------------------
3015 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3016 const TypeOopPtr* oop_type) {
3017 int rawidx = Compile::AliasIdxRaw;
3018 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3019 add_safepoint_edges(alloc);
3020 Node* allocx = _gvn.transform(alloc);
3021 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3022 // create memory projection for i_o
3023 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3024 make_slow_call_ex(allocx, env()->Throwable_klass(), true);
3025
3026 // create a memory projection as for the normal control path
3027 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3028 set_memory(malloc, rawidx);
3029
3030 // a normal slow-call doesn't change i_o, but an allocation does
3031 // we create a separate i_o projection for the normal control path
3032 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3033 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3034
3035 // put in an initialization barrier
3036 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3037 rawoop)->as_Initialize();
3038 assert(alloc->initialization() == init, "2-way macro link must work");
3039 assert(init ->allocation() == alloc, "2-way macro link must work");
3040 {
3041 // Extract memory strands which may participate in the new object's
3042 // initialization, and source them from the new InitializeNode.
3043 // This will allow us to observe initializations when they occur,
3044 // and link them properly (as a group) to the InitializeNode.
3045 assert(init->in(InitializeNode::Memory) == malloc, "");
3046 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3047 init->set_req(InitializeNode::Memory, minit_in);
3048 record_for_igvn(minit_in); // fold it up later, if possible
3049 Node* minit_out = memory(rawidx);
3050 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3051 if (oop_type->isa_aryptr()) {
3052 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3053 int elemidx = C->get_alias_index(telemref);
3054 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3055 } else if (oop_type->isa_instptr()) {
3056 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3057 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3058 ciField* field = ik->nonstatic_field_at(i);
3059 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3060 continue; // do not bother to track really large numbers of fields
3061 // Find (or create) the alias category for this field:
3062 int fieldidx = C->alias_type(field)->index();
3063 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3064 }
3065 }
3066 }
3067
3068 // Cast raw oop to the real thing...
3069 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3070 javaoop = _gvn.transform(javaoop);
3071 C->set_recent_alloc(control(), javaoop);
3072 assert(just_allocated_object(control()) == javaoop, "just allocated");
3073
3074 #ifdef ASSERT
3075 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3076 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3077 "Ideal_allocation works");
3078 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3079 "Ideal_allocation works");
3080 if (alloc->is_AllocateArray()) {
3081 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3082 "Ideal_allocation works");
3083 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3084 "Ideal_allocation works");
3085 } else {
3086 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3087 }
3088 }
3089 #endif //ASSERT
3090
3091 return javaoop;
3092 }
3093
3094 //---------------------------new_instance--------------------------------------
3095 // This routine takes a klass_node which may be constant (for a static type)
3096 // or may be non-constant (for reflective code). It will work equally well
3097 // for either, and the graph will fold nicely if the optimizer later reduces
3098 // the type to a constant.
3099 // The optional arguments are for specialized use by intrinsics:
3100 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3101 // - If 'return_size_val', report the the total object size to the caller.
3102 Node* GraphKit::new_instance(Node* klass_node,
3103 Node* extra_slow_test,
3104 Node* *return_size_val) {
3105 // Compute size in doublewords
3106 // The size is always an integral number of doublewords, represented
3107 // as a positive bytewise size stored in the klass's layout_helper.
3108 // The layout_helper also encodes (in a low bit) the need for a slow path.
3109 jint layout_con = Klass::_lh_neutral_value;
3110 Node* layout_val = get_layout_helper(klass_node, layout_con);
3111 int layout_is_con = (layout_val == NULL);
3112
3113 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3114 // Generate the initial go-slow test. It's either ALWAYS (return a
3115 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3116 // case) a computed value derived from the layout_helper.
3117 Node* initial_slow_test = NULL;
3118 if (layout_is_con) {
3119 assert(!StressReflectiveCode, "stress mode does not use these paths");
3120 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3121 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3122
3123 } else { // reflective case
3124 // This reflective path is used by Unsafe.allocateInstance.
3125 // (It may be stress-tested by specifying StressReflectiveCode.)
3126 // Basically, we want to get into the VM is there's an illegal argument.
3127 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3128 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3129 if (extra_slow_test != intcon(0)) {
3130 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3131 }
3132 // (Macro-expander will further convert this to a Bool, if necessary.)
3133 }
3134
3135 // Find the size in bytes. This is easy; it's the layout_helper.
3136 // The size value must be valid even if the slow path is taken.
3137 Node* size = NULL;
3138 if (layout_is_con) {
3139 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3140 } else { // reflective case
3141 // This reflective path is used by clone and Unsafe.allocateInstance.
3142 size = ConvI2X(layout_val);
3143
3144 // Clear the low bits to extract layout_helper_size_in_bytes:
3145 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3146 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3147 size = _gvn.transform( new (C) AndXNode(size, mask) );
3148 }
3149 if (return_size_val != NULL) {
3150 (*return_size_val) = size;
3151 }
3152
3153 // This is a precise notnull oop of the klass.
3154 // (Actually, it need not be precise if this is a reflective allocation.)
3155 // It's what we cast the result to.
3156 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3157 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3158 const TypeOopPtr* oop_type = tklass->as_instance_type();
3159
3160 // Now generate allocation code
3161
3162 // The entire memory state is needed for slow path of the allocation
3163 // since GC and deoptimization can happened.
3164 Node *mem = reset_memory();
3165 set_all_memory(mem); // Create new memory state
3166
3167 AllocateNode* alloc
3168 = new (C) AllocateNode(C, AllocateNode::alloc_type(),
3169 control(), mem, i_o(),
3170 size, klass_node,
3171 initial_slow_test);
3172
3173 return set_output_for_allocation(alloc, oop_type);
3174 }
3175
3176 //-------------------------------new_array-------------------------------------
3177 // helper for both newarray and anewarray
3178 // The 'length' parameter is (obviously) the length of the array.
3179 // See comments on new_instance for the meaning of the other arguments.
3180 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3181 Node* length, // number of array elements
3182 int nargs, // number of arguments to push back for uncommon trap
3183 Node* *return_size_val) {
3184 jint layout_con = Klass::_lh_neutral_value;
3185 Node* layout_val = get_layout_helper(klass_node, layout_con);
3186 int layout_is_con = (layout_val == NULL);
3187
3188 if (!layout_is_con && !StressReflectiveCode &&
3189 !too_many_traps(Deoptimization::Reason_class_check)) {
3190 // This is a reflective array creation site.
3191 // Optimistically assume that it is a subtype of Object[],
3192 // so that we can fold up all the address arithmetic.
3193 layout_con = Klass::array_layout_helper(T_OBJECT);
3194 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3195 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3196 { BuildCutout unless(this, bol_lh, PROB_MAX);
3197 inc_sp(nargs);
3198 uncommon_trap(Deoptimization::Reason_class_check,
3199 Deoptimization::Action_maybe_recompile);
3200 }
3201 layout_val = NULL;
3202 layout_is_con = true;
3203 }
3204
3205 // Generate the initial go-slow test. Make sure we do not overflow
3206 // if length is huge (near 2Gig) or negative! We do not need
3207 // exact double-words here, just a close approximation of needed
3208 // double-words. We can't add any offset or rounding bits, lest we
3209 // take a size -1 of bytes and make it positive. Use an unsigned
3210 // compare, so negative sizes look hugely positive.
3211 int fast_size_limit = FastAllocateSizeLimit;
3212 if (layout_is_con) {
3213 assert(!StressReflectiveCode, "stress mode does not use these paths");
3214 // Increase the size limit if we have exact knowledge of array type.
3215 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3216 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3217 }
3218
3219 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3220 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3221 if (initial_slow_test->is_Bool()) {
3222 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3223 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3224 }
3225
3226 // --- Size Computation ---
3227 // array_size = round_to_heap(array_header + (length << elem_shift));
3228 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3229 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3230 // The rounding mask is strength-reduced, if possible.
3231 int round_mask = MinObjAlignmentInBytes - 1;
3232 Node* header_size = NULL;
3233 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3234 // (T_BYTE has the weakest alignment and size restrictions...)
3235 if (layout_is_con) {
3236 int hsize = Klass::layout_helper_header_size(layout_con);
3237 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3238 BasicType etype = Klass::layout_helper_element_type(layout_con);
3239 if ((round_mask & ~right_n_bits(eshift)) == 0)
3240 round_mask = 0; // strength-reduce it if it goes away completely
3241 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3242 assert(header_size_min <= hsize, "generic minimum is smallest");
3243 header_size_min = hsize;
3244 header_size = intcon(hsize + round_mask);
3245 } else {
3246 Node* hss = intcon(Klass::_lh_header_size_shift);
3247 Node* hsm = intcon(Klass::_lh_header_size_mask);
3248 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3249 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3250 Node* mask = intcon(round_mask);
3251 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3252 }
3253
3254 Node* elem_shift = NULL;
3255 if (layout_is_con) {
3256 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3257 if (eshift != 0)
3258 elem_shift = intcon(eshift);
3259 } else {
3260 // There is no need to mask or shift this value.
3261 // The semantics of LShiftINode include an implicit mask to 0x1F.
3262 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3263 elem_shift = layout_val;
3264 }
3265
3266 // Transition to native address size for all offset calculations:
3267 Node* lengthx = ConvI2X(length);
3268 Node* headerx = ConvI2X(header_size);
3269 #ifdef _LP64
3270 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3271 if (tllen != NULL && tllen->_lo < 0) {
3272 // Add a manual constraint to a positive range. Cf. array_element_address.
3273 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3274 if (size_max > tllen->_hi) size_max = tllen->_hi;
3275 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3276 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3277 }
3278 }
3279 #endif
3280
3281 // Combine header size (plus rounding) and body size. Then round down.
3282 // This computation cannot overflow, because it is used only in two
3283 // places, one where the length is sharply limited, and the other
3284 // after a successful allocation.
3285 Node* abody = lengthx;
3286 if (elem_shift != NULL)
3287 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3288 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3289 if (round_mask != 0) {
3290 Node* mask = MakeConX(~round_mask);
3291 size = _gvn.transform( new(C) AndXNode(size, mask) );
3292 }
3293 // else if round_mask == 0, the size computation is self-rounding
3294
3295 if (return_size_val != NULL) {
3296 // This is the size
3297 (*return_size_val) = size;
3298 }
3299
3300 // Now generate allocation code
3301
3302 // The entire memory state is needed for slow path of the allocation
3303 // since GC and deoptimization can happened.
3304 Node *mem = reset_memory();
3305 set_all_memory(mem); // Create new memory state
3306
3307 // Create the AllocateArrayNode and its result projections
3308 AllocateArrayNode* alloc
3309 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3310 control(), mem, i_o(),
3311 size, klass_node,
3312 initial_slow_test,
3313 length);
3314
3315 // Cast to correct type. Note that the klass_node may be constant or not,
3316 // and in the latter case the actual array type will be inexact also.
3317 // (This happens via a non-constant argument to inline_native_newArray.)
3318 // In any case, the value of klass_node provides the desired array type.
3319 const TypeInt* length_type = _gvn.find_int_type(length);
3320 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3321 if (ary_type->isa_aryptr() && length_type != NULL) {
3322 // Try to get a better type than POS for the size
3323 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3324 }
3325
3326 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3327
3328 // Cast length on remaining path to be as narrow as possible
3329 if (map()->find_edge(length) >= 0) {
3330 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3331 if (ccast != length) {
3332 _gvn.set_type_bottom(ccast);
3333 record_for_igvn(ccast);
3334 replace_in_map(length, ccast);
3335 }
3336 }
3337
3338 return javaoop;
3339 }
3340
3341 // The following "Ideal_foo" functions are placed here because they recognize
3342 // the graph shapes created by the functions immediately above.
3343
3344 //---------------------------Ideal_allocation----------------------------------
3345 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3346 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3347 if (ptr == NULL) { // reduce dumb test in callers
3348 return NULL;
3349 }
3350 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3351 ptr = ptr->in(1);
3352 if (ptr == NULL) return NULL;
3353 }
3354 if (ptr->is_Proj()) {
3355 Node* allo = ptr->in(0);
3356 if (allo != NULL && allo->is_Allocate()) {
3357 return allo->as_Allocate();
3358 }
3359 }
3360 // Report failure to match.
3361 return NULL;
3362 }
3363
3364 // Fancy version which also strips off an offset (and reports it to caller).
3365 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3366 intptr_t& offset) {
3367 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3368 if (base == NULL) return NULL;
3369 return Ideal_allocation(base, phase);
3370 }
3371
3372 // Trace Initialize <- Proj[Parm] <- Allocate
3373 AllocateNode* InitializeNode::allocation() {
3374 Node* rawoop = in(InitializeNode::RawAddress);
3375 if (rawoop->is_Proj()) {
3376 Node* alloc = rawoop->in(0);
3377 if (alloc->is_Allocate()) {
3378 return alloc->as_Allocate();
3379 }
3380 }
3381 return NULL;
3382 }
3383
3384 // Trace Allocate -> Proj[Parm] -> Initialize
3385 InitializeNode* AllocateNode::initialization() {
3386 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3387 if (rawoop == NULL) return NULL;
3388 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3389 Node* init = rawoop->fast_out(i);
3390 if (init->is_Initialize()) {
3391 assert(init->as_Initialize()->allocation() == this, "2-way link");
3392 return init->as_Initialize();
3393 }
3394 }
3395 return NULL;
3396 }
3397
3398 // Trace Allocate -> Proj[Parm] -> MemBarStoreStore
3399 MemBarStoreStoreNode* AllocateNode::storestore() {
3400 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3401 if (rawoop == NULL) return NULL;
3402 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3403 Node* storestore = rawoop->fast_out(i);
3404 if (storestore->is_MemBarStoreStore()) {
3405 return storestore->as_MemBarStoreStore();
3406 }
3407 }
3408 return NULL;
3409 }
3410
3411 //----------------------------- loop predicates ---------------------------
3412
3413 //------------------------------add_predicate_impl----------------------------
3414 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3415 // Too many traps seen?
3416 if (too_many_traps(reason)) {
3417 #ifdef ASSERT
3418 if (TraceLoopPredicate) {
3419 int tc = C->trap_count(reason);
3420 tty->print("too many traps=%s tcount=%d in ",
3421 Deoptimization::trap_reason_name(reason), tc);
3422 method()->print(); // which method has too many predicate traps
3423 tty->cr();
3424 }
3425 #endif
3426 // We cannot afford to take more traps here,
3427 // do not generate predicate.
3428 return;
3429 }
3430
3431 Node *cont = _gvn.intcon(1);
3432 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3433 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3434 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3435 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3436 C->add_predicate_opaq(opq);
3437 {
3438 PreserveJVMState pjvms(this);
3439 set_control(iffalse);
3440 inc_sp(nargs);
3441 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3442 }
3443 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3444 set_control(iftrue);
3445 }
3446
3447 //------------------------------add_predicate---------------------------------
3448 void GraphKit::add_predicate(int nargs) {
3449 if (UseLoopPredicate) {
3450 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3451 }
3452 // loop's limit check predicate should be near the loop.
3453 if (LoopLimitCheck) {
3454 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3455 }
3456 }
3457
3458 //----------------------------- store barriers ----------------------------
3459 #define __ ideal.
3460
3461 void GraphKit::sync_kit(IdealKit& ideal) {
3462 set_all_memory(__ merged_memory());
3463 set_i_o(__ i_o());
3464 set_control(__ ctrl());
3465 }
3466
3467 void GraphKit::final_sync(IdealKit& ideal) {
3468 // Final sync IdealKit and graphKit.
3469 sync_kit(ideal);
3470 }
3471
3472 // vanilla/CMS post barrier
3473 // Insert a write-barrier store. This is to let generational GC work; we have
3474 // to flag all oop-stores before the next GC point.
3475 void GraphKit::write_barrier_post(Node* oop_store,
3476 Node* obj,
3477 Node* adr,
3478 uint adr_idx,
3479 Node* val,
3480 bool use_precise) {
3481 // No store check needed if we're storing a NULL or an old object
3482 // (latter case is probably a string constant). The concurrent
3483 // mark sweep garbage collector, however, needs to have all nonNull
3484 // oop updates flagged via card-marks.
3485 if (val != NULL && val->is_Con()) {
3486 // must be either an oop or NULL
3487 const Type* t = val->bottom_type();
3488 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3489 // stores of null never (?) need barriers
3490 return;
3491 ciObject* con = t->is_oopptr()->const_oop();
3492 if (con != NULL
3493 && con->is_perm()
3494 && Universe::heap()->can_elide_permanent_oop_store_barriers())
3495 // no store barrier needed, because no old-to-new ref created
3496 return;
3497 }
3498
3499 if (use_ReduceInitialCardMarks()
3500 && obj == just_allocated_object(control())) {
3501 // We can skip marks on a freshly-allocated object in Eden.
3502 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3503 // That routine informs GC to take appropriate compensating steps,
3504 // upon a slow-path allocation, so as to make this card-mark
3505 // elision safe.
3506 return;
3507 }
3508
3509 if (!use_precise) {
3510 // All card marks for a (non-array) instance are in one place:
3511 adr = obj;
3512 }
3513 // (Else it's an array (or unknown), and we want more precise card marks.)
3514 assert(adr != NULL, "");
3515
3516 IdealKit ideal(this, true);
3517
3518 // Convert the pointer to an int prior to doing math on it
3519 Node* cast = __ CastPX(__ ctrl(), adr);
3520
3521 // Divide by card size
3522 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3523 "Only one we handle so far.");
3524 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3525
3526 // Combine card table base and card offset
3527 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3528
3529 // Get the alias_index for raw card-mark memory
3530 int adr_type = Compile::AliasIdxRaw;
3531 Node* zero = __ ConI(0); // Dirty card value
3532 BasicType bt = T_BYTE;
3533
3534 if (UseCondCardMark) {
3535 // The classic GC reference write barrier is typically implemented
3536 // as a store into the global card mark table. Unfortunately
3537 // unconditional stores can result in false sharing and excessive
3538 // coherence traffic as well as false transactional aborts.
3539 // UseCondCardMark enables MP "polite" conditional card mark
3540 // stores. In theory we could relax the load from ctrl() to
3541 // no_ctrl, but that doesn't buy much latitude.
3542 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3543 __ if_then(card_val, BoolTest::ne, zero);
3544 }
3545
3546 // Smash zero into card
3547 if( !UseConcMarkSweepGC ) {
3548 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3549 } else {
3550 // Specialized path for CM store barrier
3551 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3552 }
3553
3554 if (UseCondCardMark) {
3555 __ end_if();
3556 }
3557
3558 // Final sync IdealKit and GraphKit.
3559 final_sync(ideal);
3560 }
3561
3562 // G1 pre/post barriers
3563 void GraphKit::g1_write_barrier_pre(bool do_load,
3564 Node* obj,
3565 Node* adr,
3566 uint alias_idx,
3567 Node* val,
3568 const TypeOopPtr* val_type,
3569 Node* pre_val,
3570 BasicType bt) {
3571
3572 // Some sanity checks
3573 // Note: val is unused in this routine.
3574
3575 if (do_load) {
3576 // We need to generate the load of the previous value
3577 assert(obj != NULL, "must have a base");
3578 assert(adr != NULL, "where are loading from?");
3579 assert(pre_val == NULL, "loaded already?");
3580 assert(val_type != NULL, "need a type");
3581 } else {
3582 // In this case both val_type and alias_idx are unused.
3583 assert(pre_val != NULL, "must be loaded already");
3584 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3585 }
3586 assert(bt == T_OBJECT, "or we shouldn't be here");
3587
3588 IdealKit ideal(this, true);
3589
3590 Node* tls = __ thread(); // ThreadLocalStorage
3591
3592 Node* no_ctrl = NULL;
3593 Node* no_base = __ top();
3594 Node* zero = __ ConI(0);
3595 Node* zeroX = __ ConX(0);
3596
3597 float likely = PROB_LIKELY(0.999);
3598 float unlikely = PROB_UNLIKELY(0.999);
3599
3600 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3601 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3602
3603 // Offsets into the thread
3604 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3605 PtrQueue::byte_offset_of_active());
3606 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3607 PtrQueue::byte_offset_of_index());
3608 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3609 PtrQueue::byte_offset_of_buf());
3610
3611 // Now the actual pointers into the thread
3612 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3613 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3614 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3615
3616 // Now some of the values
3617 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3618
3619 // if (!marking)
3620 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3621 BasicType index_bt = TypeX_X->basic_type();
3622 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3623 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3624
3625 if (do_load) {
3626 // load original value
3627 // alias_idx correct??
3628 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3629 }
3630
3631 // if (pre_val != NULL)
3632 __ if_then(pre_val, BoolTest::ne, null()); {
3633 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3634
3635 // is the queue for this thread full?
3636 __ if_then(index, BoolTest::ne, zeroX, likely); {
3637
3638 // decrement the index
3639 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3640
3641 // Now get the buffer location we will log the previous value into and store it
3642 Node *log_addr = __ AddP(no_base, buffer, next_index);
3643 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, StoreNode::release);
3644 // update the index
3645 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw);
3646
3647 } __ else_(); {
3648
3649 // logging buffer is full, call the runtime
3650 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3651 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3652 } __ end_if(); // (!index)
3653 } __ end_if(); // (pre_val != NULL)
3654 } __ end_if(); // (!marking)
3655
3656 // Final sync IdealKit and GraphKit.
3657 final_sync(ideal);
3658 }
3659
3660 //
3661 // Update the card table and add card address to the queue
3662 //
3663 void GraphKit::g1_mark_card(IdealKit& ideal,
3664 Node* card_adr,
3665 Node* oop_store,
3666 uint oop_alias_idx,
3667 Node* index,
3668 Node* index_adr,
3669 Node* buffer,
3670 const TypeFunc* tf) {
3671
3672 Node* zero = __ ConI(0);
3673 Node* zeroX = __ ConX(0);
3674 Node* no_base = __ top();
3675 BasicType card_bt = T_BYTE;
3676 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3677 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3678
3679 // Now do the queue work
3680 __ if_then(index, BoolTest::ne, zeroX); {
3681
3682 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3683 Node* log_addr = __ AddP(no_base, buffer, next_index);
3684
3685 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, StoreNode::release);
3686 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, StoreNode::release);
3687
3688 } __ else_(); {
3689 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3690 } __ end_if();
3691
3692 }
3693
3694 void GraphKit::g1_write_barrier_post(Node* oop_store,
3695 Node* obj,
3696 Node* adr,
3697 uint alias_idx,
3698 Node* val,
3699 BasicType bt,
3700 bool use_precise) {
3701 // If we are writing a NULL then we need no post barrier
3702
3703 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3704 // Must be NULL
3705 const Type* t = val->bottom_type();
3706 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3707 // No post barrier if writing NULLx
3708 return;
3709 }
3710
3711 if (!use_precise) {
3712 // All card marks for a (non-array) instance are in one place:
3713 adr = obj;
3714 }
3715 // (Else it's an array (or unknown), and we want more precise card marks.)
3716 assert(adr != NULL, "");
3717
3718 IdealKit ideal(this, true);
3719
3720 Node* tls = __ thread(); // ThreadLocalStorage
3721
3722 Node* no_base = __ top();
3723 float likely = PROB_LIKELY(0.999);
3724 float unlikely = PROB_UNLIKELY(0.999);
3725 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
3726 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
3727 Node* zeroX = __ ConX(0);
3728
3729 // Get the alias_index for raw card-mark memory
3730 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3731
3732 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3733
3734 // Offsets into the thread
3735 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3736 PtrQueue::byte_offset_of_index());
3737 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3738 PtrQueue::byte_offset_of_buf());
3739
3740 // Pointers into the thread
3741
3742 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3743 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3744
3745 // Now some values
3746 // Use ctrl to avoid hoisting these values past a safepoint, which could
3747 // potentially reset these fields in the JavaThread.
3748 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3749 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3750
3751 // Convert the store obj pointer to an int prior to doing math on it
3752 // Must use ctrl to prevent "integerized oop" existing across safepoint
3753 Node* cast = __ CastPX(__ ctrl(), adr);
3754
3755 // Divide pointer by card size
3756 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3757
3758 // Combine card table base and card offset
3759 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3760
3761 // If we know the value being stored does it cross regions?
3762
3763 if (val != NULL) {
3764 // Does the store cause us to cross regions?
3765
3766 // Should be able to do an unsigned compare of region_size instead of
3767 // and extra shift. Do we have an unsigned compare??
3768 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3769 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3770
3771 // if (xor_res == 0) same region so skip
3772 __ if_then(xor_res, BoolTest::ne, zeroX); {
3773
3774 // No barrier if we are storing a NULL
3775 __ if_then(val, BoolTest::ne, null(), unlikely); {
3776
3777 // Ok must mark the card if not already dirty
3778
3779 // load the original value of the card
3780 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3781
3782 __ if_then(card_val, BoolTest::ne, young_card); {
3783 sync_kit(ideal);
3784 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
3785 insert_mem_bar(Op_MemBarVolatile, oop_store);
3786 __ sync_kit(this);
3787
3788 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3789 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
3790 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3791 } __ end_if();
3792 } __ end_if();
3793 } __ end_if();
3794 } __ end_if();
3795 } else {
3796 // Object.clone() instrinsic uses this path.
3797 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3798 }
3799
3800 // Final sync IdealKit and GraphKit.
3801 final_sync(ideal);
3802 }
3803 #undef __
3804
3805
3806
3807 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
3808 if (java_lang_String::has_offset_field()) {
3809 int offset_offset = java_lang_String::offset_offset_in_bytes();
3810 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3811 false, NULL, 0);
3812 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
3813 int offset_field_idx = C->get_alias_index(offset_field_type);
3814 return make_load(ctrl,
3815 basic_plus_adr(str, str, offset_offset),
3816 TypeInt::INT, T_INT, offset_field_idx);
3817 } else {
3818 return intcon(0);
3819 }
3820 }
3821
3822 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
3823 if (java_lang_String::has_count_field()) {
3824 int count_offset = java_lang_String::count_offset_in_bytes();
3825 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3826 false, NULL, 0);
3827 const TypePtr* count_field_type = string_type->add_offset(count_offset);
3828 int count_field_idx = C->get_alias_index(count_field_type);
3829 return make_load(ctrl,
3830 basic_plus_adr(str, str, count_offset),
3831 TypeInt::INT, T_INT, count_field_idx);
3832 } else {
3833 return load_array_length(load_String_value(ctrl, str));
3834 }
3835 }
3836
3837 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
3838 int value_offset = java_lang_String::value_offset_in_bytes();
3839 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3840 false, NULL, 0);
3841 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3842 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3843 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
3844 ciTypeArrayKlass::make(T_CHAR), true, 0);
3845 int value_field_idx = C->get_alias_index(value_field_type);
3846 return make_load(ctrl, basic_plus_adr(str, str, value_offset),
3847 value_type, T_OBJECT, value_field_idx);
3848 }
3849
3850 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
3851 int offset_offset = java_lang_String::offset_offset_in_bytes();
3852 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3853 false, NULL, 0);
3854 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
3855 int offset_field_idx = C->get_alias_index(offset_field_type);
3856 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
3857 value, T_INT, offset_field_idx);
3858 }
3859
3860 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
3861 int value_offset = java_lang_String::value_offset_in_bytes();
3862 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3863 false, NULL, 0);
3864 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3865 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3866 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
3867 ciTypeArrayKlass::make(T_CHAR), true, 0);
3868 int value_field_idx = C->get_alias_index(value_field_type);
3869 store_to_memory(ctrl, basic_plus_adr(str, value_offset),
3870 value, T_OBJECT, value_field_idx);
3871 }
3872
3873 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
3874 int count_offset = java_lang_String::count_offset_in_bytes();
3875 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3876 false, NULL, 0);
3877 const TypePtr* count_field_type = string_type->add_offset(count_offset);
3878 int count_field_idx = C->get_alias_index(count_field_type);
3879 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
3880 value, T_INT, count_field_idx);
3881 }